Detection and Characterization
of
Figurative Language Use
in WordNet
Wim Peters
MA
Natural Language Processing Group
Department of Computer Science
University of Sheffield
July, 2004
Thesis submitted to the University of Sheffield
For the degree of Doctor of Philosophy
Detection and Characterization
of
Figurative Language Use
in WordNet
Wim Peters
Natural Language Processing Group
Department of Computer Science
University of Sheffield
Supervised by Professor Yorick Wilks
July, 2004
Abstract
For a computer to understand natural language, it requires knowledge of words and their
meanings. Achieving computer understanding of texts involves the use of lexical databases that
contain this knowledge, which can then be used in the interpretation process. Resources that are
available contain certain amounts of lexical knowledge in an explicit form. This is readily
retrievable from each resource in question. Furthermore, they also contain much knowledge in an
implicit form, for instance in the definitions associated with word senses in a machine-readable
dictionary, and the taxonomic structures in which word senses are sometimes organised.
The focus of this thesis is on making one type of such knowledge explicit, namely systematic
relationships between word senses. For example, the word university can mean a building, the
organization that is located in that building, and the people working for that organization. This
phenomenon is called metonymy, more specifically regular polysemy, and it occurs significantly
in language use. This type of linguistic knowledge can be captured and used for natural language
processing applications.
This thesis describes the discovery of such knowledge in WordNet by exploiting its hierarchical
structure. Candidate words with systematically related senses are selected, and the glosses
associated with the word senses under consideration are analysed in order to capture a semantic
relation between the senses. The patterns are compared with a semantically annotated corpus in
order to establish whether these patterns actually occur in text. Finally, these newly found
relations are integrated into knowledge fragments that form an extension of the knowledge
structures already explicitly available in WordNet. The applied method is not specific to
WordNet, but applicable to any resource that contains taxonomic information associated with
word senses.
i
The result is an extended knowledge base with increased amounts of background knowledge,
which helps reduce the knowledge acquisition bottleneck, and can be exploited for any text
comprehension process.
ii
Contents
Abstract.
Contents.
List of Tables
List of Figures
Acknowledgements
i
ii
vi
vii
viii
1 Introduction
1.1 Language and Dictionaries
1.2 The Role of Lexical Regularity in Natural Language Analysis and Processing
1.3 Objectives of this Research
1.4 Outcomes of this Research
1.5 Outline of Thesis Structure
1
1
1
3
3
4
2 Lexical Semantics and Figurative Language
2.1
Words and Senses
2.1.1 Homonymy and Polysemy
2.2
Views on Word Meaning
2.2.1 Context Dependency
2.2.2 Context Independency
2.3
Representation of Senses in Lexical Resources
2.4
Figurative language: Metonymy, Synecdoche and Metaphor
2.4.1 Metonymy
2.4.2 Synecdoche
2.4.3 Metaphor
2.4.4 Similarities and Differences between Metonymy and Metaphor
2.5
The Meaning of Metonymy
2.5.1 Associative Relations
2.5.2 Diachronic Relations
2.5.3 Metonymy as an Intra-Linguistic or Extra-Linguistic Notion
2.6
Psychological Perspectives
2.6.1 Figurative Language as a Cognitive Device
2.6.2 Child Development
2.6.3 The Frame Model
2.6.4 Frame Construal during Comprehension
2.7
Metonymy and Regular Polysemy
2.7.1 Lexicalization
2.7.2 Regular Polysemy
2.7.3 Lexical Rules
2.7.4 Typology of Metonymy and Regular Polysemy
2.7.5 Activation and Salience of Regular Polysemic Patterns
2.8
Conclusion
6
6
6
8
8
10
12
13
14
16
18
19
21
21
22
23
24
25
26
27
29
31
31
33
34
37
41
43
3 Regular Polysemy and Computational Linguistics
3.1
Computational Approaches to Capture, Detect and Resolve Regular
Polysemy
3.2
The Need for Data
3.3
Lexical Knowledge Bases
45
45
51
53
iii
4 Thesauri
4.1 Roget-type Thesauri
4.2 Manually Created Thesauri
4.3 Automatically Created Thesauri
4.4 WordNet and EuroWordNet
4.5 Thesaurus Word Senses and Dictionary Word Senses
55
55
58
58
60
63
5 Regular Polysemy Patterns in WordNet
5.1 The Occurrence of Attested Regular Polysemic Patterns.
5.2 The Automatic Extraction of Regular Polysemy Patterns from WordNet
5.2.1 Corelex
5.2.2 Extraction of more Specific Patterns from WordNet
5.2.3 Size of the Resulting Set of Patterns
66
66
68
69
68
71
6 Reduction of the Extracted Data Set on the Basis of Resource-internal Criteria 71
6.1 Size and Nature of the Extracted Data Set
6.2 Manual Evaluation in Order to Create a Gold Standard
6.2.1 Metonymic and Metaphoric Patterns
6.3 Automatic Data Reduction
6.4 Evaluation against the Gold Standard
6.5 Evaluation of Precision and Conclusion
7 Evaluation of the Data Set against a Semantically Annotated Text Corpus
7.1 Word Senses in Documents
7.2 Using Semcor for the Evaluation of RP Patterns
7.3 Discussion
71
72
74
76
85
86
87
87
89
95
8 Regular Polysemy as a Cross-Linguistic Phenomenon
8.1 The Distinction of Word Senses across Languages
8.2 Previous work
8.3 EuroWordNet
8.4 Lexical Triangulation
8.5 Evaluation
8.6 Universality of Regular Polysemy
8.7 Coverage and Extendibility
8.8 Conclusion
97
97
98
99
100
102
105
107
110
9 Building Knowledge Fragments
9.1 Establishing Semantic Relations between Word Senses Captured by a
Regular Polysemic Pattern
9.1.1 The Algorithm
9.1.2 Evaluation
9.1.3 Propagation through EuroWordNet
9.1.4 Discussion
9.2 Extension of Ontological Fragments
9.2.1 Extraction of Topical Knowledge
9.2.2 Comparison of Classification Schemes
9.2.3 Propagation through EuroWordNet
9.2.4 Measuring the Strength of Association
9.2.5 Conclusion
111
111
112
117
118
121
123
124
134
134
136
137
iv
10 Conclusions and Future Directions of Work
10.1 Original Contribution
10.2 Results
10.3 A Growing Knowledge Base
10.4 Future Work
10.4.1 Refinement of the Pruning Criteria Described in Chapter 6
10.4.2 Refinement of the Extraction Process
10.4.3 The Analysis of Text Corpora
10.4.4 The disambiguation of the Verbal Relations
10.5 The Separation of Regular Polysemic and Metaphoric Patterns
10.6 Directionality of Metonymic Relations
10.7 From Metonymy to Metaphor
138
138
139
139
141
141
141
142
143
144
145
147
Appendix A:
Appendix B:
Appendix C:
Appendix D:
Appendix E:
Appendix F:
Corelex Basic Types
Gold Standard Regular Polysemy Patterns
Success Rate of Relations Extracted in Section 9.1
Relations between Word Senses Extracted in Section 9.1
Words Subsumed by the Hypernym Pairs in Table 9.3
Evaluation of CASS Derived Rela tions between Gold Standard Hypernym
Pairs
Appendix G: Salience Scores for Gold Standard Patterns
149
151
167
173
183
187
Bibliography
202
190
v
List of Tables
4.1 Main Semantic Relations for Nouns
63
5.1 Regular Polysemic Patterns
66
5.2 Candidate RP Patterns for ‘university’
69
6.1 Randomly Evaluated Metonymic Patterns
74
6.2 Metaphoric Patterns
75
6.3 Unique Beginners in WordNet1.6
75
6.4 Incremental PCscore and Pruning Results
84
6.5 Scoring of Pruning Criteria
85
7.1 Mapping of Data Set onto Semcor Sense Collocations
92
8.1 Conceptual Coverage of English, Dutch and Spanish Wordnets
101
8.2 Experimental Results Summary
110
9.1 WordNet Concepts that Co-occur with music in Hypernymic Pairs
124
9.2 Correctly Extracted Relations with Bag-of-Words Technique
127
9.3 Remaining Valid Relations for ‘music’ Collocates from the Fully Pruned
128
Data Set
9.4 Incorrectly Extracted Relations with Bag-of-Words Technique
129
9.5 Additional Conceptual Collocates of ‘music’ for which no Explicit Relation
has been Found
130
9.6 Conceptual Collocates for ‘Discipline’ and ‘Fabric’ Ordered by Hypernymy
133
10.1 Output from the Cass Parser
142
vi
List of Figures
2.1 Taxonomy and Partonomy
17
2.2 Cognitive Linguistic View on Metonymy
24
2.3 Fragment of the ‘Monarchy’ Frame
28
2.4 Categorization Levels of Contiguity
41
4.1 synonymy, hyponymy and meronymy relations in WordNet
60
4.2 Architecture of the EuroWordNet Database
62
5.1 Words in WordNet Covered by the Pattern ‘Profession/Discipline’
68
5.2 The Regular Polysemic Pattern ‘Music/Dance’
70
5.3 The Regular Polysemic Pattern ‘Passage/Structure’
70
6.1 Multiple Inheritance
79
6.2 ‘Cloak/Dress’
81
6.3 ‘Building/Enclosure’
81
9.1: hypernymic chains for the senses of ‘tea’ captured by the regular polysemic
112
pattern ‘Plant-Beverage’
9.2 Graphical Representation of the Relation Extraction Algorithm
114
9.3 Hypernymic WordNet Chains for ‘Herring’
115
9.4 Expanded Ontological Fragment for ‘Discipline’-‘master’-‘Profession’
120
9.5 Expanded Ontological Fragment for ‘Person’-‘speak’-‘Language’
121
9.6 Vickery’s Hierarchical Faceted Classification Scheme for the Domain of ‘Music’
127
9.7 Graphical Representation of the Knowledge Fragment Provided by the Hypernymic
130
Pairs from Table 8.3
9.8 Extended Knowledge Fragment for ‘Music’
132
9.9 Additional Semantic Information from the Direct Semantic Context in EuroWordNet
135
vii
Acknowledgements
I would like to thank my supervisor, Prof. Yorick Wilks, for his advice and support over the
years.
Further, a word of thanks to my advisory panel, Dr. Louise Guthrie, Dr. Mark Hepple and Dr Guy
Brown, for their comments and guidance.
Finally, I want to express my gratitude to all the people who have enabled me to see this through
to the end.
viii
Chapter 1
Introduction
1.1
Language and Dictionaries
In this constantly changing world, the ways in which people use language changes according to
the needs that arise for successful communication. To enable this, everyday language needs to
adapt to new ideas and situations. Word meanings shift and change constantly in order to
accommodate this process of adaptation. Modern language dictionaries only provide a snapshot
of language. The vocabulary and word senses that are found in these resources give a static
picture of language use at a particular point of time (the date of publication). This so-called
synchronic information is sometimes enriched with information about the history of the
provenance of words, and the development of their senses. Dictionaries differ considerably in
their sense distinctions. For instance, the noun head has 32 senses in WordNet (Fellbaum, 1998),
but 54 in Ldoce (Proctor, 1978). Within individual dictionaries, senses overlap and their
boundaries cannot be drawn sharply (Kilgarriff, 1997).
1.2 The Role of Lexical Regularity in Natural Language Analysis and
Processing
In order to understand natural language, computers need to be able to access detailed information
about the meaning of words. For that purpose, they need to rely on richly structured lexicons.
Dictionaries have been used extensively for computer aided language processing tasks such as the
determination of the meaning of a word in a particular context (word sense disambiguation). The
problem of the fluidity of meaning described above makes it difficult to choose from a fixed set
1
of senses, because the senses quite often overlap, or, as in figurative la nguage, one sense is used
instead of another, related sense.
One way to approach the problem of fluidity of meaning, and cater for potential new word senses,
is to distinguish regularities in word sense combinations. Some word senses display the special
property that they are related in systematic ways. These regular combinations of senses apply, to
a certain level, to sets of words instead of individual words.
This systematicity of relatedness between word senses has been recognized by grammarians since
antiquity, and (ab)used by rhetoricians over the centuries. In scholastic theory, a number of socalled figures of speech (or ‘tropes’ from the Greek tropein, (to turn)) have been distinguished.
One of these exponents of so-called figurative language use, metonymy, makes use of this
relatedness between word senses. For instance, a word such as university can mean three different
things: the building, the organization and the people working for that organization. We say that
there is a metonymic relation between these three senses. This relation also applies to a number of
other words such as school, college, establishment and ministry. These words all share the same
sense distinctions. In this case, the metonymy displays regularity (it is applicable to a set of
words), and is therefore called regular polysemy.
Detection and resolution of metonymy can help to capture related senses, and help improve many
tasks involving natural language processing (NLP). Metonymic regularities have been exploited
successfully in various areas such as question answering (Stallard, 1993), anaphora resolution
(Markert and Hahn, 2002; Harabagiu, 1998) and information retrieval (Chugur, 2000).
Beyond NLP, the importance of making these relations explicit is evident for a number of related
disciplines. As described above, it is desirable for language technology so that knowledge
structures expressed by these linguistic phenomena can be captured and inferences that use the
relations between word-meanings can be drawn.
2
It is desirable for lexicography, so that the editor can make informed and consistent decisions
about presenting or indicating particular relations. It is desirable for linguistics simply because the
relations between metonymically related word senses are central to the language system.
1.3 Objectives of this Research
In this thesis I show that it is possible to use the information that is implicitly available in
electronic lexical semantic resources to assist the detection and analysis of figurative la nguage.
This involves the description of a process that defines constraints over sense collocations within
the WordNet electronic thesaurus (Fellbaum, 1998). The adopted methodology exploits the
taxonomic information that is explicitly present in that resource. The advantage of this approach
is that it is data driven, i.e. it is based on the analysis of a lexical semantic resource. The amount
of regularities has not been defined beforehand, but emerges in a bottom-up fashion based on the
distribution of word senses across the whole lexicon. The extracted regularities further extend the
already rich lexical structure of WordNet.
1.4 Outcomes of this Research
The thesis reaches a positive conclusion about the feasibility of extracting implicit information
from WordNet, and linking this information back into the knowledge base in explicit form. The
result is a knowledge base, which contains knowledge fragments that form an extension of
WordNet. The knowledge fragments consist of 4800 regular polysemic patterns, i.e. pairs of
related senses that are to a significant level refle cted in WordNet. These sense combinations form
the basic building blocks for knowledge structures at a higher level than combinations between
two individual senses.
3
A quantitative measure of this level of association between word senses gives an indication of the
probability of the link between the senses involved. Further, the database contains relations
between the word senses involved insofar these could be extracted from the WordNet glosses.
1.5 Outline of Thesis Structure
Chapter 1 Introduction sets out the scope of this thesis, discusses the motivations behind this
work, and describes the overall objectives and concepts.
Chapter 2 Lexical Semantics and Figurative Language introduces the various linguistic notions under examination in this thesis, most notably metonymy and regular polysemy.
Chapter 3 Regular Polysemy and Computational Linguistics discusses several approaches to
the detection and processing of regular polysemy in the field of computational linguistics.
Chapter 4 Thesauri describes the nature and content of different types of thesauri, the type
of lexical resource with which this thesis is concerned.
Chapter 5 Regular Polysemy Patterns in WordNet describes the detection of regular polysemic patterns in WordNet, and the methodology applied in the data analysis.
Chapter 6 Reduction of the Data Set on the Basis of Resource-internal Criteria describes the
analysis of the data set that results from the automatic extraction process described in the
previous chapter. The large amount of potential lexical regularity is reduced on the basis
of six criteria.
Chapter 7 Evaluation of the Data Set Against a Semantically Annotated Text Corpus describes a methodology that radically decreases the data for evaluation purposes according
to criteria that address dimensions of linguistic characterization other than taxonomical
systematicity. Sense collocations in a semantically annotated corpus make it possible to
evaluate the relation between figurative language use and discourse topicality.
Chapter 8 Regular Polysemy as a Cross-Linguistic Phenomenon explores the question
4
whether the obtained regularities have a higher level of universality, i.e. if the same
regularity can be detected across languages. For this purpose we have analyzed sense
distinctions from three different language resources in EuroWordNet (English, Dutch and
Spanish) and their behaviour with respect to regular polysemy.
Chapter 9 Building Knowledge Fragments investigates the feasibility of determining the nature of the relations between metonymically related senses. The regular polysemic
constructions are organized into knowledge fragments that form an extension of
WordNet.
Chapter 10 Conclusions and Future Directions of Work describes how ontological fragments
extracted in the previous chapters can be used for various applications. Further, future
extensions and refinements of the applied methods are discussed and how the obtained
lexical knowledge can be integrated into an incrementally growing knowledge base,
which will be useful for a variety of NLP applications.
5
Chapter 2
Lexical Semantics and Figurative Language
This chapter introduces the linguistic concepts that are the building blocks of this thesis. It gives
an overview of the various theoretical perspectives on word meaning, and then concentrates on
figurative language use, more specifically metonymy and regular polysemy.
2.1 Words and Senses
Lexical semantics, the field within which this thesis is embedded, is the science that occupies
itself with capturing and describing the meaning of words. One of the main issues within this
field is lexical ambiguity. Ambiguity arises when the same word can be used to mean different
things. Traditionally, two main types of ambiguity are distinguished, homonymy and polysemy.
2.1.1 Homonymy and Polysemy
Strictly speaking, homonyms are etymologically unrelated word senses that happen to be represented by the same string in a language. An example is ‘pen’ as an enclosure for confining
livestock or a writing implement with a tip from which ink flows. Word meanings that stand in
homonymic opposition are considered to be contrastive, and they are not predictable on each
other’s basis. This is the case for ‘pen’ above. Another example, ‘bark’, has the senses ‘the sound
made by a dog’ or ‘the strong outer covering of a tree’1 .
Polysemy, on the other hand, refers to a word having different senses that are in some way
related. The relations between these senses can be systematic or idiosyncratic, and are generally
considered to be complementary. For instance, line has, among others, the following two senses:
1
Taken from the Longman Dictionary of Contemporary English (Proctor, 1978).
6
“a line of people” and “a line drawn on a piece of paper” (Ravin & Leacock, 2000, p.2). These
are etymologically related through the semantic process of metaphor (see section 2.2.3) in the
sense that queuing people resemble a line drawn on a piece of paper.
Sometimes word senses express different aspects of the same situation or event. An example of
this is the dictionary entry2 of ‘sunset’:
1. the time in the evening at which the sun begins to fall below the horizon
2. atmospheric phenomena accompanying the daily disappearance of the sun
3. the daily event of the sun sinking below the horizon
Every sense describes a different part of one event. Sense one concentrates on the time aspect,
sense 2 covers the atmospheric phenomena, and sense 3 describes the event of the sun sinking.
Another example is the verb ‘marry’ 3 :
1. to take (a person) in marriage
2. (of a priest or official) to perform the ceremony of marriage for (2 people)
3. to cause to take in marriage
Each sense takes the perspective from another participant in the ‘marrying’ event: the
bride/groom, the priest and the parent.
The difference between homonymy and polysemy is not clear-cut. Judgment of semantic relatedness is subjective, and in many cases people will not agree on whether two senses are
related. Further, the criterion of semantic relatedness relies on a diachronic description of word
meaning. This is not always adopted by everybody in their judgment. Even linguists fail to reach
agreement in many cases. Also, it sometimes goes far beyond any common perception of
relatedness. For instance, it is doubtful whether many people would identify any relation between
the two following senses of line, taken from WordNet:
- A formation of people or things one after another.
2
3
Taken from WordNet (Fellbaum,1998); see section 3.4.
Taken from the Longman Dictionary of Contemporary English (Proctor, 1978).
7
- A single frequency (or very narrow band) of radiation in a spectrum.
This causes Apresjan to conclude that polysemy and homonymy are relative concepts (Apresjan,
1973), in the sense that there is no clear-cut distinction between the two.
2.2 Views on Word Meaning
Overall, there are two opposing extreme points of view with respect to the nature of word
meaning. The first one sees meaning as fully reflected in, and even constituted by its contextual
relations. The second sees meaning as wholly context independent and fixed. These opposing
views can be seen as the end points on a scale on which the opinions of several scholars are
plotted in the next sections.
2.2.1 Context Dependency
We start on the end of the scale that considers meaning as completely context dependent. This is
in line with the Firth’s postulate that the meaning of a word (at one level) can be known by the
company it keeps (Firth, 1957).
In a purely statistical approach to word meaning, Schütze (1998) states that words do not have
any semantic content, only semantic similarity to other words, which is measured by the
similarity of the contexts in which they appear. This means that word senses are in fact significant
clusters of usage in large text corpora. These clusters cannot be described a priori, and therefore
word senses can only be identified, but need to be manually examined and la belled in a later
stage.
Hanks (1987) adopts the same approach in the field of lexicography, and observes on the basis of
corpus analysis that homonymous senses of the same word co-occur with different sets of words,
as in the case of bank, with its senses ‘slope of land alongside a river’ and ‘financial
institution’:
8
‘‘On the one hand, bank co-occurs with words and expressions such as money, notes, loan,
account, investment, clerk, official, manager, robbery, vaults, working in a, its actions, First
National, of England, and so forth. On the other hand, we find bank co-occurring with river,
swim, boat, east (and of course West and South, which have acquired special meanings of their
own), on top of the, and of the Rhine.’’ (Hanks 1987, p. 127)
While this works for homonymous senses, it doesn’t for polysemous, where the distinctions
between the senses are more vague. For instance, in order to capture the following senses of bank:
‘a bracken bank’, ‘a bank of lights and speakers’ and ‘a river bank’, a lexicographer can adopt
various strategies. He or she can select a particular level of semantic granularity and lump these
senses together or split them into distinct senses (see next section), and select meaning
components for semantic description that are not necessarily activated in each sense, e.g. ‘is land’
versus ‘is elevated’ versus ‘situated beside water’ (Hanks, 2000).
Cruse (1986) approaches the variability of word senses from a lexical semantic point of view, and
describes how the linguistic meaning of a word is made up of the meanings of other words and
how it is influenced by the context:
“We can picture the meaning of a word as a pattern of (dis-)affinities with all the other words in
the language with which it is capable of contrasting semantic relations in grammatical contexts.”
(1986, p.16)
Cruse introduces the idea of sense modulation: A context can select the sense of a lexical unit if
there are multiple senses, but a single sense can be modulated by the context as well. Each sense
of a lexical unit is made up of semantic traits given by the other lexical units defining it.
According to Cruse, a semantic trait can have five statuses (degrees of necessity): criterial,
expected, possible, unexpected and excluded. To modulate a sense, a context can highlight certain
semantic traits of a lexical unit in a particular situation. For example, the pregnant cat brings the
female trait of cat from possible to criterial. This view emphasizes the changeability and fluidity
of word meaning. Cruse (1986, p. 68) distinguishes a gradient of establishment of senses. By this
9
he means that a lexical form can be associated with some senses, which are potential rather than
explicitly represented in the lexicon. Context can stimulate rules which generate an appropriate
sense. Meaning components can be potentially present, but not active (Hanks, 2000). Cruse
argues furthermore for the existence of sense-spectra, in which the senses of a lexical form lie
along a continuum, with no clear boundaries between them, and in some cases without an
encompassing sense. The sunset senses listed above form a good example of this.
Slightly further down the scale is Pustejovsky (1995), who argues against the view that word
meanings are fixed and inflexible, where lexical ambiguity must be treated by multiple word
entries in the lexicon. Rather, according to him, the lexicon can be seen as a generative system,
where word senses are related by logical operations defining the well-formedness rules of the
semantics. The lexical knowledge that he proposes as necessary for ambiguity resolution is seen
as factored out at four different levels of lexical representation: argument structure, event
structure, lexical inheritance structure and qualia structure (see Pustejovsky (1991) for details).
The qualia structure is an exact copy of the Aristotelian causes and forms a semantic network
structure that determines the ways in which a word is related to other words in the lexicon. A set
of generative devices connects the four levels, providing for the compositional interpretation of
words in context.
2.2.2 Context Independency
On the part of the scale covered so far the opinion is firmly against the checklist theory of word
senses (Fillmore and Atkins, 2000; Hanks, 2000). The conviction prevails that the view of word
meaning being based as a one-to-one fixed mapping relation between a word and a well-defined
object or concept must be abandoned for a view of meaning as having fuzzy boundaries, as being
elastic and context-sensitive. The other extreme sees meaning as wholly context independent and
fixed.
10
According to the main proponent of this view, Anna Wierzbicka, each word has only one core
sense (Wierzbicka, 1989). Context can only augment or refine but not alter this core sense, which
can be described by means of a limited number of semantic primitives. In partial agreement with
this, Wilks observes that senses are lexicon or procedure dependent, i.e. vary greatly between
dictionaries and applications, and translates this into a sense tagging procedure that lumps all
LDOCE word sense distinctions at the homonymic and part of speech level (Wilks and
Stevenson, 1997). This constitutes a back-engineering from context-specific senses in order to
establish single sense words at homograph level.
The inadequacy of a sense enumerative lexicon that contains a finite number of senses for any
polysemous lexical entry calls for mechanisms that capture the creation of novel senses, which
denote deviant usage from and form extensions of this single core sense. It is an established fact
that senses of a word extend in time in a non-random manner (Wilks et al., 1996). Within
cognitive psychology, this can be explained in terms of the prototypical character of word
meaning (Rosch, 1973). A range of psychological experiments have demonstrated that people
view some instances of a concept as ‘prototypical’ of it, and others as more or less marginal.
When applied to word meaning, different senses are seen to correspond to non-standard type of
instances of a prototypical concept. For instance, Lakoff (1987) analyses the structure of the
concept MOTHER, finding the concept to have ‘radial structure’, with various senses connected,
as by the spokes of a wheel, to the prototypical mother at the hub. From this point of view the
basic sense is then a gestalt (Koch, 1999) that corresponds to the prototypical meaning, out of
which, in various steps of significance, related senses emerge. The core sense can correlate to
one, or a cluster of senses from the list that are somehow marked by frequency of occurrence or
salience (see section 2.7.3 and 2.7.5).
11
2.3 Representation of Senses in Lexical Resources
Lexical resources are containers of linguistic knowledge. They come in the form of e.g. machine
readable dictionaries or lexical databases. If they encode word meanings they somehow need to
capture the fluidity of word meaning. When examining the ways in which multiple meanings are
represented in a dictionary or lexicon and how they are related to each other, it can be observed
that all resources supply lists of senses that are considered complementary and discrete (see also
the checklist theory in the previous section). Any links between senses are provided either by
means of nesting or by an explicit link within the dictionary definition. The principles that govern
these relations and the mechanisms that allow the creation of new senses are left implicit in the
structure and definitions.
There is little agreement among lexicographers as to the exact number of senses a word has or
should have. A similar scale as described in the previous section applies. At one end, there are the
‘splitters’, who prefer to factor out the meaning of a word into a maximally fine-grained list of
complementary senses. At the other end are the ‘lumpers’, who prefer coarse-grained sense
distinctions (Evens, 1988). Psycholinguistic evidence (Jorgenson, 1990) indicates that people
generally tend to distinguish around three senses for polysemous words. This renders many
lexical entries questionable with respect to the psychological reality and complementarity of their
sense distinctions, and reveals that they have been created by proponents of the splitters.
Different dictionaries have different numbers of senses for the same word. For instance, the noun
line has 39 senses in LDOCE, and 29 in WordNet. If we look at nouns only, WordNet (version
1.6) contains 94473 nouns, of which 12564 are polysemous. In this resource no distinction has
been made between homonymy and polysemy. Ldoce has 22590 nouns, of which 526 are
homonyms, and 7536 are polysemous. The ambiguity levels of both resources differ
considerably, from 13% for WordNet to 33% for LDOCE. Further, the observation that mapping
between resources is difficult (Fillmore and Atkins, 2000) seems to indicate that the senses are
12
not complementary, and that their boundaries are vague and overla pping. Kilgarriff (1992)
proposes another criterion for distinguishing word senses: a usage type (i.e. word meaning) merits
listing in a dictionary when it is Sufficiently Frequent and Insufficiently Predictable (SFIP). We
have encountered the term predictability earlier in our discussion of homonymy. Now it is
extended to cover polysemy as well. The claim is that complete predic tability leads to
lexicographic redundancy. Kilgarriff uses the example of the meaning of e.g. ‘house’ as an object
and as a picture of it. The latter should not be included in the dictionary. The SFIP criterion is
well in line with the change that has taken place in lexicography in the last thirty years. Before
then it was, and still is to some extent, customary that new words and word senses were suggested
to lexicographers by the general public. Lexicography was much more an introspective process.
The emergence of computer power and large electronic text corpora that act as representative
repositories of language use has influenced lexicographic methods. These resources enabled the
identification of clusters of usage on the basis of a grouping large numbers of instances. The
assessment of the nature of these statistically derived clusters, and therefore the qualitative
assessment of their predictability, however, remains a manual task.
2.4 Figurative Language: Metonymy, Synecdoche and Metaphor
When we use language we do not always mean the things we say in a literal sense. There are
various ways in which we can refer to or predicate on things in the world in an indirect way.
From Antiquity onwards, people have recognized the importance of these language mechanisms.
They have been recognized and classified as tropes. A trope (from the Greek "turn") is a language
device that produces a shift in the meaning of words. We owe the first insights into the nature of
tropes to Greek, Roman and medieval scholars in the art of rhetoric. Rhetoric is, in its broadest
sense, the theory and practice of eloquence, whether spoken or written, and was one of the Seven
Liberal Arts in the ancient and medieval educational curriculum, together with grammar, logic,
13
geometry, arithmetic, astronomy, and music. It is concerned with a consideration of the
fundamental principles according to which oratorical discourses are composed: invention,
arrangement, style, memory, and delivery, and involves the categorization of figures of speech.
From early on a number of tropes have found institutional definitions, and metonymy,
synecdoche and metaphor belong to this class of tropes. Other m
i portant tropes are irony,
overstatement ('hyperbole'), understatement ('litotes'), and euphemism or circumlocution
('periphrasis').
2.4.1 Metonymy
Various definitions of metonymy (from Greek “transferred title”) have been coined in the
literature. Although they differ from a theoretical perspective, there is a high degree of agreement
on the semantic principles underlying the process of metonymic extension (the forming of a new
metonymic sense of a word or the shift to an existing metonymic sense). The first use of the term
'metonymy' can be found in the works of classical scholars. In chapter 21 of his Poetics, Aristotle
distinguishes between four classes of metaphor, which include what was later to be called
metonymy and synecdoche. The first real definition of metonymy was given in the Rhetorica ad
Herennium:
Denominatio est quae ab rebus propinquis et finitimis trahit orationem, qua possit intellegi res
quae non suo vocabulo sit appellata. (Anonymous [Cicero] ,1954, IV 32, 43).
“Metonymy is the type of language use that takes its expression from neighbouring and related
things by means of which a thing can be understood that is not designated by its own proper
name.” (my translation)
By means of “neighbouring and related things” above, the author introduces an associative or
cognitive element into his definition. Analogously, Croft (1993, p.347) defines metonymy as “a
shift of a word meaning from the entity it stands for to a ‘contiguous’ entity.” The term
“contiguity” is a central concept in the definition of metonymic relatedness. In its literal sense, it
14
starts off as a spatial notion where two entities connect by literally touching each other, and is
extended to conceptual contiguity, or semantic relatedness. According to Nunberg (1978),
metonymy is a case of "deferred reference", in which a speaker uses a description of an entity and
succeeds in referring to it. Metonymic uses are considered a subcategory of "local" word uses, i.e.
uses which "a speaker believes are generally perceived as rational against a system of beliefs that
is available only to a sub-section of the community" (Nunberg 1978:186).
Lakoff and Johnson (1980) define it as a form of figurative speech, in which one expression is
used to refer to the standard referent of a related expression.
Radden (1999, p.21) defines metonymy as “a non-literal figure of speech in which the name of
one thing is substituted for that of another related to it. It has been described as a cognitive
process in which one conceptual entity, the vehicle, provides mental access to another conceptual
entity.”
Nehrlich and Clarke (1988) see it as the transfer of an old word to fit a new concept or object
related to the old one in space or time, by form or function.
Overall, according to scholarly tradition, metonymy in its basic form establishes a semantic
relation between two concepts that are associated with the same word form. It has primarily a
referential function in the form of substitution and transfer, that is, it allows us to use one entity to
stand for another. It also serves the function of enhancing understanding, and determines the
aspect we are focusing on (see section 2.5 for further discussion).
It is interesting to note that metonymy can often be seen as a form of textual ellipsis, where the
original phrase containing the explicit mention of the two entities involved, and the relation
between them (e.g. “the government of France”) is shortened into a metonymic expression
(“France”) with an implicit relation between government and country (Givon, 1967, Markert and
Nissim, 2002). Another example is the sentence “The bathtub is running over”, for “the water in
the bathtub is running over”. The question is whether metonymy is derived from explicit textual
representation by means of ellipsis, or whether the metonymic formation is made possible on the
15
basis of non-linguistic mental representations in the form of implicit conceptual propositions.
These then give in their turn rise to inferences that are exploited by the linguistic realization of
metonymy.
Whether textually based or conceptually motivated, metonymy does provide an abbreviation
mechanism (Warren, 1999), that highlights the implicitness of human communication (Grice,
1975). In my view, both theories represent related perspectives on the same phenomenon.
Overall, however, I am convinced that metonymy is more an extralinguistic than intralinguistic
phenomenon (see section 2.5.3). The explicit linguistic realization of every possible
metonymically related concept and the relation itself, on the basis of which ellipsis works in order
to create metonymic expressions, is not likely to be the default case. Economy of language use
and conformity to Gricean maxims (see section 2.6.4) support this hypothesis. Also, the
possibility of having several linguistic realizations that explicitly paraphrase the same metonymic
shift points to the conceptual basis of metonymy. For instance, “I have read Mao” can be
interpreted as an abbreviation of “I have read the works/books/writings/theories of Mao.” It
seems that in the abbreviated sentence the implied concept is more prominent than any specific
textual realization. It is, in my opinion, the case that fully explicit linguistic realisations are more
like paraphrases of an implicit conceptual inference. Words have the meanings we give to them
and we give them meanings by explanations (Wittgenstein, 1953).
2.4.2 Synecdoche
Traditional rhetoric has assumed that synecdoche (from Greek “accepted together”) describes
partonomies or meronymies, i.e. the whole for the part, or the part for the whole. Expressions like
“We need some new blood in the organization” cause an interpretational shift from blood as a
part of a person to the person him/herself.
According to Seto (1995, 1999), this definition fails to distinguish between two fundamentally
different semantic relations: the whole -part and the genus-species relations. The latter, also called
16
isa-relations, can be conceptualized as set inclusion, and this causes confusion between the
whole-part and isa relation. This is illustrated in figure 2.1 below.
A
A
B
B
D
C
E
...
D
E
C
...
Figure 2.1 Taxonomy and Partonomy
A taxonomic tree structure as depicted in the left of the figure can be seen as equivalent to the set
inclusion diagram on the right, on the basis of the reasoning that in a tree every node is a part of
the tree itself. In extensional terms, the set A comprises all elements from C, D and E, and
therefore these are included in A.
Cruse (1986) also seems to equate meronomy with taxonomy:
“Any taxonomy can be thought of in part-whole terms […] a class can be looked on as a whole
whose parts are its sub-classes”. (Cruse, 1986, p.179)
Consequently, no meaningful distinction can be drawn between metonymy and synecdoche, and
synecdoche is in fact seen as a special case of metonymy:
“We are including as a special case of metonymy what traditional rhetoricians have called
synecdoche, where the part stands for the whole […]” (Lakoff and Johnson, 1980, p.36)
Seto (1999) points out that partonomy is based on the perception of contiguity in the real world
(i.e. extension), while taxonomy is based on the perception of categorical hierarchy (i.e. intension
in terms of e.g. hypernymy and hyponymy) in our mind. He formally distinguishes between Erelations and C-relations. E-relations are contiguous relations between one entity and another
(e.g. “All hands on deck.”), whereas C-relations are conceptual relations relying on taxonomy,
and cover categorical transfers based on the semantic inclusion between a more comprehensive
17
and a less comprehensive category, i.e. between a hypernymic and a hyponymic concept, or
between genus and species. He considers the expression “We need some new blood in the
organization” as a metonymic E-relation (contiguity between entities in the real world), and each
case of synecdoche as a C-relation (taxonomic relation between concepts).
Seto’s main contribution consists of his separation of referential and conceptual meronymy.
According to his categorization the former belongs to the realm of metonymy, while the la tter is
captured by synecdoche.
There are two forms of synecdoche in this narrow sense:
1) taxonomic extension/generalization from species to genus, e.g. bread is generalized to
foodstuff in “earn one’s daily bread.”
2) taxonomic restriction/specification from genus to species, e.g. mortal for human.
On the basis of the discussions and definitions above we can conclude that synecdoche comprises
two specific subtypes of contiguity, meronymy and taxonomic generalization/specification.
Metonymy expresses any type of contiguity. In the remainder of this thesis, we will only consider
metonymy as the main class, and implicitly assume that it subsumes synecdoche.
2.4.3 Metaphor
Metaphor (from Greek "transference") is principally a way of conceiving one thing in terms of
another, and its primary function is understanding (Lakoff and Johnson, 1980, p. 36).
Metaphorical transfers express a semantic similarity relation between two concepts that yields a
conceptual mapping, a list of correspondences, between the source and target domains of a
metaphor (Lakoff, 1993). For example, anger can be compared with fire (Lakoff, 1987, p.388).
In short, the traditional meaning of metaphor is an implied comparison between two unlike things
that actually have something important in common. It expresses the unfamiliar (the tenor/source)
18
in terms of the familiar (the vehicle/target). When Neil Young sings, "Love is a rose," rose is the
vehicle for love, the tenor (Papafragou, 1995).
2.4.4 Similarities and Differences between Metonymy and Metaphor
In the literature the three types of figurative language use described above are much discussed
with respect to their similarities and differences. The ones that have been observed in the
literature are discussed below, without making any general statement about the (non-)identity of
both types of figurative language use.
One similarity is that both metonymy and metaphor are highly pervasive in language: there is
linguistic evidence of the widespread use of metaphor and metonymy in many la nguages (see
chapter 8). A second similarity is that both metonymy and metaphor are significant in the
development of new word senses and language change (see section 2.5).
A third similarity is that both types of language use involve non-literal meaning. Fourth, they
both involve an explicit source expression (that which is mentioned) which suggests an implicit
target (intended item of communication) (Warren, 2001). Fifth, both metonymy and metaphor
have a role in understanding because they focus on certain aspects of what is being referred to
(Lakoff and Johnson, 1980). Sixth, the concepts that are related by means of contiguity in the
case of metonymy or similarity in the case of metaphor are semantically orthogonal, i.e. they
belong to concepts that are not close to each other within a taxonomy, i.e. do not belong to the
same taxonomic branch. This enables the treatment of metonymy as instances of preference
breaking (Wilks, 1975; see section 3.1). For instance, in the metonymic pattern place/people,
place is a kind of location whereas people are a kind of living entity. The metaphor “LUST IS
HEAT” (Lakoff, 1987) links lust as a kind of desire to heat as a physical property. The semantic
proximity of the senses involved is expressed by the contiguity-versus-similarity relation.
However similar these two linguistic phenomena might appear, most modern linguists agree that
metaphor and metonymy are two distinct constructions arising from two distinct cognitive
19
operations (Warren, 2001). The first difference between them is that, historically, the basic
distinction between metaphor and metonymy is that metaphor indicates similarity and metonymy
reveals contiguity. This implies that metaphor is similarity-based, whereas metonymy is entitybased. This is corroborated by Lakoff and Johnson’s definition of metaphor in section 2.4.3 and
their definition of metonymy as “primarily a referential function, that is, it allows us to use one
entity to stand for another" (Lakoff and Johnson, 1980, pp.36-37).
Lakoff (1987) approaches the difference between metaphor and metonymy against the
background of an already more elaborate cognitive framework, using the notion of (cognitive)
‘domain’ (any kind of conceptualization relative to which semantic structures are characterized)
and ‘idealized cognitive model’ (a kind of gestalt that gives structure to reality, henceforth ICM).
For Lakoff, “a metonymic mapping occurs within a single conceptual domain which is structured
by an ICM, whereas metaphorical mappings involve a source domain and a target domain. The
second difference is that, in the case of metaphors, there is often more than one connector
between source and target, whereas in the case of metonymy there is never more than one relation
connecting source and target (Warren, 2001), if one does not take into account the phenomenon
of chaining, in other words, where metonymies are found in other metonymies. An example of
this phenomenon is the sentence “You will find better ideas than that in the ilbrary”, which
contains a chain of metonymies: ideas are expressed in words, words are printed on pages, pages
are in books and books are found in a library (Fass, 1991, p. 56). Thirdly, metaphors can form
themes that can be sustained through large sections of text. In terms of domain mapping, once
such a mapping has been established, all aspects of one domain can be used to describe the other
domain. For instance, for the metaphor “LUST IS HEAT”, verbs from the “HEAT” domain such
as “burn”, “sizzle” or “glow” can be used in order to describe the “LUST” domain. Conversely,
metonymy is a local phenomenon whose scope does not exceed the phrase level. Fourth, one can
use a non-metonymic reading and a metonymic reading together without causing zeugma (i.e. the
use of a word to cover more than two senses, though only one is appropriate in the context), as
20
shown by Warren’s examples: “Caedmon is a poet and difficult to read” is acceptable, whereas
“the mouse is the favorite food of cats and a cursor controller.” is not (Warren, 2001). Fifth, the
association taking us from source to target in metonymy has normally a different experiential
basis than the association taking us from source to target in metaphor. Metonymic association is
dependent on us having experienced source and target more or less simultaneously, whereas
metaphorical association does not require simultaneity in its quest for partial similarity (Warren,
2001). And finally, some scholars see a difference between the two in terms of regularity.
According to Apresjan (1973), the formation of metaphor by means of the same word is not
characterized by regularity of word sense combinations in a lexicon. Regularity is, in his view,
more typical of metonymy.
2.5 The Meaning of Metonymy
Overall, the semantic contribution of metonymy has historically been divided over two seemingly
disparate areas of semantic description. On the one hand there are the associationist relations that
exist between metonymically related word senses. On the other, semantic description is
concerned with the historical, diachronic aspects of the metonymic process, which are to be found
in the linguistic classification of sense development over time.
2.5.1 Associative Relations
Scholars have identified numerous patterns that underlie sense extensions (e.g. Lakoff and
Johnson, 1988; see section 2.4.1). One example is the association between university as a building
and an organization, where the relation can be specified as something like ‘is housed in’. The
rather vague classical rhetorical notion of closeness has been specified in later epochs to
semantically more specific and structural models. Closeness has been interpreted as a form of
semantic contiguity, in the sense that in many cases there are systematic relations between
21
metonymically related concepts that can be regarded as cognitive patterns in developmental sense
extensions (Nehrlich, 1992; see section 2.6), slots in idealized cognitive models (Lakoff, 1987) or
conceptual frames (Fillmore, 1977; see 2.6.3).
For example, in the sentence ‘The colonies
revolted against the crown.’ crown is used as a symbol for the monarchy (and therefore links to
the government), as well as denoting the traditional head ornament worn by the monarch.
Several types of contiguity relation can be distinguished in this case: a spatio-temporal locational
contiguity where the location of the crown on the head is prominent, a functional contiguity
inherited from the functional notion of monarch (the supreme power of a country) where the
crown represents royal power, and another functional contiguity between royal power and the
government of a country. The descriptive framework within which the contiguity relation as an
explanatory device for the phenomenon of metonymy is synchronic: the word senses and their
relations are considered from a present-day perspective.
2.5.2 Diachronic Relations
The other area of semantic description is concerned with the diachronic aspects of the metonymic
process, which are to be found in the linguistic classification of semantic change.
Polysemy in general develops through lexicalizations of a semantic change ultimately triggered
by a contextual and/or expressive ad hoc use of a word in discourse. Traditionally, cla ssification
of semantic change is divided up into four classes (e.g. Ullman, 1962):
•
specialization
•
generalization
•
metonymy
•
metaphor
Together, these two areas of semantic description may account for the intricate chains of
synchronic and diachronic sense development that can be observed. For instance, for metonymy
22
the relevant phenomenon is traditionally called "catachresis", which is, use of an inappropriate
term in order to fill a vocabulary gap. The following example is taken from (Koch, 1999) who
cites a DHLF entry (DHLF, 1992) of prison. In Old French, the original meaning was ‘the act of
seizing’. A resultative associative relation led to the interpretation of ‘captivity’, which in its turn
was linked to ‘place where prisoners are held’ by means of a spatial contiguity. Eventually,
Middle French obtained a functional link to ‘penalty of imprisonment’.
2.5.3 Metonymy as an Intra-Linguistic or Extra-Linguistic Notion
Another question we can ask is if metonymy is a relation between word senses or concepts, in
other words, if metonymy is an intra-linguistic or extra-linguistic phenomenon. From an intralinguistic perspective, if metonymy occurs between word senses, then every possible metonymic
extension needs to be present or available as semantic information in the lexical entry. As it is,
only conventionalized extensions become lexicalized, and are integrated into the dictionary (see
section 2.7.2). It may be obvious that it is not our knowledge of words but of the world that
determines contiguities. Therefore, metonymy is not a problem of linguistic structure, but a
problem that concerns the relation between language and the extra-linguistic world. Contiguity
has to be considered as constituting a conceptual, extra-linguistic and not an intra-linguistic
relationship (Koch, 1999; Blank, 1999). Systematic cases of metonymy (also called regular
polysemy, see section 2.7.2) can be captured in dictionaries by means of combinations of sense
distinctions that are valid for several words, and then take the form of lexical rules (see section
2.7.3). Still, these have been distilled from detected regularities in world knowledge.
Within the framework of cognitive linguistics, which considers the structure of language as
directly influenced by cognitive principles, words are seen the instantiations of concepts.
Metonymy is regarded as conceptual substitution. Radden and Kövecses (1999) distinguish a
number of relations between word forms and concepts, of which the following is the one we will
concentrate on in the remainder of this thesis.
23
Concept A (e.g. VEHICLE)
Word Form A (e.g. bus)
Concept B (e.g. DRIVER)
Word Form A (e.g. bus)
Figure 2.2 Cognitive Linguistic View on Metonymy
This represents cases of polysemy, where a sense extension of a word is conceived on the basis of
perceived contiguity between two concepts. They give the example ‘the buses are on strike’,
where the VEHICLE reading of ‘bus’ (concept A) is semantically contiguous with the DRIVER
reading (concept B) for which it is replaced in the example sentence.
Apart from polysemy as a metonymy-forming device, other processes of lexicalization play a
role, such as the phenomenon of ‘blocking’ where a metonymic extension is not possible because
of the existence of another word that covers the contiguous concept. For instance, the English
‘beef’ blocks the metonymic extension for ‘cow’ from animal to food (see section 2.7 for a more
detailed discussion). Another way in which metonymic polysemy is pre-empted is the
lexicalization of a contiguous concept by means of morphological processes such as
compounding and derivation. For instance, lemon (FRUIT) is generally not extended to mean
TREE on the basis of a ‘part of’ relation between the fruit and the tree. In order to express this
concept compounding is used in English to coin lemon tree.
2.6 Psychological Perspectives on Figurative Language
"A language without metaphor and metonymy is inconceivable: these two forces are inherent in
the basic structure of human speech." (Ullmann, 1962)
24
2.6.1 Figurative Language as a Cognitive Device
It is by now a commonplace in the pragmatic and psycholinguistic literature that the so-called
’figures of speech’ are not mere linguistic devices serving ornamental or literary purposes but
correspond to mental ‘figure’ grounded in cognition. Many scientists share the assumption that
metonymy is a cognitive phenomenon underlying much of our ordinary thinking and that the use
of metonymy in language is a reflection of its conceptual status (see e.g Lakoff, 1987, Gibbs,
1994). This is a distinct move away from the idea that non-literal languages constitutes a
deviation from a linguistic norm, which at the same time offers a way of a way of accounting for
the productivity of tropes in a variety of languages.
From a cognitive linguistic point of view the three figures of speech (metonymy, metaphor and
synecdoche) are seen as universal semantic and cognitive procedures or strategies which make it
possible to communicate concepts in a cost effective way by exploiting semantic relatedness.
From this perspective, then, figurative language involves basic cognitive processes rather than
deviant usage. These are seen as fundamental ways in which language conveys mental
representations. Seto (1999) portrays them as the corners of his so-called cognitive triangle.
Warren (1999) describes metonymy as an abbreviation device for conceptual access.
Lakoff's "metonymic model" has the following characteristics (Lakoff, 1987:84-5; Papafragou,
1995):
a) There is a "target" concept A to be understood for some purpose in some context.
b) There is a conceptual structure containing both A and another concept B.
c) B is either part of A or closely associated with it in that conceptual structure. Typically, a
choice of B will uniquely determine A, within that conceptual structure.
d) Compared to A, B is either easier to understand, easier to remember, easier to recognize, or
more immediately useful for the given purpose in the given context.
e) A metonymic model is a model of how A and B are related in a conceptual structure; the
relationship is specified by a function from B to A.
25
When such a conventional metonymic model exists as part of a conceptual system, it is claimed,
B may be used to stand, metonymically, for A.
In general, figurative language forms a central component of human thought, and plays an
important role in mental processes such as reasoning and inferencing. In this light, linguistic
communication can be seen as a process whereby people try to maximize their communic ative
success by minimizing their linguistic effort (cf. Wilks, 1975, Sperber and Wilson, 1986). It
conforms to the Gricean maxims, which describe rational means for conducting co-operative
exchanges (Grice, 1975; Levinson, 1983). Metonymy is a tool by means of which this goal can be
achieved, especially by means of the maxims of quantity (express as much as is necessary, no
more, no less) and relevance.
2.6.2 Child Development
From a developmental point of view, figurative competence is considered one of the basic
strategies children employ in order to extend their range of expression. The meaning extensions
they convey on the basis of this strategy seems to underlie the general use of metonymy as an
abbreviation device to shorten conceptual and linguistic distances in child as well as adult
language use (Nehrlich et al., 1999). Over-extensions based on the metonymical stretching of a
word are considered to be symptoms of cost-effective communication (Nehrlich et al., 1999).
Studies in language acquisition have investigated the phenomenon of overextension in child
language. Clarke (1993) states that children do not start out already knowing the meanings of the
words they are attempting to pronounce and use. They must first assign some meaning to them,
and that takes time. In fact, children's earliest mappings of meanings onto forms diverge from
adult usage in a variety of ways. Her studies and those of (Bowerman, 1978) revealed that
children apply overextension to categories that are perceptually similar, conceptually contiguous
or spatio-temporally related. For instance, ball will be extended to any circular object such as
doorknob and moon. Clark distinguishes two types of overextension: overinclusion (the extension
26
of a term to other members of the same taxonomy e.g. apple for apples and oranges, and
analogical extensions use of a particular term for objects from other taxonomie s on the basis of
perceptual similarity, e.g. comb for centipede.
Nehrlich adds another one on the basis of research by (Huttenlocher and Smiley 1993): spatiotemporal or functional contiguity. This type of metonymy manifests itself when e.g. a child says
toy while meaning the bag in which toys are normally kept. In tandem with Seto’s cognitive
triangle Nehrlich distinguishes three types of overextension: Synecdochical which exploit
conceptual contiguity relations within a taxonomy such as genus-species; metonymical:
perceptual, spatio-temporal and functional contiguity; and metaphorical: analogical extensions
based on perceptual similarity. The process of overextension stops when the structure of the
child's vocabulary mirrors the adult's. Metonymy is subsequently restricted to the use of words
the place of already known words, in contrast to covering up vocabulary gaps as in overextension.
This process gives rise to the forms of metonymy as we know them in adult language.
2.6.3 The Frame Model
The frame model, elaborated in artificial intelligence, psychology and linguistics in the last two
decades, provides us with a well established formalism for the representation of the cognitive
principles behind figurative language. Within this model, frames are defined as non-linguistic,
conceptual wholes, and can be found under various denominations in the literature, such as
scenes, scenarios, domains, and Idealized Conceptual Models (Minsky, 1975; Fillmore, 1975);
Lakoff, 1987; Schank and Abelson, 1977). Frames contain causal and relational information, and
can be used to represent a variety of knowledge about a wide variety of objects, actions and
events (Fillmore, 1968). The semantic contiguity that exists between senses that are related
through regular polysemy (see section 2.4.1) is consequently defined as the relation that exists
between elements of a frame or between the frame as a whole and its elements (Koch, 1999). For
27
example, a simple frame for the topic/domain ‘Monarchy’ (which we encountered in section
2.5.1) will contain the following elements and relations:
- monarch represents monarchy
- monarch wears crown
- monarch possesses royal power
- power is exercised by government
Traversal through the frame will yield the conceptually allowed metonymic extensions, such as’
crown represents monarchy’, or ‘crown possesses royal power’. This is illustrated in figure 2.3.
monarch
represents
wears
posesses
monarchy
crown
exercises
royal
power
government
Figure 2.3: Fragment of the ‘Monarchy’ Frame
Frame-type data structures are often invoked to account for inferential aspects of language
comprehension. The basic idea is that both linguistic and non-linguistic cues evoke frames which
can be bound to contextually available elements. Coulson and Kutas (1998) found positive
evidence that words cue the retrieval of frames from long-term memory, and these data structures
are then exploited in the construction of mental models of the semantic text representation. In
other words, lexical processing triggers the creative construction of cognitive models in working
memory by means of the integration of elements and relations consistent with the contextually
28
evoked frame. These models create a capacity for dynamic inferencing based on general
knowledge represented as static frames. These structures require the representation of causal and
relational information, slot/filler organization, and the existence of default values.
However, traditional implementations of frames have fallen out of favour for being overly brittle.
Wilensky (1986), for instance, argues that scripts are rigid data structures that cannot
accommodate events that are out of the ordinary. The conclusion must be that knowledge of
(proto)typical scenarios is necessary for comprehension, but far from sufficient. According to
psycholinguistic research, this knowledge is defeasible, i.e. it can be discarded at the cost of
additional processing (Rumelhart, 1987).
2.6.4 Frame Construal During Comprehension
The interpretation of a text is based on the activation of relevant frames or prototypical scenarios
from long-term memory. New information is then mapped onto these frames. The result of this
dynamic process is an integrated knowledge structure that represents the meaning of a text
(Coulson and Kutas, 1998). When the new information that comes in from the text does not fit a
prototypical scenario, the text needs to be reinterpreted in terms of frame construal. Overall, there
are three main situations in which this may occur. The first two are mentioned by Coulson and
Kutas (1998).
A) The violation of default assumptions contained in a frame.
In the sentence “Everyone had so much fun diving from the tree into the swimming pool we
decided to put in a little water.”, the word water is surprising, because, by default, we assume that
the pool is filled with water. A particular knowledge structure that forms part of a frame is
overwritten. In this case, the predication “the pool is filled with water” is negated.
B) Frame shifting
29
The interpretation process needs to substitute one frame for another. In the sentence “By the time
Mary had had her fourteenth child, she’d finally run out of names to call her husband.”,
the word husband causes a shift from a frame concerning baby names to a frame of names of
abuse.
C) Frame slot highlighting
This process involves metonymic shift: the focus shifts from one frame element to another. The
default interpretation process of a word selects the conceptually most salient sense that occupies a
slot position in a frame. Linguistic cues such as syntactic class (e.g. countability), preference
breaking, and discourse topic, as well as extra-linguistic cues, trigger the foregrounding (i.e.
salience shifting) of another slot filler within the frame. This filler is integrated into the
conceptual representation of the text. For instance, if we hear the utterance “Turn down the
potatoes please”, the phrase must refer to the heat under the pan that contains the potatoes. The
foregrounding in terms of conceptual salience (see section 2.7.3) is triggered on the basis of
extra-linguistic knowledge. There seems to be a repair from the side of the hearer on the basis of
Grice’s maxims (Levinson, 1983) of relevance and informativity. In order to maintain a
maximum level of text understanding the hearer will shift her focus of attention to the most
relevant and informative frame element.
There is a difference between this type of shift (C) on the one hand, and types A and B on the
other. It is plausible, as in the case of types A and B, that frame slot highlighting leads to delays
in sentence processing. The overall interpretation of the sentence might at first be primed for the
most frequent sense of a polysemous word. Psycholinguistic research, however, has found that
lexical complexity does not influence lexical access (Cutler, 1983). All senses of a polysemous
word are accessed simultaneously, and therefore salience shift from one frame element to another
does not necessarily cause a delay in interpretation time.
30
2.7 Metonymy and Regular Polysemy
2.7.1 Lexicalization
As described in sections 2.4 to 2.6, metonymy is in principle a productive process and can cover
any type of sense extension, be it systematic or ad-hoc (Clark, 1993). An example of an ad-hoc,
or nonce formation of a metonymic contiguity is the famous ‘ The ham sandwich is waiting for
his check’ (Nunberg, 1978), where a waiter refers to the customer by the dish he has ordered.
When this is said by a waiter in a restaurant, the metonymic expression can be used to identify a
customer who has ordered a ham sandwich. This use is justified only against a specific set of
beliefs shared by the waiters, according to which customers can be identified through their orders.
Against these beliefs, the mapping from orders to customers has become useful for uniquely
picking out a referent and distinguishing it from other possible candidates; in psychological
terms, it has acquired high cue-validity. The speech context enables us to understand these
utterances, and many metonymies come into brief existence in that way. The scope of extension
is in principle unlimited. Taylor (1995), who is active within the paradigm of prototype
semantics, points out that, in the absence of constraints, meanings can be infinitely chained via
family resemblance, so that everything ends up associated with everything else. According to
him, it might nonetheless be the case that certain instances of meaning extension are more
frequent, more typical and more natural, than others. In other words, we should be looking for
recurrent processes of sense extension, both within and across languages, rather than attempting
to formulate prohibitions on possible meaning extensions. If the relationship between concepts
involved in the metonymic extension is important enough for any type of speech community, be it
two waitresses in a restaurant or a whole country, the situationally constrained ad-hoc relationship
becomes increasingly conventionalized through processes such as pragmatic strengthening
(Traugott and Koenig, 1991, pp.193ff), or the conventionalization of inferences. This may
eventually lead to lexicalization, i.e. the explicit inclusion of the metonymic sense in the lexicon.
31
This process bears a great resemblance to the historically motivated semantic changes described
in section 2.5.2 for the French prison, where the link between ‘act of capturing’, ‘captivity’ and
‘place for prisoners’ became conventionalized. Whether or not a particular reading for a word is
lexicalized depends on the frequency of occurrence of the sense, and the level of convention that
the similarity or contiguity relation has attained in use. The consolidation of sense extensions into
attested sense distinctions and, even further, into regular polysemy, shifts and changes, reflecting
various degrees of lexicalization.
As already indicated above, any discussion of regular polysemy is never far from the
lexicon/pragmatics divide. Consider the words turkey and ostrich. Both may be used to denote a
bird, or the food which is the meat of the bird. However the use of turkey to mean the food is
familiar (to the speaker community of which the author is part) in a way in which the use of
ostrich to mean the food is not. The food reading of turkey will be found in dictionaries, the food
reading of ostrich will not. While both words can mean the food, this is lexicalized only for
turkey, and we shall say that this is a sense of the word, whereas for ostrich, there is no such
sense; that is, there is no such meaning listed in the lexicon and we resort to pragmatic processing
in order to interpret ostrich as food. In general, it tends to be the case that systematic semantic
relations between word senses are lexicalized, i.e. they are explicitly listed in dictionaries and as
such independent of a pragmatic situation. For example, The White House is on the one hand an
institution and on the other a building. The semantic relation between the two senses is ‘is housed
in’. This relation is also applicable between conceptually equivalent lexicalized senses of other
words, such as school, college, academy and hospital. Therefore, the constellation of lexicalized
senses and conventionalized relations is called regular polysemy.
Overall, figurative language does not consist of a number of natural classes but constitutes rather
a continuum of cases. The end points of this continuum are on the one hand novel, one-off uses,
and on the other fully conventionalised uses that have entered the lexicon (Papafragou, 1995).
32
The intermediate space is taken up by figurative uses with varying degrees of
conventionalization.
2.7.2 Regular Polysemy
In the previous section we established that regular polysemy is the phenomenon that, as a subset
of metonymy, captures the regularity and systematicity of sets of instantiations of relations
between word senses. It can be defined as a subset of metonymically related senses of the same
word displaying a conventional as opposed to novel type of semantic contiguity relation. This
relation holds, by definition, for related senses of two or more words (Apresjan, 1973), i.e.
represents a lexicalized pattern, not a nonce formation (a pragmatically defined novel metonymy),
and can therefore be called regular. The word senses involved are related in systematic and
predictable ways. It is this subtype of metonymy that we concentrate on in this thesis.
In principle, a pattern is considered regular if the combinations of related senses are valid senses
for more than one word. Apresjan defines regular polysemy as follows:
’Polysemy of the word A with the meaning ai and aj is called regular if, in the given language,
there exists at least one other word B with the meanings bi bj , which are semantically
distinguished from each other in exactly the same way as ai and aj and if ai and bi , aj and bj are
non-synonymous. Polysemy is called irregular if the semantic distinction between ai and aj is not
exemplified in any other word of the given language.' (Apresjan, 1973, p.16)
He suggests that many types of regular polysemy are productive, i.e. that if a word has a meaning
of a particular type (e.g. the type of ai and bi above), it can also be used with the meaning of
another type (e.g. the type of aj and bj ). From this observation, the idea of representing this kind
of polysemy with statements (rules) expressing the relationship between semantic types
developed. Regular polysemy is Apresjan’s term. Closely related phenomena are logical
polysemy (Pustejovsky, 1995), sense extension (Copestake and Briscoe, 1996) and transfers of
33
meaning (Nunberg, 1996). According to Apresjan, the semantic regularity is a distinctive feature
of metonymic transfers.
2.7.3 Lexical Rules
From a theoretical point of view, there are two extreme viewpoints on regular polysemy with
respect to its representation. On the one hand, Copestake (1996) suggests regular polysemy
should be represented as a set of processes defined in the lexicon by means of lexical rules.
Nunberg (1996), on the other hand, takes the pragmatic tack, representing the productive
processes outside the lexicon as pragmatically defined meaning transfers (e.g. Nunberg, 1996).
The boundary between lexical and pragmatic explanations of regular polysemy is determined by
synchronic analysis of language in use on the one hand and lexicographic practice on the other.
The latter represents the historical picture of the development of a word's senses and is heavily
influenced by the type of resource it wants to produce. Dictionary users and size determine the
choice between lumping and splitting word senses (Evens, 1988, Kilgarriff, 1997), and no matter
how fine-grained a dictionary’s inventory of senses is, it can never be exhaustive, nor can it fully
cover the dynamic meaning potential of language. Ostler and Atkins (1991) argue that knowledge
of regular lexical semantic relationships forms part of linguistic knowledge, and cannot simply be
a reflection of regularities in the world or “speakers' free play with analogy” (p. 77), because of
the interaction of lexical factors with the application of the relationships. An example of this is
the phenomenon of blocking (see also section 2.5.2), which occurs when the application of a rule
is blocked through the prior existence in the lexicon of a word which is already in the place of the
potential output. Blocking has two subtypes: semantic pre-emption, where it is impossible to
metonymically extend the meaning of a word (e.g. cow as animal to cow as meat) because another
lexical item already exists that occupies that semantic niche (beef), and lexical pre-emption,
where the word form to be generated already exists in the lexicon with a different sense. There
are many exceptions to the constraints imposed on lexical rules and cognitive principles. Apart
34
from blocking other overriding factors may come into play, such as rhetorical effects (e.g. ‘the
pen is mightier than the sword’) and social-communicative effects (e.g. euphemism). Ostler and
Atkins argue for explicit representation of the relationships in terms of Lexical Implication Rules
(LIRs) which generate derived lexical entries from base lexical entries. In this sort of rule, we
find the starting point for a dynamic view of the (computational) lexicon. By making
generalisations about potential syntagmatic alternations which a class of words can undergo, one
aspect of creativity in language use can be accounted for. For example, if a system captures a rule
such as ‘a countable noun which refers to an animal can be used as an uncountable noun which
refers to the meat from that animal’, it would have a basis for distinguishing the senses of dog in
John walked his dog and We ate dog for dinner last night, and further would allow for novel
instances of this ‘animal’
‘meat’ rule, such as ‘I tasted aardvark yesterday’. This sense of
aardvark cannot be considered established and a part of the fixed lexicon, but can still be
interpreted given knowledge of the lexical generalization.
The use of lexical rules depends on a structured lexicon. That is, it depends on a lexicon in which
various generalizations are captured about the semantic classes to which words in the lexicon
belong (e.g. animal nouns). This stems from the need to constrain the input to the lexical rules, to
restrict their application to those words that participate in the regular relationships. It is difficult
to imagine a semantically motivated way of defining the appropriate input to a rule without
representing the semantic relationships between words. Lexical rules would be extremely difficult
to formulate under an unstructured multiple lexical entry view of the lexicon, in which entries are
not grouped into classes.
An important aspect of lexical rules is their directionality. Constraints over semantic classes can
be formulated for lexical implication rules in order to capture the scope of the potential
metonymic extension. The view of the directionality of lexical relationships has, however, been
challenged by, among others, Nunberg (1996) and Bredenkamp et al. (1996). One central
question which arises if directionality of lexical rules is assumed is how to determine which of the
35
senses is the base sense and which is the derived sense. The base sense can be chosen on the basis
of the criterion of salience.
According to Langacker, the choice of vehicle and target in default cases of metonymy appears to
be motivated or constrained by cognitive principles of salience:
“Other things being equal, various principles of relative salience generally hold: human > nonhuman; whole > part; concrete > abstract; visible > non-visible; etc”. (Langacker, 1993, p.30).
Radden and Kovecses (1999) distinguish three general determinants of conceptual organization
that interact and overlap in certain ways:
a)
Human experience. Our anthropocentric view of the world accounts for e.g.
human > non-human: PRODUCER FOR PRODUCT (I’ve got a Ford) and
CONCRETE FOR ABSTRACT: brain for intellect.
b)
Perceptual selectivity. For instance, IMMEDIATE OVER NON-IMMEDIATE accounts for selecting stimuli in our immediate spatial, temporal or causal
environment: “I’ll answer the phone” for “I’ll answer the person speaking at the other end
of the line” is motivated by spatial immediacy.
c)
Cultural preferences. Certain members of a category are more salient than other
members, therefore TYPICAL OVER NON-TYPICAL: coughing is a typic al symptom
of a cold, so “You’ve got a bad cough” can refer to a cold.
As a rule, more than one cognitive principle does in fact apply to a particular case of metonymy.
For example, ARTIST FOR HIS/HER WORK as in “I love Shakespeare” is motivated by a
bundle of cognitive principles: HUMAN OVER NON-HUMAN and CONCRETE OVER
ABSTRACT. Most instances of metonymy, however, are not ‘fully’ motivated in the sense that
not all cognitive principles converge. Rather, we have a continuum of motivation ranging from
fully motivated default metonymies to weakly or unmotivated non-default metonymies. For
36
instance, “I’ll answer the phone” is consistent with the principle IMMEDIATE OVER NONIMMEDIATE, but is in conflict with the principle HUMAN OVER NON-HUMAN.
As pointed out by Kilgarriff (1992:89-90), there are cases in which one sense of a systematic
relationship is most salient for certain instantiations of it, while the other sense is generally more
salient for other instantiations. An example is found in the tree/wood relationship: the name of a
tree can be used to refer to its wood. For oak, ash, and many other trees/woods, the tree sense is
generally most salient (unless one is a carpenter). However, for teak and mahogany, the wood
sense seems to be most salient. Similarly, for turkey, the meat sense is more salient than the
animal sense, while the reverse is true for dog.
Concluding, the use of lexical rules in NLP systems can increase the flexibility of the lexicon in
those systems. Lexical rules define the space of regular, productive sense extensions which can be
used to generate new senses from existing ones. The addition of frequency probabilities to the
lexicon can aid in preventing non-established senses from being used in generation, and can
influence the interpretation process (Bredenkamp et al., 1996). It needs to be borne in mind that
they apply to default cases and describe, apart from the conceptual processing, the associated
lexical mechanisms.
2.7.4 Typology of Metonymy and Regular Polysemy
There are several ways in which instantiations of metonymy can be classified. Traditional and
more recent typologies give more or less complex lists of types of metonymy, like ‘inventor for
the thing invented’ and ‘container for content’ (cf., e.g. Quintilian (Meister, 1886), Lakoff and
Johnson (1980), Ostler and Atkins (1991), and Pustejovsky, (1995)). Most of these relations have
been arrived at by examination of a limited quantity of linguistic material (texts, dictionaries) or
introspection. Below are a few examples.
1. container/containerful
cup
37
2. animal/food
lamb, chicken
3. animal/skin
crocodile
4. plant/food
banana
5. product/producer
newspaper, Honda
6. substance/colour
jade, amber
7. object/shape
pyramid
8. language/people
Spanish
9. music/dance
waltz
10. figure/ground
door/window
11. place/people
city, New York
12. source material/object silver
13. place/product
china, camembert
Norrick (1981) produced a classification that comprises 18 types of metonymic principles:
1. cause -effect
a. cause-effect
b. producer - product
c. natural source - natural product
d. instrument- product
2. II Acts and major participants
a. object-act
b. instrument- act
c. agent - act
d. agent - instrument
3. III part- whole
a. part - whole
b. act - complex act
c. central factor - institution
4. IV container- content
a. container - content
b. locality - occupant
c. costume - wearer
5. V. experience - convention
a. experience - convention
b. manifestation -definition
6. possessor - possession
a. possessor - possession
38
b. office holder - office
Bredin remarks that this is in fact not a typology of metonymy, but an enumeration. In fact, the
list above is indeed non-exhaustive (Bredin, 1984). It does not reflect the true range and
distribution of the phenomenon of figurative language use.
Another way of classifying metonymies is to subsume them under more general types of semantic
contiguity, such as spatial, temporal and causal. Typologies of this kind have been proposed by
Ullman (1962) and Warren (1999). Warren (1999) distinguishes five types of contiguity, each of
which comprises several subtypes: COMPOSITION (e.g. source material/object); CAUSATION
(e.g. product/producer); POSSESSION (e.g. animal/skin and substance/colour); LOCATION (e.g.
place/people) and REPRESENTATION. These general classes subsume in their turn more
specific classes of metonymy such as the ones listed above. For instance, the causal dimension
comprises relations such as AGENT-ACTIVITY (‘to butcher a cow’) and ACTIVITYINSTRUMENT (‘the flight is waiting to depart’). Many of these higher level types rely on case
roles that substitute for each other by means of metonymy. An examples is AGENT for
INSTRUMENT (‘Blair bombed Basra’). The idea that metonymy is a case role substitution is
elaborated on by Fass (1991). From the point of view of case role or valence semantics, this is a
valid approach. Blank (1999) objects that in order to see the case role effect, we need concrete
sentences that must contain verbs. Lexicalized metonymies, however, show this metonymic sense
without being syntactically actualized. Thus, he concludes, we can only state that a case-role
substitution might have occurred in the past. He rejects the suggestion that metonymy relies on
syntagmatic contiguity: in sentences such as ‘The buses are on strike’ the syntactic position of the
subject remains the same, i.e. the case-role substitution is not syntactically marked. Only the
semantic role of this subject has been replaced by another role taken out of the paradigm: the
human agents that drive vehicles have been replaced by the vehicles they drive. Indeed, the
semantic relation between human agents and vehicles must be pre-existent to the metonymic use
39
of ‘buses’ in this utterance, because there is no aparent syntactic contiguity. The metonymic shift
has been made possible by the conceptual inferences that people drive buses and are capable of
striking. This syntagmatic contiguity can be found elsewhere in a text in the form of an explicit
realization of the semantic relation between human agents and vehicles. From a computational
point of view, this text can be in the form of a text corpus or a set of dictionary definitions. On the
basis of these resources the pre-established contiguity relation DRIVE can be gleaned. This is
basically the approach taken in chapter 8.
Overall, there is a scholarly tendency to reduce the different types of metonymy to a smaller
number of more general types of contiguity. Blank (1999) takes this to the extreme while
distinguishing two basic cognitive domain categories for metonymies that derive directly from the
two fundamental ways of conceptualizing real-life situations as static frames or dynamic
scenarios. He coins the terms co-presence where the referents are present at the same time, and
succession where the frame elements stand in a sort of a causal, final or consecutive relation, i.e.
do not co-occur synchronically. He claims the two general classes hang closely together with two
fundamental models of human conceptualization:
1 the synchronic model, in which all aspects of a given situation or a ‘system’ are equally present
and where time is excluded;
2 the diachronic model, where the processual, consecutive character of things and events are
highlighted (Blank, 1999, p. 179).
They conform roughly to the concepts frame (see section 2.6.3) and scenario. The latter can be
conceptualized as a contiguity schema between frames.
In summary, the typology of metonymies can be regarded as involving three levels of abstraction,
as illustrated in figure 2.4 below: two superordinate domains of contiguity (e.g. co-presence and
succession), a principally open list of types of contiguity (image schemas in Lakoff and Johnson,
1980), and concrete metonymies as they occur in la nguage use. The types of contiguity (class II)
are highly recurrent and conventionalized. They can consist of several levels of abstraction, from
40
e.g. the types distinguished by Warren to more specific classes that consist of combinations of
certain hypernyms in a taxonomy that subsume two or more words (Apresjan, 1973). The
concrete metonymies (class III) are realized at sense level, i.e. the highest level of granularity.
I Domains of contiguity
II Types of contiguity
III Concrete metonymies
Figure 2.4: Categorization Levels of Contiguity
2.7.5 Activation and Salience of Regular Polysemic Patterns
The term ‘contiguity’ as introduced in section 2.4.1 is a general term that covers a variety of
semantic relations that can be established or induced between the word senses or concepts
involved in regular polysemic patterns. Pustejovsky (1995) distinguishes between two different
types of sense combinations in a regular polysemic pattern, which he calls “dotted types”. In one
type both aspects of the meaning are activated simultaneously in the comprehension process. For
instance, a book is a physical object (with pages etc.) and a source of information. These two
knowledge fragments are accessed from lexical memory (or long term memory as in the previous
section) at the same time when an utterance containing book is being processed in discourse.
Within the interpretation, the combination of the two semantic labels acts as a complex object in
discourse processing, i.e. the discourse comprehension process uses their combination and not the
labels individually. In dot notation, this is expressed by a closed dot: book : ‘information •
physical’. In the other type, only one of the meaning aspects is triggered. For instance, rabbit has
an animal and a food sense. This is expressed in dot notation as ‘animal ? food’. In this case,
either the ‘animal’ or the ‘food’ reading is activated through the context, i.e. on the basis of the
41
dynamic conceptual model built up during text processing. However, both components remain in
the background in the discourse interpretation process.
This means that if a particular context chooses an ‘animal’ reading, the ‘food’ reading is, in
principle, available for inferencing. For instance, the following bridging assumption makes use of
the activation of this background knowledge:
4)
I bought a new hutch for the rabbit. Christmas would have to wait a little longer.
In my opinion, this is an oversimplification. Cruse (2000) sees the level of ‘unity’ or ‘integrity’,
as he calls the level of simultaneous activation of both senses (or ‘facets’ in his terminology in
discourse), as a continuous scale. Therefore, in the case of dotted types, it is impossible to
determine the nature of the types for each combination on the basis of this two-way classification.
The meaning components are not necessarily activated simultaneously, and can coexist peacefully
as part of the word’s meaning potential (Hanks, 2000).
Buitelaar (1998) gives the following example of the impossibility to combine two incompatible
meanings (p.34):
5)
The newspaper that fired its editor fell on the floor.
However, it is conceivable to present a more elaborate specification of a particular situation in
which this utterance is felicitous:
6)
The class was looking at newspapers and discussed the ways in whic h they treated their
editors.
The newspaper that fired its editor fell on the floor.
42
We seem to have to take into account the notion of a level of activation. This level ranges from
‘simultaneous’ to ‘deferred’ in discourse processing, and depends on linguistic cues such as
syntactic class (e.g. countability: this explains the explicit choice between animal and food),
preference breaking, discourse topic and nature of context. I contend that the strength of the
associations between senses determines the level of activation in discourse. It seems that the
weaker the combinations of senses are, the more contextual clues they need in order to force both
readings to be taken into account in the interpretation process. The default strength and salience
of the associations or contiguities between senses that are captured by regular polysemy is
reflected by their lexicalization and frequency of occurrence (see section 8.2.4). They form the
lexical reflection of the prototypical situations that function as the initial frame-based default
interpretation that eventually leads to the linguistically and situationally determined meaning.
This frequency of lexicalization gives us a view on figurative language use as a continuum from
nonce formation to fully conventionalized and lexicalized word sense. It is my impression that
lexicalization and conventionalization of regular polysemic patterns cover the same axis in
semantic description. The two different dot types form the ends on the scale of
conventionalization of regular polysemic patterns. This scale reflects sense formation as a
dynamic process, which is in constant flux, and therefore the lexicon with its sense distinctions is
merely a snapshot of synchronic language use. We will find instances of figurative language use
from any position on the scale.
2.8 Conclusion
In this chapter we have described the nature of regular polysemy, and embedded the term in its
linguistic, psychological and historical context. We have established that regular polysemy is a
systematic and productive phenomenon in language. This warrants any computational attempt to
automatically capture it, in order to enable a computer to exploit this systematicity for automatic
43
language processing. The next chapter gives an overview of computational approaches to regular
polysemy, and discusses the importance of its representation in a lexical knowledge base.
44
Chapter 3
Regular Polysemy and Computational Linguistics
This chapter discusses several approaches to the detection and processing of regular polysemy in
the field of computational linguistics.
Regular polysemy has generated a great amount of interest in natural language processing
because it is in most cases regular, productive and frequent. Markert and Hahn (2002) found 17%
of all utterances to be metonymic expressions in 27 German magazine texts. Metonymy
resolution can also improve many language engineering tasks. Stallard (1993) cited a 27%
performance improvement by incorporating metonymy resolution into a question answering
system about a limited domain (commercial air flights), which had to understand metonymies
such as “Which wide body jets serve dinner?”. Anaphora resolution, a crucial task in many NLP
applications, often depends on metonymy recognition as well (Markert and Hahn, 2002;
Harabagiu, 1998). In a sentence like “Germany has decided to veto America’s proposal, at the
evident peril of Berlin being considered part of an axis, this time an axis of evil.”, the coreference
chain can only be established if “Germany” and “Berlin” are recognized as metonymies for the
government of Germany. In information retrieval, Chugur (2000) examines the usefulness of
regular polysemic patterns for document clustering.
3.1 Computational Approaches to Capture, Detect and Resolve Regular
Polysemy
One of the first attempts at capturing figurative language use in an NLP system is preference
semantics (Wilks, 1975), a meaning representation scheme whereby semantic expectations are
stored as soft constraints, or preferences, rather than as strict rules. In this approach, novel
45
sentence types, such as metaphors, which are not accommodated exactly by some case frame in
the semantic lexicon, are nevertheless assigned some representation at the semantic level, rather
than being rejected out of hand as anomalous. Thus, the system exhibits a considerable degree of
flexibility in handling input of an unanticipated nature. One of the basic assumptions in this
approach is that figurative language necessarily results in preference-breaking at the semantic
level. For instance, if a non-conventional metaphor is encountered, a suitable frame is found for
which it is not entirely suited by selecting frame with the least number of preference breakages.
The interpretation mechanism of preference semantics relies on so-called pseudo-texts. A pseudotext is a contextual predicate bundle in a frame-like representation, which demonstrates the
correct semantic usage of a concept in a variety of different contexts. Pseudo-texts essentially
cover the spectrum of valid semantic usages, associations and contexts in which a concept may be
employed normally. Thus, a pseudo-text for the concept CAR might include the information that
cars consume gasoline and transport people and goods from one place to another. Whenever a
preference violation is triggered, it becomes the task of the extended sense interpretation system
to project the ill-suited case-frame structure onto a more suitable case-frame chosen from the
relevant pseudo-text. For example, when given the metaphor "My car drinks gasoline" (Wilks,
1975), a violation of the semantic preference that causes DRINK to seek an animate agent and a
potable patient is triggered, prompting the metaphor interpretation process to kick in and select
the alternative case- frame CONSUME from the pseudo-text for CAR on the basis of the
inference that DRINK is a taxonomic hyponym of CONSUME. The inference steps that need to
be taken are called projection. They consist of the replacement of a template formula by a new
one constructed by access to a pseudo-text. Projection is activated only in the case of preference
breaking, i.e. in the case that a word is encountered in a position with which a clashing preference
is associated and no proper semantic representation of the text can be created on the basis of
satisfying the preferences.
46
Some might argue that a drawback of this approach is the lack of psychological reality.
Psycholinguistic findings (Gibbs, 1994) indicate that, when processing figurative la nguage in the
form of conventional metaphors and idioms, people do not first test out a literal interpretation.
The time taken to process figurative language is no longer than the time taken to interpret
straightforwardly compositional sentences. Given these findings, one might argue that any
automated interpretation system for figurative language should therefore take all possible
interpretations into account simultaneously. The range of possible interpretations that compete for
priority can be constrained by their plausibility in the form of e.g. sense frequency and contextual
information. Preference breaking is a form of implausibility detection, and seems from that point
of view a good starting point for the detection of figurative language within a computational
approach.
Fass (1991) builds upon the preference-breaking idea in his Collative Semantics approach to offer
a unified account of coherence, metonymy, metaphor and lexical ambiguity within a single
semantic framework. He developed a program called Meta5 that uses a hand coded lexicon with
480 senses. Each sense is enriched with sense frames: semantic information such as taxonomic
relations and preferences. According to the preferences expressed by the verbs in predicative
structures the program determines whether the subcategorized object conforms to the preference.
If this is not the case, the literal meaning expressed by the preference is not applicable, and
inference rules are applied in order to find an interpretation for the expression in terms of
metonymy, metaphor or anomaly. These rules make use of what he calls “cells”, which are in fact
identical to Wilks’ pseudotexts: triples of subject-verb-object or agent-action-patient relations that
that describe the semantic relations the word sense in question can get into. For example, in the
case of “Ted played Bach” (Fass, 1991b, p.69) the inference rule is “artist for art form”, and starts
from the taxonomic relations between ARTIST and OCCUPATION, as found in the sense frame
of ‘Bach_as_artist’. Then a cell is encountered in which the syntagmatic relation
“OCCUPATION make ART FORM” is contained. Finally, from the sense frame of
47
‘Bach_music’ the information is accessed that ‘the supertype of Bach’ is MUSIC, and from the
sense frame of MUSIC we learn that its supertype is ART FORM. Therefore the rule applies by
traversing the syntagmatic and paradigmatic networks, and the result, the semantic class MUSIC,
is matched against the preference of ‘play’. If this is equal to or subsumed by the concept
expressed in the preference, the metonymic interpretation is successful. In this case, the
preference is equal to the supertype of ‘Bach_music’. This cell matching algorithm can be applied
recursively when several metonymies have been chained, in other words, where metonymies are
founding other metonymies (see section 2.4.4). If all metonymic rules fail (Fass implements four
rules: ARTIST FOR ART FORM, PART FOR WHOLE; PROPERTY FOR WHOLE;
CONTAINER FOR CONTENT and CO-AGENT FOR ACTIVITY), there is no traversal path
through cells and sense frames that links both word senses. The word sense is subsequently
checked for metaphoricity. For this purpose, a relevant analogy in terms of structural similarities
between the sense frames and cells are detected, much in the same manner as Wilks above. If
both metonymic and metaphorical inference rules fail, the expression is considered anomalous.
Pustejovsky (1991, 1995) proposes a generative model in which lexical entries contain a range of
representative aspects of lexical meaning at different levels. One of these levels, the qualia
structure, contains in complete agreement with its originator Aristotle, four roles or causes that
make an object to be what it is. These roles are the following: formal (ontological classification);
constitutive (what an object consists of, e.g. parts and material; telic (the functionality of the
entity) and agentive (what made the entity come into being). Explicit relations link semantic
aspects of the word in different roles. For example, the entry for crown, which can have the
metonymically related senses “traditional head ornament worn by the monarch” and “symbol for
the monarchy”, would look something like this:
FORMAL:
physical_object
CONSTITUTIVE
gold, jewel
48
TELIC
represents monarchy
AGENTIVE
artifact
For the purpose of underspecified semantic tagging (i.e. semantic labelling with general
concepts), Buitelaar (1998) made regular polysemic patterns in WordNet explicit in an automatic
fashion. His CoreLex database4 , used in the Core Lexic al Engine (Pustejovsky, 1995), contains
126 semantic types, covering 39,937 nouns in 317 systematic polysemous classes. These
semantic types consist of combinations of high level concepts from the WordNet hie rarchy that
subsume two or more senses of polysemous words. Two steps were taken to derive CoreLex from
WordNet. Firstly, all polysemous nouns in WordNet were reduced to a set of Basic Types,
corresponding largely to WordNet's 'unique beginners' and 'top nodes', such as artefact, causal
agent, shape and act. Subsequently, systematic groupings of nouns were created on the basis of
their Basic Types distributions. For example, the noun banana, occurring both in a food and a
plant sense, was put in a group with other nouns exhibiting the same pattern, such as coriander,
grapefruit, plantain and mulberry. A list with the Corelex basic types can be found in Appendix
A.
In addition to the formulation of lexical rules that capture the formation of a derived sense from a
given base sense, Copestake and Briscoe (1996) and Copestake (1995) propose adding to lexical
entries conditional probabilities that reflect how likely a word is to be used in a specific sense.
Thus, although a lexical rule will generate a derived sense from a base sense, different
frequencies may be associated with the two senses, and for some word forms (e.g. teak) the
derived sense will have a higher frequency than the base sense. Under such proposals, novel
usages of a word form can be derived through productive application of a lexical rule but the NLP
system will have a measure reflecting sense usage. Established word senses will be associated
with high frequencies, while non-established senses will have low frequencies. Frequency
4
Available from http://www.cs.brandeis.edu/~paulb/CoreLex/overview.html
49
information could be used to guide parsing, through preference of high-frequency senses.
Ambiguous word forms would be parsed initially using the high-frequency sense; low-frequency
senses should only be chosen in the case of a syntactic conflict or as a result of subsequent
pragmatic processing which determines that it is the ‘correct’ sense.
The objective of the ACE program (see http://www.itl.nist.gov/iad/894.01/tests/ace/) is to develop
automatic content extraction technology to support automatic processing of natural text. The ACE
research objectives are viewed as the detection and characterization of Entities, Relations and
Events. For this purpose a 100,000 word corpus has been built and annotated with the following
types:
1. Person. Person entities are limited to humans. A person may be a single individual or a
group if the group has a group identity.
2. Organization. Organization entities are limited to corporations, agencies, governments,
and other groups of people defined by an established organizational structure.
3. Facility.
Facility entities are limited to buildings and other permanent man-made
structure and real estate improvements.
4. GSP (A Geographical-Social-Political Entity). GSP entities are geographical regions
defined by political and/or social groups.
A GSP entity subsumes and does not
distinguish between a region, its government or its people.
5. Location. Location entities are limited to geographic entities such as geographical areas
and landmasses, bodies of water, and geological formations.
Metonymies involving any of the types above are annotated by means of a combination of classes
or a role specification in the form of any of the semantic classes in the case of GSPs.
This corpus will provide useful information for developers of algorithms that cover the automatic
detection and resolution of metonymies.
50
Markert and Nissim (2002) developed an annotation scheme for location metonymies covering
eight patterns amongst which are place for people (e.g. Liverpool in “Liverpool celebrates its
election as cultural capital”), place for event (e.g. Kyoto in “Kyoto didn’t go far enough”) and
place for product (e.g. Bordeaux as wine). They have created a corpus that covers 5000
occurrences of annotated location expressions. The corpus, sampled from the BNC, covers many
domains and genres. Annotation was performed using general semantic classes (organizations,
persons, locations, and, to a lesser extent, animals) rather than individual word senses. This
division into annotation schemes by semantic classes is motivated by the fact that words
belonging to one semantic class undergo similar meaning shifts, improves inter-annotator
agreement, and allows generalizations over the data set. The corpus then in turn serves as a basis
for their metonymy resolution algorithms. They produced a prototype machine learning algorithm
with a 50% accuracy for metonymy recognition, using a feature space consisting of collocations
and syntactic subject/object. This algorithm operates on the semantic classes and is able to infer
readings for unknown words providing these belong to any of the semantic classes identified in
the training data. The difference with previous algorithms lies in the fact that the latter rely on
selectional restrictions and syntactic constraints for metonymy recognition (Fass, 1997, Wilks,
1975). Such approaches fail when there is no intrasentential anomaly, as in “I like Shakespeare”
for “I like the works of Shakespeare”. Markert and Hahn (2002) found no direct violation of
semantic restrictions in 29% of 106 metonymies from 27 German magazine texts.
3.2 The Need for Data
Many of the working systems only have a limited number of lexical entries at their disposition for
the recognition and interpretation of figurative language. Lack of availability of a large number of
patterns and participating words is the main cause of the sparse evaluation of most NLP
algorithms dealing with figurative language. Some of them are discussed in comparison to
51
constructed examples only (e.g. Pustejovsky, 1995, Fass, 1997) while others used naturally
occurring language annotated according to subjective intuitions of one individual only (e.g.
Markert and Hahn, 2002, Stallard, 1993, Verspoor 1997, Harabagiu, 1998). Large-scale data
driven analysis of these phenomena on the basis of large amounts of available linguistic material
is rare (with the exception of Buitelaar 1998, see section 5.2.1). The corpora developed in ACE
and by Markert and Nissim (2002) provide a valuable first step, but are still rather small-scale.
One reason for the restricted number of attested instances of metonymy and metaphor in texts and
other resources is that their reliable discovery is a time consuming, mostly manual task.
Violations of selectional restrictions, which may indicate the presence of a metonymic sense for a
word that occurs as a verbal argument, can be identified on the basis of the automatic acquisition
of selectional preferences (Resnik, 1997), with a reported success rate of 50%, combined with the
semantic disambiguation of content words. If the resulting success rate is high enough, this could
form the basis of an automatic approach alternative to the method described in this thesis.
Recent initiatives towards the creation of larger scale metonymic typology cover mainly the
creation of corpus material. Corpus-oriented research into metonymy relies on the annotation of a
substantial amount of texts with metonymic patterns. This has been tackled for a limited set of
metonymic patterns in projects such as ACE, where the range of metonymy is restricted to
alternations between persons, facilities, locations and organizations. Markert and Nissim (2002)
have created a corpus annotation is restricted to location names. Nevertheless, the creation of
these corpora embodies a first step towards a shift in the approach to metonymy from an intuitive,
introspective, scholarly, intellectually top-down collection of patterns, and associated recognition
and resolution techniques on the basis of a limited set of metonymic regularities, to a data-driven,
bottom-up perspective. This involves, on the one hand, recognition of new unattested instances of
existing patterns, but also the identification of implicit regular polysemic regularities in existing
lexical resources, which may coincide with or extend existing lists. Resource derived regularities
52
and generalizations can then assist the corpus annotation task by offering the annotator the
various metonymic possibilities for a given word.
3.3 Lexical Knowledge Bases
Lexical knowledge bases (LKBs) such as machine-readable dictionaries and thesauri are far from
perfect in terms of the semantic coverage they explicitly provide. They lack completeness in the
sense that they only capture lexicographically attested sense distinctions and regular
combinations of senses. Even conventional regular polysemic sense patterns are not always
included in LKBs. For instance, ‘France’ has only the country sense in WordNet (see chapter 4),
whereas ‘United States’ has the additional metonymic sense ‘government of the United States’.
In general, dictionaries are not primarily directed towards encoding semantic regularities across
the lexicon, but towards distinguishing senses that enable the user to interpret texts. WordNet, the
lexical resource under examination in this thesis, forms no exception because its sense
distinctions are dictionary-based.
Electronic lexicons, however, contain far more information than the organization of their data
suggests. Most lexical resources such as machine-readable dictionaries only provide a list of
senses, and sometimes a coarse-grained classification of senses. For instance, in the case of the
Longman Dictionary of Contemporary English (Procter, 1979) and the Cambridge International
Dictionary of English (CIDE, 1999), sense distinctions are classified according to a limited set of
high level ontological classes (e.g. animal), and domain topics (e.g. air travel). The definitions
and glosses associated with the senses distinguished in these and other resources such as
WordNet provide a wealth of implicit lexical and encyclopaedic information. This information, if
explicitly formulated, can be exploited by a computer for the purpose of natural language
understanding (Harabagiu, 2000).
53
Systematic relatedness between senses is one type of knowledge that is mostly left implicit in
resources. Overall, senses are presented in the form of lists, with sometimes a certain level of
nesting that implies further specification of an originally coarse-grained sense distinction.
The increasing availability of lexical knowledge bases that cover different languages offers new
possibilities for computational approaches to lexical structure. It is now possible to observe
regularities between senses across words and languages, and begin to form a picture of the
conceptual mechanisms involved in the process of sense distinction and the assessment of
conceptual relatedness of word senses. Before discussing our particular approach to regular
polysemy in chapter five, we present an overview of thesauri in the next chapter.
54
Chapter 4
Thesauri
The data resources under examination in this thesis are a type of thesaurus. This chapter describes
the nature and content of different types of thesauri.
The name `thesaurus' covers a range of language resources that are useful for a range of different
language engineering purposes. We work from an inclusive definition of a thesaurus: ``a resource
in which words with similar meanings are grouped together''. The varieties include at least the
following:
1
Roget, Macquarie and others, produced, as books, to help writers with word selection
2
Thesauri produced manually for use in information retrieval systems
3
`Automatic thesauri', produced by processing corpora, with similarity between words
measured (directly or indirectly) by co-occurrence.
4
WordNet and EuroWordNet
There is of course a vast literature on the use of thesauri in computational linguistics, stretching
back to the earliest days of the enterprise when Roget was hand-punched onto cards and the links
used for a disambiguation engine (Masterman, 1957), cited in (Wilks et al., 1996, p. 89) and the
extensive work of the Sedelows (Sedelow, 1992). Here the references to the literature will be
indicative.
4.1 Roget-type Thesauri
In 1852, Roget published his first version of his thesaurus. Roget’s thesaurus is organized by
categories of thinking, and arranges words and phrases according to their meanings. Words with
55
related meanings are grouped together in separate entries. The latest version of Roget's Thesaurus
contains 1043 entries. These are hierarchically ordered in terms of generality. For instance, the
first two main classes are listed below, together with their subdivisions:
1 Class I. Words Expressing Abstract Relations
1.1 Section I. Existence
1.2 Section II. Relation
1.3 Section III. Quantity
1.4 Section IV. Order
1.5 Section V. Number
1.6 Section VI. Time
1.7 Section VII. Change
1.8 Section VIII. Causation
2 Class II. Words Relating to Space
2.1 Section I. Space in General
2.2 Section II. Dimensions
2.3 Section III. Form
2.4 Section IV. Motion
Zooming in on the class 2.4 section 4 ‘Motion’ gives us a further subdivision:
1. Motion in General
264. Motion
265. Quiescence
266. Journey
267. Navigation
268. Traveler
269. Mariner
270. Transference
271. Carrier
272. Vehicle
273. Ship
2. Degrees of Motion
274. Velocity
275. Slowness
Each class contains a number of words that bear a certain similarity to each other. For instance,
‘Journey’ (class 266) contains, amongst others, the following:
journey, excursion, expedition, tour, trip, grand tour, circuit, peregrination, discursion, ramble,
pilgrimage, hajj, trek, course, ambulation, march, walk, promenade, constitutional, stroll,
saunter, tramp, jog trot, turn, stalk, perambulation
56
Landau comments on what he calls the extreme inclusiveness of thesauri:
Rarely used words, non-English words, names, obsolete and unidiomatic
expressions, phrases: all are thrown in together along with common
words without any apparent principle of selection. For example, in
the fourth edition of Roget’s International Thesaurus - one of the
best of the conceptually arranged works - we find included under the
subheading orator: Demosthenes, Cicero, Franklin D.Roosevelt, Winston
Churchill, William Jennings Bryan. Why not Pericles and Billy Graham?
When one starts to include types of things, where does one stop? There
is actually a list of insects (paragraph 414.36), which is even more
of a random sampling than that of orators. Such works are a potpourri
of everything the compiler can think of. (Landau, 1989, p.108)
The market for Roget-style thesauri is distinct from that for dictionaries. They are marketed as
aids to help writers choose the appropriate word, and for this the critical consideration is to
provide a wide range of possibilities. This is quite unlike the native-speaker dictionary market,
where the main purposes are to help with finding meanings for rare words, finding correct
spellings, and as an arbiter for word games and family disputes (Summers, 1988), which means
that the penalty for the sins sketched by Landau is not great.
We may also classify under this heading some items produced for the language learners' market.
McArthur's Lexicon (Macquarie, 1986) is, like Roget, organised hierarchically by meaning but
restrains itself to a modest 15,000 items, and for each of these it provides a definition and
sometimes an example. It is thus quite dictionary-like, and takes a dictionary-like approach to
sense division. It frequently has several numbered senses of a word under a thesaural heading,
with, for example, twenty-five meanings for turn under the thesaural heading ‘N297: twisting and
turning’. The Longman Activator (Summers, 1995) complements it: whereas McArthur first
addresses “real world” vocabulary, where the difference between two close items corresponds to
their different non-linguistic denotations, the Activator leaves real world vocabulary out entirely,
focusing instead on parts of the vocabulary, more often verbal and adjectival, where differences
be-
57
tween near-synonyms are usually better explained by differences in circumstances and patterns of
use than denotations. Many Activator headwords are phrasal, and the analyses of meanings are
designed to help learners select the appropriate English word or phrase from the set that the
Activator offers. Like McArthur's Lexicon, the Activator is hierarchically organised and provides
definitions and examples.
4.2 Manually Created Thesauri
In many specialist areas where information retrieval systems are widely used, domain-specific
thesauri have been developed. As with WordNet, the organising principles are synonymy and
taxonomy, which make it possible for searches to be broadened or narrowed, and for searches to
be matched against documents using synonyms of the search terms. Baeza-Yates et al. (1999)
present the basic relations for IR-manual thesauri as BT (broader term), NT (narrower term) and
RT (related term). IR-manual thesauri will often also use semantic relations of particular salience
in the domain, for example medical thesauri may include relations such as ‘located’, ‘prevents’
and ‘diagnoses’. A resource such as the Unified Medical Language System (UMLS) is a highly
sophisticated object incorporating a very large quantity of medical knowledge and supporting
inference of various kinds (EAGLES, 1999).
IR-manual thesauri are domain-specific, and are thereby not the core concern of this thesis. The
key difference between WordNet and IR-thesauri is, arguably, that WordNet addresses general
language.
4.3 Automatically Created Thesauri
There is a substantial body of work on the automatic generation of thesauri and related resources
from large corpora. Some of this work takes place under the heading of NLP, and some under the
heading of Information Retrieval (e.g. Sanderson and Croft, 1993). The simplest strategy for
58
automatic thesaurus generation is to look for collocational similarities between words in a text
corpus:
For each content word in the corpus
for each other content word,
find how often both occur within k words (or characters) of each other.
If there are n content words in the corpus, each word can then be represented by a vector of length
n; similarities between vectors can be computed using any of a variety of similarity measures, and
for each word we can identify the most similar words.
Schütze's ‘word space’ (Schütze, 1998; see also section 2.2.1) is defined in this way. He uses a
mathematical technique, singular value decomposition (SVD), to reduce the dimensionality of the
space. Latent Semantic Indexing (LSI) (Deerwester, 1990) also uses SVD, but applies it to a
matrix of counts of words in documents, so words label the rows of the matrix and documents
label the columns (or vice versa). (Hindle, 1990), (Lin, 1998), (Grefenstette, 1994a),
(Grefenstette, 1994b) all find and count triples of ‘grammatical-relation - word1 - word2’, rather
than simple unordered word co-occurrences. Grefenstette (1994a) distinguishes first, second and
third order affinities between words. First order affinities are between words that co-occur with
each other. Second-order affinities are between the words that co-occur with the same words.
Thus words with complementary distributions, such as tumor and tumour, have no first-order
affinity but a marked second-order one, since documents will either contain tumor or tumour but
not both, yet the two will occur in the same kinds of contexts. Spelling variants will be an
extreme case of words having a strong second-order affinity yet no first-order affinity. Third
order affinities relate to distinct senses of polysemous words: we would like to identify that bank,
in one sense, has an affinity with river, and in another, an affinity with business.
Sparck-Jones (1964) applied this notion of vector based similarity in order to find synonymous
and near-synonymous words in Roget’s thesaurus. She defined semantic similarity as the size of
59
the intersection of the semantic classes both words belong to divided by the union of all semantic
classes containing any of the two words. In the most extreme case, fully fledged synonymy would
be exemplified by two words occurring in exactly the same semantic classes in Roget.
4.4 WordNet and EuroWordNet
WordNet (Fellbaum,1998) is a lexical database produced on psycholinguistic principles. It has
been developed at Princeton University and, for ten years now, has been freely available.
WordNet is an English-language resource and contains information about nouns, verbs, adje ctives
and adverbs in English and is organized around the notion of a synset. A synset is a set of words
with the same part-of-speech that can be interchanged in a certain context. For example, {car;
auto; automobile; machine; motorcar} form a synset because they can be used to refer to the
same concept. A synset is often further described by a gloss: "4-wheeled; usually propelled by an
internal combustion engine". Finally, synsets can be related to each other by semantic relations,
such as hyponymy (between specific and more general concepts), meronymy (between parts and
wholes), cause, etc. as is illustrated in Figure 4.1.
{conveyance; transport}
hyperonym
{vehicle}
{bumper}
hyperonym
{motor vehicle; automotive vehicle}
{car door}
hyperonym
meronym
{car; auto; automobile; machine; motorcar}
meronym
hyperonym
{hinge; flexible joint}
meronym
{doorlock}
meronym
{car window}
{armrest}
{car mirror}
hyperonym
{cruiser; squad car; patrol car; police car; prowl car}
{cab; taxi; hack; taxicab; }
Figure 4.1: synonymy, hyponymy and meronymy relations in WordNet
60
In this example, taken from WordNet1.5, the synset {car; auto; automobile; machine; motorcar}
is related to:
•
a more general concept or the hypernym synset: {motor vehicle; automotive vehicle},
•
more specific concepts or hyponym synsets: e.g. {cruiser; squad car; patrol car; police car;
prowl car} and {cab; taxi; hack; taxicab},
•
parts it is composed of: e.g. {bumper}; {car door}, {car mirror} and {car window}.
Each of these synsets is again related to other synsets as is illustrated for {motor vehicle;
automotive vehicle} that is related to {vehicle}, and {car door} that is related to other parts:
{hinge; flexible joint}, {armrest}, {doorlock}. By means of these and other semantic/conceptual
relations, all word meanings in a language can be interconnected, constituting a huge network or
wordnet. Such a wordnet can be used for making semantic inferences (e.g. in answer to the
question “what things can be used as vehicles?”), for finding alternative expressions or wordings
(“what words can refer to vehicles?”), or for simply expanding words to sets of semantically
related or close words, in e.g. information retrieval. Furthermore, semantic networks give information on the lexicalization patterns of languages, on the conceptual density of areas of the
vocabulary and on the distribution of semantic distinctions or relations over different areas of the
vocabulary. In Fellbaum (1998) a detailed description is given of the history, background and
characteristics of the Princeton WordNet.
EuroWordNet is a multilingual lexical database with wordnets for several European languages,
which are structured along the same lines as the Princeton WordNet (Fellbaum 1998). Each
language module represents an autonomous and unique language-specific system of la nguageinternal relations between synsets. Equivalence relations between the synsets in different
languages and WordNet1.5 are made explicit in the so-called Inter-Lingual-Index (ILI). Each
synset in the monolingual wordnets has at least one equivalence relation with a record in this ILI,
either directly or indirectly via other related synsets. Language-specific synsets linked to the same
61
ILI-record should thus be equivalent across the languages, as is illustrated in Figure 4.2 for the
language-specific synsets linked to the ILI-record drive by means of an equivalence relation (type
II). Figure 4.2 further gives a schematic presentation of the different modules and their interrelations. In the middle, the language-external modules are given: the ILI, a Domain Ontology
and a Top Concept Ontology. The ILI consists of a list of so-called ILI-records (ILIRs) which are
related to word-meanings in the language-internal modules, (possibly) to one or more Top
Concepts and (possibly) to domains. The language-internal modules then consist of a lexicalitem-table indexed to a set of word-meanings, between which the language-internal relations are
expressed (type III). There are many semantic relations defined in EWN. The main ones and the
ones that are of interest are listed in table 4.1.
move
travel
go
English
Wordnet
Traffic
III
ride
III
Air
Traffic
2ndOrderEnt
Road
Traffic
Locatio
Dynami
drive
II
I
EN Lexical Items
ES Lexical Items
II
cabalgar
jinetear
bewegen
reizen
gaan
Domain-Ontology Top-Ontology
I
I
III
rijden
berijden
III
II
NE Lexical Items
IT Lexical Items
ILI-record
{drive}
II
III
III
conducir
Inter-Lingual-Index
guidare
cavalcare
III
III
Spanish
Wordnet
mover
transita
r
Dutch
Wordnet
I = Language Independent
link
II = Link from Language Specific to
Inter Lingual
Index
III = Language Dependent
Link
andare
muoversi
Italian
Wordnet
Figure 4.2: Architecture of the EuroWordNet Database
62
Wordnets are thesaurus-like rather than dictionary-like in that their principle mode of
organisation is the synset. They guard against Landau's complaint by only allowing specified
semantic relations between word-meanings and synsets.
Relation
Example
antonymy
synonymy
foe - friend
jester - fool
hyponymy
hypernymy
vertebrate - bird
parrot - bird
holonymy
meronymy
attribution
pertainymy
member/part/substance
member/part/substance
door -wall
bread - flour
size - large
pole – polar
Table 4.1: Main Semantic Relations for Nouns
4.5 Thesaurus Word Senses and Dictionary Word Senses
At one level, the difference between a dictionary and a Roget-type thesaurus is one of indexing:
the dictionary is organised alphabetically, the thesaurus by meaning or word group. If this were
the only difference, a computational environment that offered both indexing possibilities would
remove the distinction, and a resource such as WordNet, which offers both options, would be
equally dictionary and thesaurus. But there are further differences. Firstly, most published
dictionaries give only limited space to word clusters. Most Roget-type thesauri do not include
definitions, and group words according to implicit rather than explicit semantic categorizations,
so the information for reading a resource either way is absent. Secondly, most existing resources
have been developed from the one perspective or the other, but not both. When a lexicographer is
producing a dictionary entry, the goal is to provide a coherent analysis that separates out the
distinct meanings and patterns of use the word has, with each part of the entry making sense in
63
relation to the others. When he is producing a thesaurus entry (at least for paper publication), the
unit that must appear coherent is the thesaurus entry or word group. Thus where a word has two
distinct but closely-related meanings, but the distinction is not salient for other words and the
senses both fall in the same thesaurus category, the compiler will not present the word twice in
the same thesaurus entry. Without definitions justifying the different senses, the presentation of
the same word twice would be confusing to the user. So a dictionary distinction may be lost in the
thesaurus. Conversely, a single dictionary meaning is commonly found in more than one section
of the thesaurus. Consider the word listless in the Macquarie Dictionary (Macquarie, 1997) and
Macquarie Thesaurus. In the dictionary, listless has two definitions:
1. feeling no inclination toward or interest in anything.
2. characterised by or indicating such feeling: a listless mood.
The difference is between the adjective describing persons (``they all seemed quite listless'') and
the adjective applied to certain other nouns (``a listless mood'', ``a listless wave of the hand'').
In the Thesaurus, listless appears in three places, within the sections headed BOREDOM,
IDLENESS and APATHY. No one would claim that these three nouns are strictly synonymous,
but they are close, and a user will be happy to find listless in any of the three sections. In none of
these three sections is there any attempt to differentiate two senses of listless, one applying to
persons, the other applying to other nouns. A comparable extension of other adjectives, as in “a
tired atmosphere”, “an idle moment”, “a bored glance”, is assumed in various parts of the
Thesaurus. So the inclusiveness of a thesaurus allows listless to be entered under different
semantic headings that are not specified in different dictionary senses of the word, while the two
dictionary definitions are not distinguished at all in the thesaurus.
64
Wordnet, with its word senses and hierarchical organization, is a popular resource for
computational linguistic purposes. In the next chapters we will concentrate on this resource and
investigate its usefulness for the detection and characterization of regular polysemy.
65
Chapter 5
Regular Polysemy Patterns in WordNet
This chapter concentrates on the occurrence of regular polysemic patterns in WordNet. After a
description of the ways in which regular polysemy is presently addressed in WordNet, and a
discussion of related research, I describe an algorithm to automatically extract RP patterns from
the WordNet hierarchical structure, and illustrate its result with a number of examples.
5.1 The Occurrence of Attested Regular Polysemic Patterns
As we saw in the previous chapter, regular polysemy (RP) has mostly been approached from a
theoretical perspective. There is a limited set of default relations identified in the literature of
which the following are a subset (cf. Ostler and Atkins, 1991, Pustejovsky, 1995; see also section
2.7.4). Most of these relations have been identified by examination of a limited quantity of
linguistic material (texts, dictionaries) or introspection. The list is by no means exhaustive. For
comparison purposes, I manually mapped each RP pattern in table 5.1 below onto the WordNet
hierarchy. The nodes that correspond with the pair members below subsume words that in the
ideal case have senses that correspond completely with the RP type in question. It is possible that
other words can be found that are subsumed by different synsets. It is also possible that the
WordNet node pairs subsume meaning relations between specific word senses that do not
correspond to the RP pattern.
RP pattern
example word
1. container/containerful cup
2. animal/food
lamb, chicken
3. animal/animal skin
crocodile
4. plant/food
banana
5. product/producer
newspaper, Honda
6. substance/colour
jade, amber
7. object/shape
pyramid
8. language/people
Spanish
in WN
Y
Y
Y
Y
Y
Y
Y
Y
words
19
172
15
393
3
33
101
257
66
9. music/dance
10. figure/ground
11. place/people
12. organization/building
13. action/result
waltz
door/window
city, New York
school
examination
Y
N
Y
Y
N
16
7
10
-
Table 5.1 Regular Polysemic Patterns
The matching of reported patterns onto WordNet synsets, however interesting, has a limited
scope. It is not possible to scale up the coverage, and basically this is just an evaluation of
WordNet to see whether it covers the patterns or not. On examination, it appears that WordNet
does pass the test, because it covers 85% of the elements of the list above. The only patterns from
the list above it cannot accommodate are figure/ground and action/result. The former is
exemplified by window, which can be the generally wooden framework (figure) or the opening in
the wall into which the frame has been fitted (ground). It is a very general pattern that subsumes
combinations of WordNet synsets such as structure/opening (containing window) and
structure/passage (containing door). The second pattern, action/result, is difficult to translate into
hierarchical terms, since all kinds of different ontological types can be the result of a particular
action or process. Result is more a functional node than a taxonomical node, and although,
admittedly, WordNet does have some of these integrated in its hierarchies, the taxonomic nodes
are in the majority. Therefore, one must attempt to describe functional nodes in terms of
taxonomic ones. For instance, one of the WordNet pairs possibly subsumed by the pattern
action/result is action/artifact.
And finally, the limited coverage of the pattern product/producer can be explained by the fact
that WordNet contains only one proper name with the desired pattern (Ford).
67
5.2 The Automatic Extraction of Regular Polysemy Patterns
from WordNet
The successful automatic extraction of candidate patterns from WordNet will enable a largerscale investigation of the phenomenon of regular polysemy in the English language. It needs to be
borne in mind, however, that this investigation is entirely resource dependent. Any extracted
patterns are the result of the lexicographic practices applied by the WordNet creators. The
analysis assumes that the sense distinctions reflecting valid RP patterns have been established in
WordNet.
As said above, in general, dictionaries are not primarily directed towards encoding semantic
regularities across the lexicon, but towards distinguishing senses that enable the user to interpret
texts. WordNet with its dictionary-based sense distinctions forms no exception; the semantic
encoding of WordNet is not aimed at the implementation of explicit patterns of regular polysemy.
However, a number of semantic regularities have been manually captured in WordNet by means
of the so-called ‘cousin’ relation, which is used to group concepts according to similarity or
relatedness of meaning. The relation is computed from a table of 88 concept pairs in WordNet1.5,
and 220 in WordNet1.6. All the hyponymic concepts of these concept pairs, i.e. all concepts that
occur below them in the hierarchy, are assumed to be involved in the semantic relation.
Exceptions are listed in a separate file. Although several regularities have been observed, there is
no explicit description of the semantic relations covered by the concept pairs. Close inspection
shows that these pairs implicitly signal a semantic relation that is not always an instance of
regular polysemy. The list is tentative, incomplete and does not offer a consistent and structured
account of recurrent patterns between sets of words. The cousin relation does, however, capture a
number of valid RP patterns, such as:
organization (a group of people who work together)
building (a structure that has a roof and walls and stands more or less permanently in one place)
68
person (a human being)
language (a systematic means of communicating by the use of sounds or conventional symbols)
container (something that holds things, especially for transport or storage )
containerful (the quantity that a container will hold)
For the concept pairs listed above, a large number of words occur both as hyponyms of the first
node and the second node. However, the cousin relation is not primarily geared towards capturing
regular polysemy. The members of some patterns only subsume one particular word form, which,
according to Apresjan’s criterion, does not constitute regular polysemy. For example, the cousin
pair food-1, nutrient (any substance that can be metabolized by an organism to give energy and
build tissue) and tableware-1 (articles for use at the table) have only the word dish in common.
Since the cousin relation applies to all words found in the subsumed synsets, the semantic relation
here is predominantly between different words. It seems that regular polysemy patterns are often
more a by-product of the creation of cousins. Cousin relations, therefore, do not necessarily
generate regular polysemous patterns, but sometimes capture semantic relations between different
words of a more schema-like nature. Within the scope of the present research we are only
interested in sense distinctions of individual words and can only take those cousin relations into
account that generate clusters that share word forms.
5.2.1 Corelex
As described in section 3.1, the Corelex classes consist of systematic combinations of high-level
hypernyms and therefore yield classes that are semantically highly underspecified. Because of the
lack of specificity of the concepts involved, their combination is less likely to produce regular
polysemic patterns that either capture a particular meaningful semantic relation between the
69
subsumed word senses, or conflates a group of more specific patterns that are lost in the bulk of
participating words. For example, we find bundle, package, packet, ragbag, deck, edition, library,
menagerie, repertory belonging to the same CoreLex type (arg, a combination of the Basic Types
artifact and group) where we find the first four words covered by the more specific hypernymic
nodes collection-1 and container-1 and the last three by collection-1 and facility-1. The
examination of the regular polysemous classes showed a number of disadvantages of CoreLex for
our purposes. Firstly, 19 of them consist of only one Basic Type and therefore do not display
regular polysemy. Secondly, the Basic Types are largely based on very high-level nodes in the
WordNet hierarchy. Thirdly, the generated classes are not always homogeneous in nature;
particularly the larger groups do not necessarily exhibit regular polysemic patterns, amd
occurrences of false friends are not infrequent. A more detailed examination of the artifact-group
set (both concepts are WordNet1.5 unique beginners, i.e. the top level nodes of individual
branches within the WordNet taxonomy) yielded a.o. the following results:
artefact-1
a man-made object
group-1
any number of entities (members) considered as a unit
institution-2 a building or complex of buildings where an organization for the promotion of some
cause is situated
institution-1 an organization founded for a specific purpose
guard-3 a device designed to prevent injury
guard-5 a group of men who escort and protect some important person
menagerie -2 the facility where wild animals are housed for exhibition
menagerie -1 a collection of live animals for study or display
70
shower-1 a plumbing fixture that sprays water over you
shower-5 a party held by friends to present gifts to a person
type -6 a small block of metal bearing a raised character on one end; produces a printed character
when inked and pressed on paper; “he dropped a case of type so they made him pick them up”
type -3 (biol) the taxonomic group whose characteristics are used to define the next higher taxon
unit-6 a combination of interrelated interacting elements designed to work as a coherent entity
unit-2 an organization regarded as part of a larger social group
The above examples show some disadvantages of restricting regular polysemy to combinations of
high level concepts. It leads to inappropriate instantiations of a pattern, as in the case of shower
and type, where there is no systematic relation between the two senses. Whereas there seems to be
no meaning relation between the various senses of type, the relation between the two senses of
shower is an isolated case of metaphorical transfer (see also section 6.2.1). It also leads to the
creation of groups that contain words that are not semantically similar. For instance, menagerie
and guard seem to display a ‘facility/collection’ and a ‘group/device’ alternation respectively.
Using high level concepts for the characterization of regular polysemy can block the
identification of subgroups that are semantically more coherent.
5.2.2 Extraction of More Specific Patterns from WordNet
In order to obtain semantically more specific and coherent groups a new approach was taken. The
developed technique (Peters, 2000) looks for patterns at any level of specificity, i.e. at higher as
well as lower levels in the hierarchies. It identifies sense combinations in WordNet, where the
71
senses involved display a potential regular polysemic relation, i.e. where the senses involved are
candidates for systematic relatedness.
In order to obtain these candidate patterns WordNet has been automatically analysed by
exploiting its hierarchical structure for nouns. Wherever there are two or more nouns with senses
in one part of the hierarchy, which also have senses in another part of the hierarchy, then we have
a candidate pattern of regular polysemy. The patterns are candidates because there seems to be an
observed regularity for two or more words. The methodology followed was an operalization of
Apresjan’s definition of regular polysemy (described in section 2.7.2) in order to select candidate
RP patterns. The technique exploits WordNet’s hierarchical structure (Peters, 2000): wherever
there are two or more words with one sense in one part of the hierarchy which also have a sense
in another part of the hierarchy, then we have a candidate pattern of regular polysemy. Since the
noun hierarchy is the most structured taxonomy in WordNet the method is restricted to nouns.
An example of the result of this query over WordNet can be found in figure 5.1 on the next page.
The words architecture, law, literature, politics and theology all have a sense subsumed by the
synset {profession} and another sense under {discipline}. The location of these concepts in the
hierarchies can be deduced from their hypernyms that are indicated by arrows.
Words do not necessarily occur within synsets at the same level in the hierarchy, as is the case
with the words in figures 5.2 and 5.3 below. These are graphical representations of the regular
polysemic pattern music - dance and passage - structure respectively. Although all words occur
as direct hypernyms of these hypernymic pairs, the senses of words like bolero and arch do not
occupy the same level in the two WordNet hierarchies. In other words, the senses of the
participating words can occur at any hierarchical level below each member of the hypernymic
pair.
67
psychological feature
act
cognition
activity
content
occupation
knowledge domain
profession
(the principal activity
in your life)
discipline
(a branch of knowledge)
architecture
law
literature
politics
theology
hypernym combination
words whose senses occur under both hypernyms
Figure 5.1: Words in WordNet Covered by the Pattern ‘Profession/Discipline’
5.2.3 Size of the Resulting Set of Patterns
The data set that results from the automatic extraction process described above is very large:
320000 word sense pairs occur as hyponyms of 57177 combinations of hypernyms. In total there
are 8347 nouns involved with 21885 senses. Many of these word sense pairs are subsumed by
more than one hypernym pair that functions as the characterization of the candidate regular
polysemy pattern. This is caused by the fact that the hypernyms further up both hierarchical
chains combine with each other into distinct hypernym pairs. The result is therefore the Cartesian
product of all hypernyms in chain one and all hypernyms in chain two. For instance, apart from
the music - dance pattern in figure 5.2, other possible combinations are, for instance, music social dancing and music-ballroom dancing.
Another example concerns patterns associated with the word university:
sense 2: “where a seat of higher learning is housed, including administrative and living quarters
as well as facilities for research and teaching”
68
sense 3: “a large and diverse institution of higher education created to educate for life and for a
profession and to grant degrees”
A subset of these patterns is listed in table 5.2, where we see the combination of the hypernymic
node structure (at level 4 in the hierarchy, indicated in column 2) with several other hypernyms at
various levels within the same hierarchy (see column 6).
hyper1
level hyper1_word hyper1_gloss
hyper2
level hyper2_word
hyper2_gloss
103431817
4 structure
A thing
constructed
105962976
2 social group
people sharing
some social
relation
103431817
4 structure
A thing
constructed
105997592
3 organization
a group of people
who work together
103431817
4 structure
A thing
constructed
106002286
4 institution
103431817
4 structure
A thing
constructed
106144776
5 educational
institution
an organization
founded and
united for a
specific purpose
an institution
dedicated to
education
Table 5.2: Candidate RP Patterns for ‘university’
69
Unique Beginner 1
Unique Beginner 2
Artefact
Action
Music
Landler 1
Dance
Country dancing 1
Rumba 2
Bolero 1
Social dancing 1
Waltz 1
Swing 9
Gavotte 2
Gavotte 1
Swing 8
Folk dancing 1
Ballroom dancing 1
Folk music
Landler 2
Rumba 3
Bolero 3
Polka 1
Polka 2
Waltz 2
Figure 5.2: The Regular Polysemic Pattern music – dance
structure
passage
access
area
passage way
arcade 2
arch 4
entrance
corridor
arcade 1
room
arch 3
gallery 1
gallery 7
Figure 5.3: the regular polysemic pattern passage - structure
70
Chapter 6
Reduction of the Extracted Data Set on the Basis of
Resource-internal Criteria
This chapter describes the analysis, automatic reduction and evaluation of the data set that results
from the automatic extraction process described in the previous chapter. The analysis covers a
manual evaluation, the creation of a gold standard set of regular polysemic patterns, the
application of six criteria to reduce the large set of candidate regular polysemic patterns, and an
evaluation of this reduction process.
6.1 Size and Nature of the Extracted Data Set
The set of 57177 candidate regular polysemic patterns that results from the extraction process
described in the previous chapter is wholly dependent on WordNet’s taxonomic structure. It
cannot be regarded as useful as it is, for several reasons. Firstly, this set is too big. It is to be
expected that it contains many invalid candidates and pairs that are too general to yield
semantically distinguishable and coherent patterns (as is the case with Corelex, see section 5.2.1).
It is our aim to extract a set of maximally valid regular polysemic classes, not a set reflecting
every potential semantic regularity in the distribution of WordNet sense distinctions.
Secondly, WordNet’s taxonomic structure reflects lexicographers’ consensual vie wpoints on the
distinctions between word senses and conceptual classes. In some cases, the combination of two
hypernymic nodes from different taxonomic branches into a candidate regular polysemic pattern
highlight structural shortcomings in WordNet’s taxonomy, where the difference between
taxonomic and functional concepts is blurred. For instance, two senses of a word such as knife are
subsumed by functional concepts lexicalized by tool and weapon. Contrarily, the two senses of
hutch are subsumed by a taxonomic (structure) and a functional concept (enclosure). The lack of
71
discrimination between these two types of hypernymic concepts is illustrative of the ontological
confusion in WordNet, but, in any case, these different types of concepts reflect different aspects
of the same gestalt meaning (Koch, 1999).
In the next phase, a random evaluation of the set of candidate regular polysemic patterns is
performed, which gives an indication of the semantic nature and coverage of the patterns. The
evaluation is performed in three stages: In the first stage, manual evaluation results in a first
characterization of the data set, and the creation of a gold standard set of patterns for automatic
evaluation purposes. In the second stage, six criteria are applied to the set in order to
automatically filter out as many unsuitable candidates as possible, and retain a maximum
coverage of the gold standard set. Finally, the accuracy of the method is evaluated by means of an
examination of a random set of patterns.
6.2 Manual Evaluation in Order to Create the Gold Standard
A manual evaluation was performed along the lines of two semantic criteria in order to determine
the systematicity of the relations between the hypernymic pairs and therefore the validity of the
candidates. These criteria are semantic guidelines that closely interact, and help us in this first
phase to get an idea of what the extracted data set contains. Also, they give us a first picture of the
implicit encoding of regular polysemy in WordNet.
1) Semantic homogeneity: The same semantic relation should apply between the senses of the
participating words. For instance, the hypernymic pair act-building has nine participating words
such as chapel (“a service conducted in a chapel” versus “a place of worship that has its own
altar”). Apart from elements such as these there are words like architecture (“the profession of
designing buildings” versus “the building”) and theatre “(the art of writing and producing plays”
versus “the building”). The different types of activity, captured by one sense of each of these
72
words, does not constitute a semantically homogeneous set. Whereas designing, writing and
producing might be considered to be forms of creating, conducting a service is not. The
conclusion must be that “activity” is a concept that is too general to enable the distinction of a
meaningful relation in this pattern (see also criterion 2 below), and that the set of words
subsumed by this hypernymic pair does not constitute a semantically homogenous pattern,
because not all words are to be considered valid instantiations of the pattern act-building.
2) Semantic transparency: It should be possible to distinguish a meaningful semantic relation
between the two hypernyms and the subsumed sense combinations. This often entails that the
hypernymic pair should be fairly specific. For instance activity-group is a pair at a higher level in
the WordNet hierarchy than the pair group action-organization. The second pair is subsumed by
the first, because it offers us more specific information about the nature of the regular polysemy.
The evaluation was performed by randomly examining a number of candidate regular polysemic
patterns and their associated word-sense-pairs. For each word sense pair, it was assessed whether
the same relation held between each pair, i.e., whether it was a case of real regular polysemy. The
concept pairs that were evaluated as valid show a low percentage of unsuccessful instantiations in
the form of sense pairs of participating words. On average, of each valid pattern, around 10% of
the word members extracted did not pass the two criteria above. The observed regular polysemic
patterns correspond in a number of cases to relations attested in the literature (see list in section
5.1). This is particularly true for those sets containing a relatively large number of members, such
as music - dance and container – quantity. Table 6.1 lists a few examples of the results.
The regular polysemic patterns attested in the literature and the valid patterns obtained from the
manual evaluation served as the basis for the creation of a gold standard set of patterns. This gold
standard was then used as a yardstick for evaluating the data reduction criteria described in the
next section. The gold standard consists of 138 hypernymic pairs and is listed in Appendix B. It
73
needs to be borne in mind that the selection process of this gold standard set was based on
subjective judgment and intuition, but is a necessary preamble to any automatic selection
approach.
Relation
Examples
publication-publisher
musical composition – group
of singers
building –
institution/association
package/container –
collection
music –
arrangement/formation
construction – body of
people
paper, newspaper, magazine
trio, quartet, suite
No of words
covered
3
16
school, chamber, court
15
parcel, bundle, pack
5
Line, arrangement, chorus
3
house, body, camp
5
Table 6.1: Randomly Evaluated Metonymic Patterns
6.2.1 Metonymic and Metaphoric Patterns
Although most of the valid patterns under examination were judged to be instances of metonymy,
and, more specifically, regular polysemy, some of the word sense combinations reflect a
metaphorical relation. Although according to Apresjan these are generally not considered
productive, the results from the evaluation indicate a dependency on a specific part of the
WordNet hierarchy. For instance, table 6.2 lists metaphoric relations restricted to hyponyms of
the unique beginner pair, artefact-cognition. A unique beginner is the top concept of one
WordNet branch. There are nine unique beginners in WordNet, each of which forms the top of a
separate taxonomic branch.
These concepts are listed in table 6.3 below. We seem to have stumbled on a metaphorical
domain expressed by the combination of two unique beginners where properties of artificial
structures are mapped onto cognitive skills. Different types of metaphorical transfer can be
identified, such as ‘theory as a supporting structure for thought’ and ‘a melodic theme as
74
pervasive idea’. The fact that this domain is particularly productive for metaphoric extension does
seem to indicate that extensions like these can behave more or less regularly as soon as the a
correspondence between two knowledge domains has been established (cf. section 2.4.4). This
would, at least in part, invalidate Apresjan’s view.
Relation
Examples
structure – idea
musical theme – idea
concrete obstruction abstract obstruction
device - something having
influence over somebody
artistic work – idea
Concrete
representation/creation mental representation
framework, foundation, base
theme, motif, strain
barrier, roadblock, hurdle
No of words
covered
5
4
8
bait, lure, support
3
design, motif, pattern
figure, model, scene
3
8
Table 6.2: Metaphoric Patterns
Unique
Beginner
abstraction
act
event
group
phenomenon
possession
psychological
feature
state
Gloss
a general concept formed by extracting common features from specific
examples
something that people do or cause to happen
something that happens at a given place and time
any number of entities (members) considered as a unit
any state or process known through the senses rather than by intuition or
reasoning
anything owned or possessed
a feature of the mental life of a living organism
the way something is with respect to its main attributes; "the current state of
knowledge"; "his state of health"; "in a weak financial state"
Table 6.3 Unique Beginners in WordNet1.6
Both metonymy and metaphor reflect different viewpoints (Ballim et al., 1991; see also previous
section) or ways of seeing (Cruse, 2000) on the meaning of a word in discourse. For instance, in a
sentence that contains a metonymic expression such as “Prussia invaded France in 1871.”, it may
75
be important to take into account exactly how the believer may be viewing the invasion, whether
it was the Prussian army, the government or the people as a whole that did something (Ballim et
al., 1991, p. 135). This is generally applicable to all words that participate in regular polysemic
patterns: the way they are interpreted depends on the nature of the pattern and the salience of the
senses. Metaphor clearly presumes viewpoints, since it is explicitly defined as a view of one
“target” subject matter, T, that uses descriptive resources from another “source” subject matter, S,
a way of seeing T as S (Barnden, 1997). Although there is a conceptual difference between
metonymy and metaphor (see section 2.4.4), both phenomena display, up to a certain extent, the
property of lexical systematicity in WordNet as well as in other resources, and are therefore
extracted by the technique described in chapter 5.
After inspection of the data, many more instances of metonymy were found than instances of
metaphor, and this is the reason why this thesis concentrates on the former phenomenon. In this
sense, the data seems to support Apresjan’s view that metonymy is a more productive
phenomenon, and that it therefore establishes itself more in lexicons than metaphor. A potential
reason for this might be that metonymy is much more situationally and conceptually constrained
by contiguity, whereas metaphor seems to be much more free with respect to its associability, i.e.
is less constrained by contiguity in the correspondence or transfer stage, as long as there is a
certain level of similarity. The stronger the structural analogy between the domains is, the more
potential the metaphor will have to become lexicalized. The fact that metaphor, after having been
established, allows reasoning (Barnden et al., 1995), reflects that contiguous knowledge from the
source domain S can now be projected onto the target domain T in order to reinforce the
similarity, and facilitate learning and understanding.
6.3 Automatic Data Reduction
Any of the 57177 automatically extracted combinations of hypernymic concepts can in principle
76
yield valid groups, and ideally none of these combinations should be weeded out until they have
been evaluated. However, it is possible to define certain criteria that reduce the set of data to be
examined, or select subsets defined according to certain specifications. The hypothesis is that
these criteria and techniques will maximise the probability of valid patterns of regular polysemy.
Before applying these criteria, two initial reduction steps were applied in order to maximise the
validity of the candidate patterns resulting from the application of the criteria.
The first step involved the elimination of all hypernymic pairs that contain the node entity as one
of its members. This node is one of the highest nodes in the WordNet hierarchies and therefore
very uninformative. Obviously, none of the pairs belonging to the gold standard contain the entity
node. After weeding out the pairs with entity the unreduced set consists of 55702 hypernymic
pairs.
The second step involved the number of participating words. Raising the threshold from two
words per candidate pattern to three we end up with 8062 words involving 21270 senses, with
22869 candidate RP patterns. This already cuts off more than half of the data. Although this
discarded data may contain valid RP patterns, the small number of participating words does not
guarantee a higher probability of an established pattern. By applying this filter we have
eliminated many spurious candidates, and this will enable us to concentrate on the more
conspicuous regular polysemic patterns in WordNet. Below follows a description of each applied
criterion. The overall evaluation of all criteria can be found in section 6.4.
Criterion 1: Number of Participating Words
Increasing the minimal number of participating words from three to at least four reduces the data
set to 14774 candidate RP patterns with 7884 words and 10881 word senses. Although the
discarded data may contain valid RP patterns, the small number of participating words (two or
three) is a very marginal indication of an established pattern. By applying this filter we have
77
eliminated many spurious candidates, and this will enable us to concentrate on the more
conspicuous regular polysemic patterns in WordNet.
Examples of weeded out gold standard patterns (see Appendix B):
32
press – publisher
33
religious ceremony – structure
36
beverage – plant part
57
combat – emotion
103
improvement - motion
Criterion 2: Level of Hypernymic Pair Members in the Hierarchy
The discussion in section 5.2.1 concluded that the ideal patterns to aim for are the ones whose
constituent concepts are not too high up in the hierarchies. Concept pairs that are too high up tend
to subsume a variety of more informative sub patterns (i.e. a set of more specific patterns at lower
levels in the hierarchy that are subsumed by a high level concept pair), and their level of semantic
underspecification is therefore too high for our purposes. For example, the candidate pair
abstraction (a general concept formed by extracting common features from specific examples ) –
cognition (the psychological result of perception and learning and reasoning) has 489
participating words. It is, however, impossible to determine a clear-cut relation between the two
concepts in the pattern. Conversely, patterns too low down in the hierarchies fail to capture an
adequate level of semantic generalization, and cover only small numbers of participating words.
The specificity of the hypernymic concepts involved can be used as a criterion to obtain a
reduction of the data set. One indication of concept specificity is the level of the hypernyms in the
WordNet hierarchical chains. The higher the position they occupy in the hierarchy, the more
general the pattern they instantiate. The lower their position is in the hierarchy, the more specific
the concepts are and the more homogenous the semantic relation is that relates the senses. This
criterion concentrates on the highest levels in the hierarchy.
78
First, all candidate pairs containing entity (a top concept in the WordNet taxonomy) as pair
member had already been removed in the pre-processing stage (see above). The exclusion was
extended to all concepts that occupy the first or second level in the hierarchy. Matching this result
against the gold standard, only 65% of this set was retained. Rejecting only the candidate pairs
containing concepts of level 1 (i.e. the unique beginners; see table 6.3 above) resulted in a
percentage of 92.8% retained gold standard patterns. On the basis of this, the latter criterion was
chosen.
Examples of weeded out gold standard patterns (see Appendix B):
22
group – belief
24
act – ability
55
possession – distribution
101
feeling – state
124
act - appearance
Criterion 3: Semantic Overlap
According to this criterion, the level of semantic overlap between the nodes in each hypernymic
pair determines its validity. The nodes are regarded as semantically overlapping if they share a
common hypernym. This can happen in two ways:
a) Within in the same chain. The algorithm used in the analysis looks for word sense
combinations that occur within different hierarchical chains in WordNet. In around 200
cases, however, the senses involved belong to the same chain, because one of the word
senses involved has two hypernyms (multiple inheritance). Overall, there are 872
instances of multiple inheritance relations for nouns in WordNet.
Sense 1
arnica – (used especially in treating bruises)
ð tincture – (a medicinal extract in an alcohol solution)
ð medicine, medication, medicament, medicinal drug – (something that treats or prevents or
79
alleviates the symptoms of disease)
ð drug – (something that is used as a medicine or narcotic)
ð artifact, artefact – (a man-made object)
ð object, physical object – (a physical (tangible and visible) entity
ð entity, something – (anything having existence (living or nonliving))
Sense 3
arnica – (an ointment used in treating bruises)
ð ointment, unguent, balm, salve – (used for healing or soothing)
ð remedy, curative, cure – (a medicine or therapy that cures disease or relieve pain)
ð medicine, medication, medicament, medicinal drug – (something that treats or
prevents or alleviates the symptoms of disease)
ð drug – (something that is used as a medicine or narcotic)
ð artifact, artefact – (a man-made object)
ð object, physical object – (a physical (tangible and visible) entity; “it was full of
rackets, balls and other objects”)
ð entity, something – (anything having existence (living or nonliving))
ð treatment – (care by procedures or applications that are intended to relieve illness or
injury)
ð care, attention, aid, tending – (the work of caring for or attending to someone or
something)
ð work – (activity directed toward making or doing something)
ð activity – (any specific activity or pursuit)
ð act, human action, human activity – (something that
people do or cause to happen)
Figure 6.1: Multiple Inheritance
For example, the pattern remedy (a medicine or therapy that cures disease or relieve pain)
versus drug (something that is used as a medicine or narcotic) subsumes two senses of
arnica (see figure 6.1 above). Sense two is subsumed by two hypernymic branches
starting with remedy and drug. The hypernymic path associated with sense 1 and 3
overlap at the node of ‘medicine’. Therefore ‘remedy’ is a hyponym of two unique
beginners: ‘object’ and ‘act’.
In the first case it is a hyponym of drug, but in the second it belongs to a different
hypernymic chain. Because of this reason the pattern remedy – drug comes out of the
automatic analysis, although in one of the possibilities the senses belong to the same
chain.
80
b) In different chains that come together at a higher point in the hierarchy. Because they
share a common hypernym, combinations of this kind can be expected to express more
closely related concepts than the concept pairs whose members belong to two different
taxonomic chains. For instance, the pattern cloak-dress subsumes 3 words. Both concepts
meet at the node garment:
garment
overgarment
cloak
dress
Figure 6.2: ‘Cloak/Dress’
Another example concerns the pattern implement (“instrumentation (a piece of equipment
or tool) used to effect an end”) versus natural object (“an object occurring naturally; not
made by man”) (e.g. stick). Both concepts meet at the high level node object.
entity, something
object, physical object
artifact
instrumentation
natural object
implement
Figure 6.3: ‘implement/natural object
Many of the patterns rejected by this criterion contain members whose hypernymic structures
converge at high levels in the WordNet taxonomy, for instance at the nodes abstraction (level 1),
psychological feature (level 1), causal agent (level 2) and change (level 3). It can be argued that
these concepts are too general to block these candidate patterns, because relevant regular
81
polysemic information can still be captured by some of them. However, manual inspection shows
that a vast majority of these candidate patterns are not suitable for RP, and therefore the criterion
was applied.
The result of this pruning process is a set of 15288 remaining candidate RP patterns (5034 words,
12417 word senses).
Examples of weeded out gold standard patterns (see Appendix B):
13
facility – structure
16
decoration – structure
41
fabric – covering
103
improvement – motion
117
written communication - message
Criterion 4: Informativity
This criterion defines a score for the level of informativity of the hypernymic pair on the basis of
its number of participating words: if a hypernym pair (hp1) is the hypernym of another hypernym
pair (hp2), and hp2 subsumes at least 90% of the words participating in hp1, then hp1 is
discarded. This criterion serves to weed out the less informative pairs that occur higher up than
necessary in the taxonomy. The more specific pairs occur at a more fine-grained level in the
hierarchies, and therefore describe the RP pattern in a more precise way. The requirement that
these more specific classes contain at least 90% of the class defined by the pair higher up in the
hierarchies guarantee a minimal loss of coverage. For example, the pattern edible fruit – fruit tree,
containing words such as cherry, pear and pond apple, is subsumed by the hypernymic pair
produce-angiospermous tree. The first pair has 88 participating words, whereas the latter has 90.
The more specific pattern edible fruit – fruit tree thus covers 96% of the more general pattern.
The additional semantic information available for this pattern (fruit as a subset of produce; fruit
82
tree as a subset of the angiospermae) is deemed to be a good reason for the exclusion of the
semantically more underspecified pattern. Overall, the application of this criterion resulted in a
set of 16514 hypernymic pairs.
Examples of weeded out gold standard patterns (see Appendix B):
25
object – aspect
43
person – job
78
natural object – location
101
state - feeling
135
food –artifact
Criterion 5: Semantic Similarity
Resnik (1999) applies notions from information theory to establish a measure of semantic
similarity between WordNet concepts. This measure takes into account the distance between two
concepts and their most specific common hypernym (also called most informative subsumer) in
terms of node traversal, as well as the position of subsumer/subsumed concepts in the hierarchy.
The theory behind this is that semantic similarity is higher at more specific levels in the
hierarchies.
Most gold standard pairs (124 of 138) have a zero score. This is the expected score because this
means that the concepts in the candidate pattern do not belong to the same taxonomic branch in
WordNet. The highest score of the 16 remaining from the gold standard is 2.68. Determining the
criterion as the selection of all candidate pairs with an informativity score of 2.68 or smaller
seems to constitute a feasible cut-off point. The result is a set of 22476 candidate patterns.
Criterion 6: Proportional Coverage
Another possible indication of the validity of candidate RP patterns is to compare the number of
83
words participating in each pattern (pc1) with the number of words subsumed by each member
concept in the hypernymic pair (pc2 and pc3). The following algorithm was used to determine the
proportional coverage (PC) score: pcscore =
pc1/((pc2+pc3)/2)) = 2pc1/(pc2+pc3). The
hypothesis is that a higher pcscore is an indication of a more significant candidate RP pattern,
because it has a higher level of proportional representation in the combination of the member
concepts of the hypernymic pairs. The highest pcscore in the gold standard is 0.31, whereas the
lowest pcscore is 0.0003. Filtering out all candidate pairs with a pcscore of less than 0.0003
yields a set of 22869 valid candidate pairs. This means that there is no reduction of the initial set
of candidate patterns (see p.77). The overall range of this pcscore is between 1 (100% coverage)
and 0.00007. The low value for this initial cut-off point shows that the proportional coverage
criterion does not yield clear-cut results. The incremental increase of the pcscore threshold value
gives the results listed in table 6.4. In order to keep this criterion in line with the other ones in
terms of retaining score against the gold standard (which calls for a retaining score of more than
85%, see table 6.5 below), a pcscore of 0.002 was chosen as the cut-off point.
Pcscore
Resulting number of
candidate RP patterns
Percentage
of
reduction
0.001
0.002
0.003
0.004
0.005
19037
15012
12412
10430
9106
17%
34%
46%
54%
60%
Percentage
of retained
Gold
Standard
patterns
95%
87%
80%
75%
72%
Table 6.4: Incremental PCscore and Pruning Results
Examples of weeded out gold standard patterns (see Appendix B):
6
person - weaponry
89
life form - fabric
84
25
object - aspect
49
life form – garment
108
person - ability
6.4 Evaluation against the Gold Standard
The scoring results of each criterion against the gold standard is shown in table 6.5. We started
off with a set of 22869 candidate patterns resulting from the preprocessing stage described in the
beginning of section 6.3.
Criterion
Resulting number of Percentage
candidate RP pairs
of reduction
1
2
3a
3b
4
5
6
14774
19141
22714
15288
16514
22476
15012
35%
16.3%
0.7%
33.2%
27.8%
1.7%
34.3%
Number of
retained Gold
Standard
patterns
130
129
138
122
118
138
122
Pe rcentage of
retained gold
standard
patterns
94.2%
93.5%
100%
88.4%
85.5%
100%
88.4%
Table 6.5: Scoring of Pruning Criteria
If we regard the criteria as equal in pruning performance, the overlap between all criteria results
in a set of 4659 candidate patterns. Matched against the gold standard set it only retains 46.4% of
all gold standard pairs. We have to keep in mind that for this manual selection task the
assignment of valid candidate pairs from a Cartesian product list is always up to a certain point
subjective. The cross-tabulation of two WordNet chain fragments makes it difficult to pinpoint
one particular candidate pattern as the best. It seems therefore reasonable to ontologically relax
the matching criterion against the gold standard: if any pair in the sets listed above does not
belong to the gold standard, but is a direct hypernym or direct hyponym of any gold standard pair,
then it is also counted as a hit. This means that both members of the pair should be a direct
85
hypernym or hyponym of any gold standard pair. However, this relaxation of the criteria, from
direct hypernyms or hyponyms to hypernyms or hyponyms three nodes away from any original
gold standard pair, did not increase the number of retained gold standard patterns.
Overall, the reduction of the data set by taking the intersection of the sets yielded by the six
criteria above is quite dramatic: 91.9% of the original dataset has been discarded. The small
percentage of retained gold standard pairs, however, decreases the reliability of this set as
containing the most probable valid regular polysemic patterns.
6.5 Evaluation of Precision and Conclusion
The reduced data set, produced by the reduction techniques described in the previous section,
requires an evaluation in order to establish how well the reduction methodologies have worked.
In order to do this 100 randomly chosen patterns from the pruned set of 4659 patterns were
manually evaluated. 64% of these were judged to be valid instances of regular polysemy. We may
therefore estimate a precision performance of 64% for this set of regular polysemic patterns. The
conclusion is that the filtering on the basis of the six criteria is reasonably accurate.
A question we can ask ourselves now is whether the result is actually useful in processing text, in
other words, do we get reasonable coverage of instances of regular polysemy in language use in
the form of a text corpus? This question will be addressed in the next chapter.
86
Chapter 7
Textual Evidence: Comparing the Extracted Data Set
with a Semantically Annotated Text Corpus
This chapter describes the first in a series of two strategies, by which the automatically reduced
data set resulting from the techniques in chapter 6 is extrinsically evaluated by means of other
resources.
Until now, the criteria used in the definition of regular polysemic patterns in WordNet and their
automatic extraction and filtering, applied in the previous chapters, are restricted to the domain of
monolingual taxonomical systematicity. The evaluation of these methods has, until now, relied on
intrinsic criteria, as described in the previous chapter. In order to further investigate the validity
of the regular polysemic patterns, we need to establish an estimate of the validity of the data by
means of extrinsic criteria, i.e. by comparing the extracted regular polysemic patterns with data
from other resources. For this purpose we have opted for two different approaches. The first
concentrates on discourse topicality, which is described in this chapter, and involves the use of a
semantically annotated corpus, more specifically, a text corpus that has been annotated with
WordNet senses. The corpus that fits this requirement is Semcor (see section 7.2 below).
7.1 Word Senses in Documents
It has been argued (Gale et al, 1992) that words tend to be used in only one sense within a
document. This seems a valid observation for homonymous words, in the sense that where the
word senses are not related in any way, only one of them tends to be used in a document.
However, it has been observed that senses of the same word do co-occur within the same
87
document (Krovetz, 1998). When these word senses co-occur they are semantically close or
related.
Topical distinctions, i.e. where the word senses involved occur in different domains of discourse,
tend to coincide with homonymous sense distinctions occurring in different documents. For
instance, financial documents will refer to bank as an institution rather than a sloping piece of
land. Sense co-occurrences (multiple occurrences of the same word, but with different senses)
within the same document do not normally reflect topical shifts (Markert and Nissim, 2002),
because they highlight related or semantically close aspects of the word in question. The semantic
proximity between two senses of the same word that co-occur within a topical boundary can take
on several forms. Amongst these there are the following types of semantic relatedness:
1) Generalization versus specification, where one sense can be seen as the hypernym of the
other. For instance, hammer (small mallet used by a presiding officer or a judge) is in
WordNet1.5 a hyponym of hammer (a hand tool with a heavy rigid head and a handle;
used to deliver an impulsive force by striking).
2) Vagueness, where the boundaries between the senses are not well defined, and senses
have a great amount of semantic overlap. For example, band has two senses: “a group of
musicians playing popular music for dancing” and “instrumentalists not including string
players”. They are both direct hyponyms of musical group.
3) Regular polysemy, where the senses of a word are related in a systematic way.
For our purposes we will concentrate on the last case, where instances of regular polysemic
patterns occur within the same document. The evaluation of the regular polysemic patterns
extracted from WordNet involves a comparison of these patterns with sense co-occurrences in a
corpus. This should evaluate the adequacy of the regular polysemic patterns, and the importance
of regular polysemic patterns used within documents, as opposed to other types of semantic
proximity.
88
7.2 Using Semcor for the Evaluation of RP Patterns
The evaluation technique to be applied involves the use of Semcor (Landes et al., 1998), a subset
of the Brown corpus that has been semantically tagged with WordNet1.6 senses. This corpus
consists of around 360,000 word forms distributed over 127 documents. In order to be able to
evaluate the data from the set of 4659 candidate regular polysemy patterns, the words that
participate in them, together with their senses, were mapped onto sense co-occurrences in Semcor
documents. We define sense co-occurrences here as multiple occurrences of the same word, but
with different senses, within the same Semcor document. There are 184 words from the data set
containing regular polysemic patterns whose participating word senses co-occur in two or more
Semcor documents. We chose a subset of 23 words with senses co-occurring in more than seven
Semcor documents for the evaluation of this method. We postulated that these high frequency
sense co-occurrence pairs would maximise the chance for valid regular polysemic patterns. Eight
of these words were covered by the extracted regular polysemic patterns: business, door,
development, fact, hour, life, problem, world. Seven of these were considered valid cases of
regular polysemy. These are listed below.
business
RP pattern:
group action (action taken by a group of people)
organization (a group of people who work together)
sense 1: the activity of providing goods and services involving financial and commercial and
industrial aspects; "computers are now widely used in business"
sense 2: a commercial or industrial enterprise and the people who constitute it; "he bought his
brother’s business"; "a small mom-and-pop business"; "a racially integrated business concern"
89
door
RP pattern:
structure (a thing constructed; a complex construction or entity; "the structure
consisted of a series of arches"; "she wore her hair in an amazing construction of
whirls and ribbons")
way (any road or path affording passage from one place to another; "he said he
was looking for the way out")
sense 1: a swinging or sliding barrier that will close the entrance to a room or building; "he
knocked on the door"; "he slammed the door as he left"
sense 2: the space in a wall through which you enter or leave a room or building; the space that a
door can close; "he stuck his head in the doorway"
hour
RP pattern:
measure (how much there is of something that you can measure)
clock time (the time as given by a clock; "do you know what time it is?"; "the
time is 10 o’clock")
sense 1: a special and memorable period; "it was their finest hour"
sense 2: clock time; "the hour is getting late"
problem
RP pattern:
message (what a communication that is about something is about)
condition (a condition or state at a particular time: "a condition (or state) of disrepair"; "the current status of the arms negotiations")
90
sense 1: a state of difficulty that needs to be resolved; "she and her husband are having
problems"; "it is always a job to contact him"; "urban problems such as traffic congestion and
smog"
sense 2: a question raised for consideration or solution; "our homework consisted of ten problems
to solve"
fact
RP pattern:
information (knowledge acquired through study or experience or instruction)
message (what a communication that is about something is about)
sense 1. a piece of information about circumstances that exist or events that have occurred.
sense 2. a statement or assertion of verified information about something that is the case or has
happened.
development
RP pattern:
change of state (the act of changing something into something different in
essential characteristics)
process
(a sustained phenomenon or one marked by gradual changes;
"events now in process"; "the process of calcification begins
later for boys than for girls")
sense 1: act of improving by expanding or enlarging or refining; "he congratulated them on their
development of a plan to meet the emergency"; "they funded research and development"
sense 2: a process in which something passes by degrees to a more advanced or mature stage; "the
development of his ideas took many years"; "the evolution of Greek civilization"; "the
slow development of her skill as a writer"
life
RP pattern:
cognition (the psychological result of perception and learning and reasoning)
being (the state or fact of existing: "a point of view gradually coming into
being";)
sense 1: a characteristic state or mode of living; "social life"; "city life"; "real life"
91
sense 2: the experience of living; the course of human events and activities; "he could no longer
cope with the complexities of life"
Word
Pattern
business
door
hour
problem
fact
development
life
group action/organization
structure/way
measure/clock time
message/condition
information/message
change of state/process
cognition/being
Other WordNet words
participating in the same
pattern
21
13
4
12
13
22
5
Success Rate
76%
100%
100%
84%
54%
54.5%
80%
Table 7.1: Mapping of Data Set onto Semcor Sense Co-occurrences
For this successful set of patterns, the other words from WordNet participating in the regular
polysemic patterns (column 3 of table 7.1) were then manually evaluated, and the suitability of all
words for each pattern was manually judged. The success rate in column 4 of table 7.1 indicates
the percentage of words that participate in the same pattern derived from WordNet as the word in
the first column and whose sense combinations have been evaluated as valid instantiations of the
regular polysemic pattern.
For instance, the pattern group action (action taken by a group of people) - organization (a group
of people who work together) has 21 participating words: delegation, mission, line, intelligence,
industry, resistance, hands, government, defence, expedition, association, authority, battery,
faith, commercial enterprise, confederation, deputation, polity, business, offence, alliance. After
manual inspection the sense co-occurrences of 16 of these words were deemed to be valid
instantiations of this regular polysemic pattern. This results in a success rate of 76% for the
pattern group action – organization in terms of participating words in WordNet.
The only unsuccessful word is world. The associated pattern is natural object (an object
occurring naturally; not made by man) / content (the sum or range of what has been perceived). It
92
was impossible to determine a meaningful relation between these two concepts, especially with
respect to the participating senses of world:
sense 1: the 3rd planet from the sun; the planet on which we live; "the Earth moves around the
sun"; "he sailed around the world"
sense 2: all of your experiences that determine how things appear to you; "his world was
shattered"; "we live in different worlds"; "for them demons were as much a part of reality as trees
were"
If we want to evaluate the importance of regular polysemic patterns in terms of occurrence within
Semcor documents, as opposed to other types of semantic proximity, we arrive at a percentual
estimate of the occurrence of regular polysemy in Semcor on the basis of the following
computation: 7 (number of words judged as having frequently co-occurring senses in Semcor
related by regular polysemy) divided by 23 (number of words with frequently co-occurring senses
in Semcor) = 30.4%. Since we take only into account the candidate patterns that are actually
present in Semcor, the recall (total number of candidate patterns found in Semcor) is 100%. The
applied methodology has a precision (number of correctly identified regular polysemic patterns in
Semcor) of 7/8=87.5%, with an F-measure of 93%.
The remaining 15 words whose senses co-occur in seven Semcor documents or more, do not
participate in any of the 4659 patterns. The sense combinations belonging to eight of these fifteen
words can be considered cases of taxonomic generalization/specification, i.e. where one sense is a
more specific instantiation of the other. The generalization/specification relation is often not
reflected in the WordNet taxonomy. A few examples are:
day:
sense 1: time for Earth to make a complete rotation on its axis.
93
sense 2: the period of time taken by a particular planet (e.g. Mars) to make a complete rotation on
its axis.
child
sense 1: a human offspring (son or daughter) of any age; "they had three children"; "they were
able to send their kids to college"
sense 2: a young person of either sex (between birth and puberty); "she writes books for
children"; "they''re just kids"; "`tiddler'' is a British term for youngsters"
case
sense 1: an occurrence of something; "it was a case of bad judgment"; "another instance occurred
yesterday"; "but there is always the famous example of the Smiths"
sense 2: a special set of circumstances
man
sense 1: the generic use of the word to refer to any human being; "it was every man for himself"
sense 2: an adult male person (as opposed to a woman); "there were two women and six men on
the bus"
moment
sense 1: a particular point in time; "the moment he arrived the party began"
sense 2: an indefinitely short time; "wait just a moment"; "it only takes a minute"; "in just a bit"
As a final observation, none of the senses were deemed metaphorically related.
94
7.3 Discussion
The number of examined words (23) is rather small and therefore it is difficult to ascribe any
statistical significance to the results. Nevertheless, it is possible to draw the tentative conclusion
of an expected precision score of 87.5%.
Around 70% of this set of words with frequently co-occurring senses within the same Semcor
document is not covered by the data set. One question that can be raised with respect to this group
is whether they should be incorporated as instantiations of regular polysemic patterns, i.e.
whether they are members of any group of words whose senses are related to each other in the
same way. Manual evaluation, although a difficult and subjective task, reveals that the senses
involved seem to have such a high degree of overlap that they can be regarded rather as clusters
of generalization, where two or more senses are lumped together, than as systematically related
concepts from separate branches within an ontology.
Also, it needs to be borne in mind that Semcor has been manually annotated, and that five of the
seven words that were not covered by either a regular polysemic pattern from WordNet or the set
of generalization-specification sense combinations described in the previous section have more
than six senses. The more polysemous the words, the more fine-grained the sense distinctions are,
and the more difficult it is to assign senses properly and consistently. Increasing numbers of
senses generally indicate a higher level of granularity of the sense distinctions. The makes the
proper and consequent assignment of senses a more difficult and negatively influences interannotator agreement (Véronis, 2001). Differences in sense assignments in Semcor adversely
affect the match with the regular polysemic patterns extracted from WordNet.
The outcome of the small-scale evaluation in section 7.2 supports the conclusion of (Krovetz,
1998) that related senses co-occur within the same document. This co-occurrence of multiple
senses within a discourse can be used as evidence for lexical semantic relations. In the case of
Semcor, around 30% of frequent sense collocations within the same document are metonymically
95
related in terms of regular polysemy patterns in WordNet. The results for metaphor are in line
with Apresjan's assertion that metaphorical patterns are generally not instances of regular
polysemy, although our analysis of WordNet indicates the existence of some metaphorical
patterns (see section 6.2.1).
Another observation is that 30% of frequent sense collocations in Semcor are instances of regular
polysemy. This is a higher percentage than the finding that 17% of expressions in text corpora are
metonymic (Markert and Nissim, 2000; see section 3.8), but this might be an acceptable
divergence given the small size of the experiment described above. The fully pruned set of 4659
regular polysemic patterns, captures all the instances of regular polysemy in Semcor.
Finally, we can conclude on the basis of these findings that WordNet is a resource that captures
regular polysemy in a reasonably accurate way, and that it can be, up to a certain extent, useful in
processing a text corpus by providing knowledge about systematic relationships between word
senses that occur within the same document.
The next chapter examines other aspects of regular polysemy, namely the validity of the patterns
in terms of multilinguality and universality.
96
Chapter 8
Regular Polysemy as a Cross-Linguistic Phenomenon
The previous chapter explored the validity of regular polysemic patterns as represented in the
American English WordNet in terms of word sense co-occurrences in texts, the first approach to
extrinsic evaluation of the validity of the reduced data set of 4659 patterns. This chapter describes
the second approach, which concentrates on multilinguality, and exploits the multilingual
information in EuroWordNet. It explores the question whether the obtained patterns have a higher
level of universality, i.e. if they can be detected across languages. For this purpose we have
analyzed sense distinctions from three different language resources in EuroWordNet (English,
Dutch and Spanish) and their behaviour with respect to regular polysemy.
8.1 The Distinction of Word Senses across Languages
Some time ago, Resnik and Yarowsky (1997) suggested that for the purposes of word sense
disambiguation, the different senses of a word could be determined by considering only sense
distinctions that are lexicalized cross-linguistically. Intuitively, one assumes that if another
language lexicalizes a translation in two or more ways, there must be a conceptual motivation.
For instance, the noun interest has among its senses ‘financial interest in something’ and ‘a fixed
charge for borrowing money; usually a percentage of the amount borrowed’. These senses have
different translations into Dutch, respectively inzet and rente. This provides an indication that the
English senses for interest are conceptually distinct. On the other hand, child has two senses in
WordNet1.5: “a human offspring” and “a young male child”. Both senses have parallel
translations in Spanish, German and French, respectively nino, Kind and enfant. This is a strong
indication that both senses are strongly related, and have been separated without a proper
97
conceptual motivation. The overall hypothesis is that, if we look at enough languages, we would
be likely to find the significant lexical differences that delimit different senses of a word (Ide,
1999).
Identity of lexicalization can also be used as a criterion to select regular polysemic patterns from
the data set described in chapter 6 that are valid across languages. For this purpose, we applied
the following methodology. We start from two hypotheses. The first hypothesis has already been
discussed earlier in section 2.7, and postulates that regular polysemy is borne out by intra-lingual
regularities in lexicalization patterns. This means that metonymically related senses are, apart
from compounding and derivation, largely realized by the same lexicalization within one
language, i.e. they constitute different senses of the same word. The second hypothesis is that, if
metonymy is a universal phenomenon, regular polysemic patterns are reflected in identical
translational equivalences across languages. If these patterns are lexicalized in more than one
language, this would point to the possibility that metonymically related senses are not distinct
senses in themselves, but need to be linked together into some kind of semantic structure. The
semantically contiguous senses are perceived more as a whole than as separate entities. This
assumption seems to be in line with the conceptual gestalt approach of Blank (1999) where
metonymically related senses are perceived as different entry points into a knowledge frame,
which functions as a perceptual gestalt (see section 2.5).
8.2 Previous Work
The question whether regular polysemy is a cross-linguistic phenomenon has until now only been
approached by small-scale analyses. Kamei and Wakao (1992) approached the question from the
perspective of machine translation and conducted a comparative survey of the acceptability of
metonymic expressions in English, Chinese and Japanese consisting of 25 test sentences. The
98
results they report show that in some cases English and Japanese share metonymic patters to the
exclusion of Chinese, but that in others English and Chinese are alike.
Seto (1996) performed a study into the lexicalization of the container-content schema in various
languages (Japanese, Korean, Mongolian, Javanese, Turkish, Italian, Germanic and English).
This pattern is lexicalized in English by ‘kettle’:
1. A metal pot for stewing or boiling; usually with a lid
2. The quantity a kettle will hold
His observation was that the pattern is observable in all languages, and can be considered crosslinguistic. This small study seems to indicate that the regular polysemic pattern extends over
language family boundaries to such an extent that it almost seems universal. This could suggest
that the pattern is rooted in general human conceptualization, and reflects an important nonarbitrary semantic relation between concepts or objects in the world. For instance, if we describe
the relation between container and content in terms of Aristotle’s qualia structure (Pustejovsky,
1995; see section 3.1), we define the function of a container as to hold an object or substance
(telic role) and the semantic class of a container as an artifact, because it is normally brought into
existence for the purpose of containing.
More small-scale studies like the ones described above have been performed, mostly relying on
introspection and small-scale dictionary analysis. A limited number of patterns that are valid in
more than one language have been identified such as container/content and producer/product
(Peters, 2000). With the availability of WordNet and EuroWordNet it has become possible to
investigate the cross-linguistic nature of metonymy on a large scale.
8.3 EuroWordNet
EuroWordNet (EWN) (Vossen 1997; Peters 1998, see also section 4.4) is a multilingual thesaurus
incorporating wordnets from eight languages: English, Italian, Dutch, German, Spanish, French,
99
Czech, and Estonian. The wordnets have been built in various ways. They generally have been
created on the basis of language specific resources such as monolingual dictionaries, and matched
onto the original Princeton WordNet (Fellbaum 1998) in a later stage, when the links from the
language specific wordnet to the interlingual index were created. They therefore reflect la nguage
specific lexicalization patterns and semantic organization. Others have been built from the start
on the basis of a match between WordNet and bilingual dictionaries. In this case the conceptual
structure is less language specific but can be regarded as the conceptual overlap between the
structure of the English WordNet and the ontological structure associated with that particular
language.
The structure of EuroWordNet gives us, for the first time, the opportunity to examine the question
of the language independence of regular polysemy in a more systematic and automatic way. The
data obtained from this resource depends heavily on vocabulary coverage in the various la nguage
wordnets, and until the advent of EuroWordNet no lexical database with the appropriate semantic
information was available for analysis. The EuroWordNet database is the most comprehensive
multilingual thesaurus to date. This resource not only provides us with an appropriate amount of
lexical information in terms of vocabulary coverage, but also has the additional advantages that
its taxonomic building blocks are identical for all languages involved and the language specific
concepts are all linked to an interlingua which is based on the full set of the original Princeton
WordNet (version 1.5), and is referred to as the interlingual index (ILI) (Peters et al., 1998).
8.4 Lexical Triangulation
In order to determine whether regular polysemy is indeed a cross-linguistic phenomenon, one
needs to compare languages, preferably from different language families. We started with a
comparative analysis of Germanic and Romance languages. The main reason for this choice is
that the size of the corresponding wordnets is large enough to yield significant results. For our
100
analysis we used three languages: English, Dutch and Spanish, hence the term for this process:
lexical triangulation. Singling out areas where three language-specific lexicalization patterns
converge enabled us to identify metonymic patterns that supported the hypothesis that certain
relationships between senses are inherent. We extracted the sense combinations of Spanish and
Dutch words that participate in any of the potential regular polysemic patterns from the initial
large set described in chapter 5. In other words, we concentrate here on lexicalization patterns in
three different languages: sense combinations that are lexicalized by one language-specific word
in English, Spanish and Dutch.
The first step in this process was the reduction of the search space for regular polysemic patterns
in EuroWordNet. First we determined the conceptual overlap for nouns between the English,
Dutch and Spanish wordnets. Table 8.1 below shows the number of nouns in the three wordnets
involved.
Language
Number of
noun synsets
English
Dutch
Spanish
66025
28352
24073
Number of
corresponding ILI
concepts
66025
26779
24087
Table 8.1: Conceptual Coverage of English, Dutch and Spanish Wordnets
The conceptual overlap between these wordnets is computed by determining the intersection of
ILI noun concepts covered by each of the wordnets. The total overlap is 17007 ILI noun concepts.
The corresponding set of WordNet1.5 synsets contains 920 English polysemous nouns with
senses that have identical translations in Spanish and Dutch. The senses in the different wordnets
are linked through the ILI concepts by means of equivalence synonymy or near-synonymy
relations (Vossen 1997). For example, the English word church has one sense that refers to a
building and another that denotes an institution. The same sense distinctions apply to the Spanish
101
iglesia and the Dutch kerk . Further, birthday has two senses (“an anniversary of the day on which
a person was born” and “the day on which a person was born”). Both senses are mirrored for the
Spanish cumpleanos and the Dutch geboortedag.
The second step was to map the WordNet senses of the 920 English polysemous nouns with
senses that have identical translations in Spanish and Dutch onto the results from the wordnet
analysis described in section 5.2.2, and then to evaluate the cross-linguistic validity of the regular
polysemic patterns that have been projected from the English monolingual wordnet onto the
Dutch and Spanish wordnets.
8.5 Evaluation
The result of this cross-linguistic projection yields a subset of the unreduced monolingual
analysis data set described in section 5.2.2. It covers 404 distinct English nouns out of a total of
8062 (5%).
The percentage covered by the cross-linguistic data compared to the original analysis varies from
a 100% for the very small potential classes of regular polysemy (2-3 words) to 1-2% for middle
sized (30-50 words) and large classes (100+ words).
In order to create a set for manual evaluation, the set of 404 English nouns was reduced by
discarding candidate RP patterns that do not occur in the automatically pruned set described in
chapter 6. This step reduced the number of participating words to 224. At this point, 177 words
were randomly chosen from this set for manual evaluation. The evaluation consisted of the
examination of the hypernym pairs that reflect a candidate regular polysemic relation. The
guiding criteria used in this evaluation phase were semantic homogeneity (the semantic relation
that defines the candidate regular polysemic class should apply to the majority of the participating
words) and specificity of the pattern (the lower the position of the hypernymic pair in the
hierarchy, the more specific the semantic relation).
102
The result was that 109 of these 177 words displayed valid regular polysemic patterns (62%), 68
did not (38%). This means that by means of this automatic filtering method we have a 62%
success rate for identifying valid regular polysemic patterns. Below are a few examples of crosslinguistic RP classes that have satisfied the criteria of the evaluation.
Hypernymic Pair: Control (the activity of managing or exerting control over something) –
Trait (a distinguishing feature of one's personal nature)
English RP class (7 total): abstinence, sobriety, inhibition, restraint, self-control, self-denial,
self-discipline
Dutch RP class (2 total): zelfcontrole, onthouding
Spanish RP class (3 total): autodiscipline. abstinencia, abnegación, inhibición
Coverage of the intersection between all three languages: 36% of the RP class members
derived from WordNet have a cross-linguistic projection in Dutch and Spanish.
Hypernymic Pair: Fabric (something made by weaving or felting or knitting or crocheting
natural or synthetic fibers) - Covering (a natural object that covers or envelops)
English Rp class (4 total): wool, hair, fleece, tapa
Dutch RP class (1 total): wol
Spanish RP class (1 total): lana
Coverage of the intersection between all three languages: 25% of the RP class members
derived from WordNet have a cross-linguistic projection in Dutch and Spanish.
Hypernymic Pair: Plant (a living organism lacking the power of locomotion) - Edible fruit
(edible reproductive body of a seed plant especially one having sweet flesh)
English RP class (159 total): apple, boxberry, blackcurrant, banana, fig . . .
Dutch RP class (9 total): banaan, vijg, persimoen, meloen…
103
Spanish RP class (20 total): banana, plátano, melón, caqui, higo…
Coverage of the intersection between all three languages: 2.5% of the RP class members
derived from WordNet have a cross-linguistic projection in Dutch and Spanish.
Hypernymic Pair: Person (a human being) - Quality (an essential and distinguishing attribute
of something or someone)
English RP class (11 total): attraction, authority, beauty, . . .
Dutch RP class (1 total): schoonheid
Spanish RP class (4 total): belleza,atracción, autoridad, imagen
Word intersection between all three languages: 9% of the RP class members derived from
WordNet have a cross-linguistic projection in Dutch and Spanish.
Hypernymic Pair: Substance (that whic h has mass and occupies space) - Drug (something that
is used as a medicine or narcotic)
English RP class (25 total): alcohol, bromide, dragee, histamine, iodine, liquor…
Dutch RP class (2 total): broom, cocktail
Spanish RP class (10 total): bromuro, histamina, muscatel, yodo…
Word intersection between all three languages: 4% of the RP class members derived from
WordNet have a cross-linguistic projection in Dutch and Spanish.
Hypernymic Pair: Occupation (the principal activity in your life) – Discipline (a branch of
knowledge)
English RP class (6 total): architecture, literature, politics, law, theology, interior design
Dutch RP class (1 total): architectuur
Spanish RP class (2 total): arquitectura, teología
104
Word intersection between all three languages: 16% of the RP class members derived from
WordNet have a cross-linguistic projection in Dutch and Spanish.
8.6 Universality of Regular Polysemy
It is possible to view these results as an indication of the cross-linguistic validity of the regular
polysemic patterns and their level of universality relative to the language families represented by
the wordnets. The hypothesis is that if a metonymic pattern occurs in several languages, there is
stronger evidence for a higher level of universality of the regular polysemic pattern.
Of course there is interference with the coverage of the wordnets in EuroWordNet. Since the
Dutch and Spanish wordnets are only half the size of the English wordnet only limited coverage
can be expected. Still, the coverage seems to be consistently low in most cases, often not more
than 2-5%. On the basis of wordnet size only one would expect a higher coverage.
Although it is probable that many cases where language-specific wordnets do not correspond with
respect to regular polysemic patterns is mainly due to gaps in the coverage of the individual
wordnets, it is possible to envisage other explanations for the lack of identical lexicalizations in
other language wordnets than English:
1. The metonymic pattern in the English source language is language specific, and is not realised
as a polysemous word in the target language. For example, the Dutch kantoor is synonymous to
the English office in the sense “where professional or clerical duties are performed”, but its sense
distinctions cannot mirror the regular polysemic relation in English with “a job in an organization
or hierarchy”. Instead, other types of word formation are used to account for the metonymy. Quite
often the missing sense can in fact only be lexicalized by either another word or a
compound/derivation related to the word with the potentially missing sense. This is the case with
Spanish translational equivalences of the container-containerful senses of pitcher. The container
105
sense is aguamaril, whereas the containerful sense is cántaro. For the Dutch kantoor (office) the
job sense is the compound kantoorbaan (office job). Further, the Dutch vereniging has the sense
“an association of people with similar interests”. The English equivalent is club, for which there
is another sense in Wordnet (“a building occupied by a club”). This is not a felicitous sense
extension for the Dutch vereniging, because the favoured lexicalization is the compound
verenigingshuis (club house).
2. The pattern is unattested in the target language in terms of usage but forms a potential sense
extension in that language. For instance, Dutch derivational patterns as described above have the
potential to be superseded by the base word. This may happen especially in pragmatically more
restricted situations. For instance, the English calf can mean the animal as well as the leather
obtained from the animal. Dutch generally uses a derivation (kalfsleer) of the base word kalf. The
base word can be used in a situation where the leather of a calf is opposed to that of any other
type of animal: “Is this chamois? No, it’s calf.” The metonymic sense of kalf in Dutch is always
ready to be activated.
3. The metonymic pattern is in fact attested in the language, but one or more senses participating
in the patterns has not yet been captured in the wordnet. For example, embassy has one sense in
WordNet (“a building where ambassadors live or work”). The Dutch translational equivalent
ambassade has an additional sense denoting the people representing their country. This sense can
be projected to the English WordNet as a regular polysemy pattern that is also valid in English. In
fact, LDOCE (Procter,1978) only lists the sense which is missing in WordNet.
Further, the Spanish iglesia and the Dutch kerk both mean “building for worship’” and “a service
conducted in a church”. The Spanish wordnet has an additional systematically related sense for
iglesia (“institution to express belief in a divine power”) that is not shared by its Dutch
counterpart but is a valid new sense in this language.
106
4. The metonymic pattern is in fact attested in the language, but the granularity of the wordnet in
question does not distinguish between the senses, and lumps them together. For instance, the
Dutch word binnenhuisarchitectuur (interior design) has only one sense which covers two
metonymically related senses in Princeton WordNet (see below).
8.7 Coverage and Extendibility
The explanations listed in the previous section need verification by testing native speaker
intuitions about the regular polysemic gaps found in the data. This was done by projecting the
senses of the participating English words in an RP class onto Dutch and Spanish, and assessing
whether the missing senses were adequate additional senses in these two languages.
The experiment that was conducted was very small. The pattern we examined is the hypernymic
combination occupation (the principal activity in your life) – discipline (a branch of knowledge).
This regular polysemic class has five members. Two Dutch and two Spanish native speakers were
asked to judge the felicitousness of the senses that are missing in the Dutch and Spanish
wordnets. Below is a short discussion of each member.
interior design
1. the trade of planning the layout and furnishings of an architectural interior
2. the branch of architecture dealing with the selection and organization of furnishings for an
architectural interior
The corresponding Dutch word binnenhuisarchitectuur has only one sense which is linked to both
WordNet senses by means of a near-synonymy relation. This means that the Dutch wordnet is
underspecified for the distinction of these metonymically related senses and can be extended with
107
the specific sense distinctions (see explanation 4 in the previous section). This coincided with the
verdict of the Dutch jury. The Spanish WordNet has a separate translation for each sense:
interiorismo (corresponding to interior design 1) and deseno de interiores (corresponding to
interior design 2). The latter translational equivalent was considered to also have a possible
“trade” reading.
law
1. the learned profession that is mastered by graduate study in a law school and that is responsible
for the judicial system
4. the branch of philosophy concerned with the law
The Dutch word rechtswetenschap also has only one sense which is linked to both WordNet
senses by means of a near-synonymy relation. This again means that the Dutch wordnet is
underspecified for the distinction of these metonymically related senses and can be extended with
the specific sense distinctions (see explanation 4 above). The Spanish equivalent of law 4 is
jurisprudencia, whereas law 1 does not have a correspondence in the Spanish wordnet. The
profession reading was not considered a felicitous additional sense for this word. Both subjects
remarked that another word captures both meanings: leyes, which is not present in the Spanish
wordnet.
literature:
1. the profession or art of a writer
2. the humanistic study of a body of literature
The Dutch letterkunde is only linked up to sense literature no. 2. Sense no. 1 was not considered
to be a straightforward new sense for this word by the judges. The Spanish literatura lacks a
108
“profession” reading in the Spanish wordnet. This sense was considered as valid by one subject,
but rejected by the other subject.
Politics
1. the profession devoted to governing and to political affairs
2. the study of government of states and other political units
The Dutch word politicologie also has only one sense that is linked to both WordNet senses by
means of a near-synonymy relation. This again means that the Dutch wordnet is underspecified
for the distinction of these metonymically related senses and can be extended with the specific
sense distinctions. The Dutch subjects, however, were not happy with the profession reading.
The Spanish política lacks a profession reading in the Spanish wordnet. The Spanish subjects
considered this a valid sense for this word.
Theology
1. the learned profession acquired by specialized courses in religion (usually taught at a college or
seminary
2. the rational and systematic study of religion and its influences and of the nature of religious
truth
The Dutch theologie has no “profession” reading. This reading was considered valid by the Dutch
subjects. The Spanish teología has both senses in the Spanish wordnet, and this coincides with the
subjects’ intuition.
The results are summarized in table 8.2 on the next page. Overall, the projection of the word
senses onto the Dutch wordnet yields a sense extension for one word out of a possible two. For
the Spanish wordnet the same process creates valid new senses for two out of four words.
109
interior design
Senses available Sens e possible
Senses available Sense possible
in NL
in NL
in ES
in ES
lumped into one
+
-
+
+
-
-
sense
law
lumped into one
sense
literature
-
-
-
+/-
politics
lumped into one
-
-
+
+
+
+
sense
theology
-
Table 8.2: Experimental Results Summary
8.8 Conclusion
The manual evaluation shows regular polysemy patterns to be valid across the three languages
examined, and indicates that regular polysemy has a certain level of universality. Also, the results
of the experiment indicate a potential for enhancing the semantic compatibility and consistency of
wordnets. Wordnets are, up to a certain extent, automatically extendable on the basis of regular
polysemic information available from other wordnets. In our small experiment 50% of the Dutch
and Spanish words that do not display a WordNet-derived regular polysemic pattern were
successfully semantically enriched with this pattern on the basis of the opinion of the judges.
110
Chapter 9
Building Knowledge Fragments
Each pair of hypernymic nodes whose extraction and evaluation is described in the previous
chapters implicitly captures a systematic semantic relation between word senses of words that
have senses under both hypernymic nodes in the WordNet hierarchies. It does not, however, give
any explicit information about the nature of this relation. If we want to know that, for instance, an
animal can be used for human consumption (i.e. if we want to capture an explicit relation between
e.g. chicken as an animal and chicken as food), this will require an additional effort, which is
described in this chapter.
9.1 Establishing Semantic Relations between Word Senses Captured by
a Regular Polysemic Pattern
There are several ways in which the determination of the semantic relations between the related
word senses can be approached. First, one can stop here and use the unlabelled associations in a
knowledge base by employing a semantic relation such as ‘is related to’ between the senses. The
semantic characterization of the relation will then only imply that they are related in some
unspecified way. Second, these relations can be determined by means of human introspection.
The examination of the pair and the participating word senses will provide a human assessor with
enough information to postulate a relationship. However, this is a costly and time consuming
activity that is worthy of a project on its own. An example of this type of project is FrameNet
(Fillmore and Atkins, 1992), which involves as one of its phases the determination of the
structure of the frames involved, and the nature of the slots and fillers.
111
A third option is to automate this process by exploiting the semantic information available in the
WordNet glosses for the process of extracting explicit semantic relations between the word senses
involved in the regular polysemic pattern. This method is applied and evaluated in this section.
9.1.1 The Algorithm
The extraction process concentrated on the linguistic information available in the WordNet
glosses that are associated with the concepts in the taxonomic chains that contain the hypernymic
pairs and the word senses subsumed by them. First, the glosses were preprocessed. Part of speech
tags were added by means of a version of the Brill tagger (Hepple, 2000). Further, nominal and
verbal content words were lemmatized with a lemmatizer that uses the WordNet morphological
information.
tea
Sense 1
tea -- (a beverage made by steeping tea leaves in water; "iced tea is a cooling drink")
=> beverage, drink, drinkable, potable -- (any liquid suitable for drinking: "may I take your
beverage order?")
=> food, nutrient -- (any substance that can be metabolized by an organism to give energy and build
tissue)
=> substance, matter -- (that which has mass and occupies space; "an atom is the smallest
indivisible unit of matter")
=> object, physical object -- (a physical (tangible and visible) entity; "it was full of rackets,
balls and other objects")
=> entity, something -- (anything having existence (living or nonliving))
=> liquid -- (a substance that is liquid at room temperature and pressure)
=> fluid -- (a substance that is fluid at room temperature and pressure)
=> substance, matter -- (that which has mass and occupies space; "an atom is the smallest
indivisible unit of matter")
=> object, physical object -- (a physical (tangible and visible) entity; "it was full of rackets,
balls and other objects")
=> entity, something -- (anything having existence (living or nonliving))
Sense 3
tea, Camellia sinensis -- (extensively cultivated in e.g. China and Japan and India; source of tea leaves)
=> shrub, bush -- (a low woody perennial plant usually having several major branches)
=> woody plant, ligneous plant -- (a plant having hard lignified tissues or woody parts especially
stems)
=> vascular plant, tracheophyte -- (green plant having a vascular system: ferns, gymnosperms,
angiosperms)
=> plant, flora, plant life -- (a living organism lacking the power of locomotion)
112
=> life form, organism, being, living thing -- (any living entity)
=> entity, something -- (anything having existence (living or nonliving))
Figure 9.1: hypernymic chains for the senses of ‘tea’ captured by the regular polysemic pattern
‘Plant-Beverage’
For each word sense pair (synset pair) subsumed by a particular hypernymic pair (e.g. the sense
pairs of the words chicory, coffee, mate, sarsaparilla and tea for the hypernymic pattern plantbeverage), the WordNet glosses associated with these synsets and with their hypernyms were
analyzed by mapping the synset elements onto the glosses in the following way. For each of these
related sense pairs, one member is subsumed by hypernym1 and the other by hypernym 2, as
illustrated in figure 9.2. For all nouns participating in the same way as tea in the regular
polysemic patterns from the gold standard set two bags of words were created, each associated
with a sense captured by the regular polysemic pattern. The bag consisted of the word itself, its
synonyms (synset members) and all the members of the synsets that are hypernyms of the word
sense in question.
In the next step of the process the words in the bag of the first word sense were matched against
the processed gloss associated with the synset to which this sense belongs (henceforth synset 1).
If there was a match, the words from the bag of the second word sense (henceforth synset 2) were
matched against the gloss. If there was a match and the word from the synset 1 bag (word 1)
preceded the word from the synset 2 bag (word 2) within the gloss, the text between the matches
was extracted. If this span of text contained a verb, it was selected, together with any associated
prepositions. The verbs that are realized in the surface structure and link any of the words in each
bag are considered here to reflect a semantic relation. Although the analysis only takes
collocation into account, positional constraints ensure that the collocations indeed reflect
syntactic relations in the sentence structures of the glosses. A distance limit of three positions
between the matched nouns and the verb was applied in order to reduce spurious matches. It is
113
then postulated that these verbs represent an instantiation of the relation(s) holding for the regular
polysemic pattern.
The same matching process was repeated for the glosses associated with all hypernyms of synset
1. Then the whole process was repeated, looking for matches in the gloss of synset 2 and all its
hypernyms. Figure 9.2 below gives a graphical representation of the process.
Gloss synset 1 and 2
Gloss hypernyms synset 1 and 2
word 1 “* * VERB * *” word 2
Set of synonyms
and hypernyms
Stage 2 match
Set of synonyms
and hypernyms
Stage 2 match
Word sense 1
Word sense 2
Stage 1
Create bag
of words
Stage 1
Create bag
of words
Figure 9.2: Graphical Representation of the Relation Extraction Algorithm
The requirement that word 1 precedes word 2 is geared towards the extraction of transitive and
prepositional verbs, both used in active form. Passive constructions are also extracted in the form
114
of a past participle combined with ‘by’ or any other preposition (e.g. ‘used as’). The order
constraint also determines the syntactic order of the relation, i.e. which hypernymic pair member
is the subject and which is the object of an extracted verb. The relations that were extracted in this
manner for tea, and, consequently, for the pattern ‘plant-beverage’, are the following: ‘tea as a
plant is made into a beverage by steeping’, and ‘tea is a beverage’.
Another example will illustrate the process in detail. The regular polysemic pattern ‘animal-food’
is applicable to 172 words in WordNet. One of these words is herring:
Sense 1: commercially important food fish of northern waters of both Atlantic and Pacific
Sense 2: valuable flesh of fatty fish from shallow waters of northern Atlantic or Pacific; usually
salted or pickled.
The hypernymic hierarchy of the participating herring senses in WordNet 1.6 is as follows:
Sense 1
herring
=> saltwater fish
=> seafood
=> foodstuff, food product
=> food, nutrient
=> substance, matter
=> object, physical object
=> entity, something
Sense 2
herring, Clupea harangus
=> clupeid fish, clupeid
=> soft-finned fish, malacopterygian
=> teleost fish, teleost, teleostan
=> bony fish
=> fish
=> aquatic vertebrate
=> vertebrate, craniate
=> chordate
=> animal, animate being, beast, brute, creature, fauna
=> life form, organism, being, living thing
=> entity, something
=> food fish
=> fish
=> aquatic vertebrate
=> vertebrate, craniate
=> chordate
=> animal, animate being, beast, brute, creature, fauna
115
=> life form, organism, being, living thing
=> entity, something
Figure 9.3: hypernymic WordNet chains for ‘herring’
food fish is the hypernym of sense 2, and has the associated gloss “any fish used for food by
human beings”. Of the words in this gloss ‘fish’ is found in the synset1 bag (330 words, amongst
which fish, entity, life form, vertebrate, craniate) and food in the synset 2 bag (518 words,
amongst which seafood, food, substance, food product, nutrient, object). The intermediate text
span is ‘used for’ which consists of a verb and a preposition. The outcome is the relationship
‘animal used for food’. This relation is found 37 times.
Yet another example is the pattern profession (an occupation requiring special education) and
discipline (a branch of knowledge) subsumes five words: architecture, literature, politics, law
and theology. Sense 6 of ‘law’ has the gloss ‘the learned profession that is mastered by graduate
study in a law school and that is responsible for the judicial system; "he studied law at Yale"’
Bag synset 1 contains profession, bag synset2 study. In between is the verb ‘is mastered by’
which yields the relation ‘profession is mastered by discipline’ for this regular polysemic pattern.
This relation is found 2 times. One other relation was found: ‘concerned with’, which occurs only
once.
Other relational patterns that have been found using this methodology are:
writing (reading matter; anything expressed in letters of the alphabet (especially when considered
from the point of view of style and effect); "the writing in her novels is excellent")
message (what a communication that is about something is about)
36 words, e.g. account, conclusion, declaration, epitaph. The relation ‘express’ occurs once,
‘state’ occurs 24 times.
116
fabric (something made by weaving or felting or knitting or crocheting natural or synthetic
fibers)
covering (a natural object that covers or envelops)
5 words: fleece, hair, tapa, tappa, wool
The relation ‘made from’ occurs once.
person (a human being; "there was too much for one person to do")
language (a systematic means of communicating by the use of sounds or conventional symbols;
"he taught foreign languages"; "the language introduced is standard throughout the text")
257 words, e.g. Tatar, Assyrian, Hopi, Punjabi
The relation ‘speak’ occurs 132 times.
9.1.2 Evaluation
All 138 regular polysemic pairs from the gold standard set in Appendix B were used in the
evaluation of this method. The manual evaluation obtained the following results:
1. For 79 of the 138 patterns relations were found by analyzing the WordNet glosses. This is
57% recall (number of patterns for which a relation is found/number of patterns). This
drops to 38% when we compute recall for the fully pruned set of around 4700 pairs (see
chapter 6).
2. The examination of a sample of the approved 79 patterns from the gold standard set,
consisting of 31 RP pairs, showed that the valid relations are in fact valid for an average
of 92% of the words whose senses are subsumed by the hypernymic pairs.
3. Of all the 183 relations that were extracted for the gold standard hypernymic pairs, 91 are
wrong, and 92 are valid. This means a precision of 50%. The valid relations cover 28 of
the 138 patterns.
117
The 31 pairs mentioned in result 2 above and the success rate of the extracted relations between
their members can be found in Appendix C. All valid relations for the gold standard are in
Appendix D.
Please note that about 50% of the RP pairs for which a relation has been extracted from the
glosses (76 in total), have more than one relation. For instance, the relations ‘bear’, ‘cultivated
for’, ‘have’ and ‘produce’ apply to the pair ‘fruit tree - edible fruit’ (see Appendix D). All
extracted relations are considered to contribute to the definition of the semantic rela tedness
between the RP pair members. The observation that more than one relation can apply raises an
interesting theoretical issue about the nature of a metonymic relation. Multiple relations highlight
the semantic complexity of the contiguity between related senses. In the example above, ‘bear’
and ‘have might be considered to be near-synonyms, whereas ‘cultivated for’ implies human
involvement and a purpose for the produced fruit. This more complex constellation of semantic
relations defining contiguity in a regular polysemic pattern seems to form a conceptual whole
whose parts are ready to be activated depending on the nature of the context. This would be an
obvious consequence of the many ways in which entities can interact in our world, but has until
now not been addressed in the literature.
9.1.3 Propagation through EuroWordNet
Regular polysemic patterns form, together with any extracted relation between the concepts that
constitute the patterns, fragments of lexical semantic knowledge. These knowledge fragments can
be further enriched with complementary semantic information from the EuroWordNet database
(EWN, see section 4.4). Each wordnet in the EWN architecture contributes its own particular set
of language-specific concepts and relations. Although the bulk of the semantic relations found in
EWN are hyponymy and hypernymy relations between senses of words belonging to the same
part of speech,, a number of wordnets, especially the Dutch, Spanish and British English ones,
supply causal and thematic relations and relations between senses of words that belong to a
118
different part of speech. The semantic information from all wordnets taken together forms the
superset of concepts and relations available in EWN. The structure of the EWN database is based
on the synsets of WordNet version 1.5, and there is a mapping facility between version 1.5 and
1.6 (Daudé et al., 2001). Therefore, the appropriate elements can be automatically selected as
extensions of the regular polysemic knowledge fragments extracted from the 1.6 version.
Figure 9.4 on the next page illustrates this process for the regular polysemic pattern disciplineprofession.
This regular polysemic pattern is valid for five words (see section 9.1 above), and the relation
‘master’ has been extracted from the glosses. This verb was then looked up in EWN and was
assigned its appropriate sense by hand (‘get the hang of, be proficient in’). discipline is one of its
arguments, and it can fulfill the agent or instrument role. profession fills its patient role. Each
arrow represents one step in the superset of EWN
relations. The relations themselves are labeled in the square boxes. Obviously, not all relations
and relational chains available in EWN have been incorporated into the picture.
One additional relation needs to be mentioned explicitly here, because it was difficult to
incorporate into the figure: master is a subevent of acquire knowledge. This relation links the
bottom and the topmost concept in the figure.
The 'TC' relation indicates the EuroWordNet top concepts that are described in section 4.4 and in
more detail in (Rodriguez et al., 1998).
Figure 9.5 tracks the EWN links associated with the synset verbalize, utter, talk , speak, mouth
(‘express in speech’) from the RP pattern person-‘speak’-language.
119
Gloss: get the
hang of, be
proficient in
Gloss: branch of
knowledge
Gloss: an occupation requiring
special education (especially in
the liberal arts or sciences)
patient
agent
discipline
profession
Master
instr
isa
subevent of
Gloss: be able to, have the
ability to
TC
can
practise
Dynamic
Mental
Xpos
hypernym
Gloss: learn,
acquire knowledge
Gloss: the quality
of being able to
perform
Gloss: be a
student of a
certain subject
cause
ability
Subevent
of
learn
cause
caused by
teacher
agent
develop
Gloss: grow, progress,
unfold, or evolve
gradually through a
process of evolution,
natural growth, or
differentiation
acquire
knowledge
teach
Gloss: impart
skills or
knowledge
Figure 9.4: Expanded Ontological Fragment for ‘Discipline’-‘master’-‘Profession’
120
Gloss: a human
being
Gloss: express in
speech
agent
person
Gloss: a systematic means of
communicating by the use of
sounds or conventional
symbols
speak
has subevent
patient
language
agent
result
speaker
antonym
language
pronounce
Gloss: speak,
pronounce, or utter in
a certain way
be silent
cause
sound
Gloss: the sudden
occurrence of an
audible event
Figure 9.5: Expanded Ontological Fragment for person-‘speak’-language
9.1.4 Discussion
In the relation extraction process the relations have been identified as verbs in the glosses that are
associated with a subset of the words that participated in the patterns. After the identification,
these relations are generalized over the whole set of word sense pairs subsumed by the
hypernymic pair in question. For instance, the ‘used for’ relation between the senses of herring
occurs in the gloss of the direct hypernym of one of its participating senses. It is then generalized
to all words subsumed by animal-food. This may sometimes lead to overgeneralization, i.e. the
invalid extension of the relation to all members of the regular polysemic set. For example, the
relation ‘summarize’ was extracted for the pattern message (what a communication that is about
121
something is about) – speech act (the use of language to perform some act) on the basis of the
gloss ‘a legal document summarizing the agreement between parties’, in which document is in the
message bag and ‘agreement’ is in the speech act bag. It is, however, intuitively obvious that not
all messages summarize speech acts.
Another problem with this technique is the ambiguity of the synset and hypernym elements that
are found to match the glosses. This can also lead to the identification of false relationships. An
example is the occurrence of end in ‘(football) the person who plays at one end of the line of
scrimmage’. This gloss erroneously yields the relation ‘play at’ for the pattern causal agentfabric, because another of the 14 senses of end in WordNet is ‘a piece of cloth that is left over
after the rest has been used or sold’ which is a hyponym of fabric. These problematic cases
reduce the precision of the technique described in this section.
Overall, low recall was to be expected from a technique like this. Although definition texts lend
themselves well for extracting encyclopaedic knowledge in terms of sentence structure and
conveyed type of information, explicit relations that match the strict criteria applied are hard to
find. The 50% success rate for the relations that did come out is reasonably encouraging.
One further step was undertaken to investigate a possible increase in precision by taking into
account the frequency of the relation in the glosses. This frequency takes into account the number
of times the verb is extracted from any synset or hypernym definition for each pair of word
senses. Taking only the most frequent relation into account for each hypernym pair slightly
decreased the precision to 44%. Discarding the relation with the lowest frequency for each pattern
increased the precision to 56%, but reduced recall (number of pairs for which a relation is
found/number of pairs) by 9%. One drawback is that the discarding of lowest frequency relations
amounts to a loss of 34% of the valid relations found in the evaluation described in section 9.1.2
above. On the other hand, the pruning technique managed to dispose of 41% of the invalid
relations from section 9.1.2.
122
The overall recall of 57% indicates that for 59 of the 138 gold standard patterns no relation was
found. Because the patterns have been manually evaluated as valid, the relation associated with
the remaining patterns from the gold standard to which the technique failed to assign a relation
can be characterized as ‘is related to’. Overall, this relation can be used as the semantic
characterization of the relation between all word sense pairs in the fully pruned data set for which
no explicit relation can be extracted. If we want to attach any probability of success to this
process, we can take a metaphorical step that compares thesauri with corpora. The textual
information available in a thesaurus can be regarded as a text corpus. This applies to WordNet
and EuroWordNet. WordNet is a monolingual corpus, whereas EWN is a parallel multilingual
corpus. If we look at the results of the evaluation of precision for sense collocations in the Semcor
corpus (50%, see chapter 7) and identical lexicalization in different wordnets (62%, see chapter
8), we can conclude that, on the basis of these figures, a success rate of about 50% can be
expected for the automatic assignment of this general relation.
Finally, with respect to the propagation through the wordnets in EWN, both extended fragments
for the patterns discipline-‘master’-profession and person-‘speak’-language show a considerable
increase in available knowledge.
9.2 Extension of Ontological Fragments
In the previous section we saw that the glosses associated with WordNet synsets provide
information about the semantic relationships between two word senses that are linked by means
of regular polysemy.
The generalization of these relations to all words that are subsumed by each hypernymic pair
postulates that this information also determines significant semantic relations between the
concepts that form the regular polysemic hypernymic pair subsuming the word senses. For
123
instance, herring as ‘fish used for food’ is generalized to apply to the subsuming concept pair
animal-food and thus becomes ‘animal used for food’.
The same technique as described in section 9.1 can be applied to an additional set of elements
relevant to the extended conceptual schemes within which the hypernym pairs occur. This
extended conceptual scheme can be see as a domain-specific or topical knowledge structure that
captures concepts and relations that are relevant for each of the hypernym pairs described in
chapter 5 and 6.
This domain-specific knowledge structure will, off course, also apply to the senses of the words
that are subsumed by these hypernymic pairs that form the regular polysemic patterns.
9.2.1 Extraction of Topical Knowledge
I will illustrate this by means of an example. If we examine the music 1 (an artistic form of
auditory communication incorporating instrumental or vocal tones in a structured and continuous
manner), we see that this concept co-occurs in the data set with a number of hypernyms in order
to form the extracted pairs of higher level concepts that subsume senses of three or more words.
These hypernyms are listed in table 9.1 below.
Hypernym
Gloss
social relation
a relation between living
organisms; esp between people.
relation
psychological
feature
an abstraction belonging to or
characteristic of two entities or
parts together
a feature of the mental life of a
living organism
person
a human being.
performer
an entertainer who performs a
dramatic or musical work for an
audience
an artifact (or system of artifacts)
instrumentality
Examples
bass,canto,fanfare,homophony,line,obbligato,
obligato,overture,resolution,rondeau,signature,
Star-Spangled banner,statement,theme,voice
bass,canto,fanfare,homophony,line,obbligato,
obligato,overture,resolution,rondeau,signature,
Star-Spangled banner,statement,theme,voice
idea,line,macumba,melody,motif,motive,resolution
,theme,
variation
Bach,bass,Beethoven,Brahms,Chopin,Handel,
Hare Krishna,Haydn,Mozart,Stravinsky,voice,
Wagner
Bach,bass,Chopin,voice
bass,hornpipe,line
124
that is instrumental in
accomplishing some end
social dancing
dancing as part of a social
occasion
Bolero,gavotte,hornpipe,landler,mazurka,
minuet,paso doble,polka,quadrille,rumba,
samba, schottische, tarantella
substance
that which has mass and occupies
space
a group of people who work
together
an activity that diverts or amuses
or stimulates
air,bass,signature
organization
diversion
chorus,Hare Krishna,line,symphony
ballet,bolero,gavotte,harmonisation,
harmonization,hornpipe, landler,mazurka,minuet,
paso doble,pavane,polka,quadrille, rumba,
samba,saraband,schottische,tarantella
air,line,voice
communication
the activity of communicating
higher cognitive
process
cognitive processes that
presuppose the availability of
knowledge and put it to use
device
an instrumentality invented for a
bass,hornpipe,line
particular purpose.
taking a series of rhythmical steps ballet,bolero,gavotte,hornpipe,landler,mazurka,
(and movements) in time to music minuet,paso doble,pavane,polka,quadrille,rumba,
samba,saraband,schottische,tarantella
an abstraction belonging to or
air,harmony,line,resolution,signature,voice
characteristic of an entity
something that people do or
accompaniment,air,ballet,bolero,gavotte,
cause to happen
harmonisation,harmonization,harmony,hornpipe,
landler,line,mazurka,minuet, part,
pasodoble,pavane,polka,quadrille,resolution,
rumba,samba,saraband,schottische,statement,
tarantella,variation,voice
a general concept formed by
air,bass,canto,fanfare,harmony,homophony,line,
extracting common features from obbligato,obligato,overture,resolution,rondeau,
specific examples
signature,Star-Spangled Banner,statement,
theme,voice
an essential and distinguishing
air,harmony,voice
attribute of something or
someone.
any specific activity or pursuit.
accompaniment,ballet,bolero,gavotte,
harmonisation,harmonization,hornpipe,
landler,line,mazurka,minuet,part,
paso doble,pavane,polka,quadrille,rumba,
samba,saraband,schottische,tarantella,variation
any communication that encodes line,Star-Spangled Banner,statement
a message.
dancing
attribute
act
abstraction
quality
activity
signal
idea,line,resolution
ballroom
dancing
any of a variety of social dances
performed by couples in a
ballroom
bolero,minuet,paso doble,rumba,samba
communication
something that is communicated
between people or groups
bass,canto,fanfare,homophony,line,obbligato,
obligato,overture,resolution,rondeau,
signature,Star-Spangled Banner,
statement,theme,voice
125
writing
reading matter; anything
Signature, theme, rondeau, resolution, obligato,
expressed in letters of the alphabet line, canto, statement
(especially when considered from
the point of view of style and
effect); "the writing in her novels
is excellent"
Table 9.1: WordNet Concepts that Co-occur with music in Hypernymic Pairs
The combination of music with any of the concepts listed in the table above represents an aspect
of the overall schematic meaning of music, i.e. the combinations express a part of the
lexical/encyclopaedic meaning of music that is implicit in the WordNet knowledge base. This
semantic information is contained in the sum of regula r polysemic collocations in WordNet and
the hypernymic generalizations that cover the word senses. In other terms, these collocations
reflect parts of the script (Schank and Abelson, 1975) for music. For instance, the pair ‘music –
writing’ represents the sound and written symbol aspects of the domain within which they both
occur.
Of course, as in section 9.1, one can stop here and use the unlabelled associations in a knowledge
base by merely indicating that they are related in some way. We can also take it further than that
and try to establish the nature of the relation that binds each concept pair together. This will
bootstrap or complement the creation of a domain specific classification scheme.
Classification schemes for the musical domain such as Vickery’s (1975) have been created
mainly by hand. It is possible to approximate or complement this task by exploiting the
systematicity in WordNet. For instance, Vickery (1975) presents a hierarchical faceted
classification scheme as one might conceive for a subject matter indexing task for the domain of
music. This particular scheme uses verbs as units of description, which are listed between quotes
(Anick and Bergler, 1992) in the figure below. The hierarchical structure is represented by means
of indentation.
126
music production
creation "compose"
performance
production means
voice "sing"
lips "whistle"
musical instrument "play"
in a particular manner
brushing across strings "strum"
sounding strings with a pick "pick"
purpose
entertaining an audience "perform"
skill improvement "practice"
Figure 9.6: Vickery’s Hierarchical Faceted Classification Scheme for the Domain of ‘music’
When we examine the data file we find a number of regular polysemic patterns in which ‘music’
collocates with another concept into a hypernymic pair that covers senses of three or more words.
These are listed in table 9.2 below.
The relations are determined on the basis of the same technique as presented in the previous
section, where two bags of words containing synonym and hypernym members are mapped onto
the glosses associated with the word senses and their hypernyms. If there is a verb between them,
this is regarded as the relation that connects the two concepts. Manual evaluation of the result of
this method indicated a coverage (recall) of 80%, i.e. 80% of the hypernym pairs involved are
actually linked by a verb in the glosses. However, the percentage of correct relations between the
concepts (precision) is only 22%.
hyper1
music
music
music
music
music
music
music
hyper2
Activity
Creator
Person
Artist
causal agent
Artist
Composer
relation
accompany
accomplish
make
accomplish
accomplish
make
accomplish
127
hyper1
music
music
music
music
music
music
hyper2
Composer
Musician
Musician
causal agent
Act
Person
relation
make
accomplish
make
make
accompany
accomplish
Table 9.2: Correctly Extracted Relations with Bag-of-Words Technique
Note that several pairs are hierarchically related. For instance, the pairs music-artist and musicperson are characterized by the same semantic relations because artist is a hyponym of person.
The reason why there are quite a few sets of hyponymically related pairs in this table is because
the procedure was applied to all pairs that subsume three or more words. In other words, only
pruning criterion 1 was applied to the original data set in order to obtain these results (see chapter
6). When the full set of pruning techniques is applied a smaller number of remaining pairs with
valid relations result from the analysis (see section 6.3 for a detailed description). These are listed
in table 9.3. Regretfully, the fully pruned data set has lost some of the finer grained semantic
information, such as the pair music-composer. The specific nature of the concept composer makes
it useful for a knowledge base. On the other hand, any human being is capable of making music,
and this information, which is still available, will be useful if more general inferences are
required.
hyper1
hyper2
relation
music
music
music
causal agent accomplish
causal agent make
Person
accomplish
music
music
Person
Activity
make
accompany
Table 9.3: Remaining Valid Relations for ‘music’ Collocates from the Fully Pruned Data Set
128
hyper1
hyper2
relation
music
think
music
higher cognitive
process
Thinking
music
music
music
Content
Idea
communication
change
change
represent
music
someone
music
written
communication
Message
music
music
music
Activity
Message
social dancing
someone
someone
render
music
music
music
Activity
Activity
Diversion
change
render
be about
think
create
Table 9.4: Incorrectly Extracted Relations with Bag-of-Words Technique
After manual evaluation we are able to construct a part of the semantic space of music on the
basis of the conceptual collocates and the relations between them (see figure 9.7). This miniature
knowledge frame covers the senses of all 34 words that are subsumed by all members in the
participating hypernym pairs in table 9.3:
accompaniment, Bach, ballet, bass, Beethoven, bolero, Brahms, Chopin, gavotte, Handel,
Hare Krishna, harmonisation, harmonization, Haydn, hornpipe, landler, line, mazurka,
minuet, Mozart, part, paso doble, pavane, polka, quadrille, rumba, samba, saraband,
schottische, Stravinsky, tarantella, variation, voice, Wagner
A full listing of these words and their participating senses can be found in Appendix E.
A greater semantic coverage can be obtained by taking all hypernym collocates of music into
account (see table 9.1) that are not yet covered by the knowledge frame in figure 9.7, and defining
a general ‘is related to’ relation between these hypernyms and music.
129
Causal agent
make
accomplish
isa
Music
Person
make
accomplish
accompany
Activity
Figure 9.7: Graphical Representation of the Knowledge Fragment Provided by the Hypernymic Pairs from
Table 9.3
When matched against the fully pruned dataset (see section 6.3) the following 10 additional
concepts can be added with which music is conceptually related. These are listed in table 9.5.
Because no explicit relation has been found for these pairs in the analysis of the WordNet
glosses the relation that links the pair elements can be let underspecified or be made explicit
after manual inspection. The result is the relation defined in the third column of table 9.5.
The default relation ‘involve’ is used here where it seems impossible to define a specific relation.
Hyper word
Hyper gloss
attribute
an abstraction belonging to or
characteristic of an entity
cognition
the psychological result of perception and
learning and reasoning
communication something that is communicated between
people or groups
content
the sum or range of what has been
perceived
Relation
involve
is structured by
involve
is structured by
130
Hyper word
Hyper gloss
diversion
an activity that diverts or amuses or
stimulates
message
what a communication that is about
something is about
social dancing dancing as part of a social occasion
social relation a relation between living organisms; esp
between people
writing
reading matter; anything expressed in
letters of the alphabet (especially when
considered from the point of view of
style and effect); "the writing in her
novels is excellent"
written
communication by means of written
communication symbols
Relation
accompany
is structured by
accompany
involve
is structured like
is structured like
Table 9.5: Additional Conceptual Collocates of ‘music’ for which no Explicit Relation has been Found
The extended knowledge fragment of the domain describing things that have to do with music is
depicted in figure 9.8. Some of these pair members are related to others by means of hyponymy.
This is indicated by the ISA relation between these elements.
131
Causal agent
Attribute
Social
relation
isa
isa
involve
Communication
Person
make
accomplish
Music
isa
Cognition
Is
structured
like
Writing
structure
accompany
isa
isa
Content
Activity
isa
structure
Diversion
Message
isa
Social
dancing
Figure 9.8: Extended Knowledge Fragment for music
Examples of conceptual collocates for discipline and fabric are listed in table 9.6. Collocates that
are grouped stand in a hypernymic relation to each other. The most general concept is always
listed first. Manual evaluation of the concepts that collocate with discipline showed a recall of
45.5% (a relation was extracted for 5 out of a total of 11 patterns), and a precision of 83% (5 out
of 6 relations were valid).
132
ARGUMENT1
DISCIPLINE
RELATION
INVOLVE
DISCIPLINE
DEAL WITH
ARGUMENT2
NATURAL
PHENOMENON
ACTIVITY
DISCIPLINE
DISCIPLINE
DISCIPLINE
CONCERNED WITH
CONCERNED WITH
CONCERNED WITH
ACTIVITY
OCCUPATION
PROFESSION
DISCIPLINE
MASTER
PROFESSION
SHELTER
TENT
MADE OF
MADE OF
FABRIC
FABRIC
COVERING
CLOTHING
OVERGARMENT
COAT
MADE OF
MADE OF
MADE OF
MADE OF
FABRIC
FABRIC
FABRIC
FABRIC
FABRIC
FABRIC
WOVEN WITH
WOVEN WITH
CORD
THREAD
FABRIC
FABRIC
FABRIC
MADE FROM
MADE FROM
MADE FROM
SUBSTANCE
MATERIAL
WOOL
FABRIC
MADE WITH
LINEN
FABRIC
FABRIC
MADE FROM
MADE FROM
BODY COVERING
HAIR
FABRIC
FABRIC
MADE OF
MADE OF
ANIMAL FIBER
NATURAL FIBER
CONSUMER GOODS
COMMODITY
MADE OF
MADE OF
FABRIC
FABRIC
ACT
DEED
ACCOMPLISHES
ACCOMPLISHES
FABRIC
FABRIC
Table 9.6: Conceptual Collocates for ‘Discipline’ and ‘Fabric’ Ordered by Hypernymy
133
9.2.2 Comparison of Classification Schemes
If we compare this fragment with Vickery’s classification scheme in Figure 9.6, we find that the
two schemes are largely complementary. The only overlap consists of the production/creation
nodes (‘compose’) in Vickery’s classification, and the ‘make, accomplish’ relation between
person and music in figure 9.8. It needs to be borne in mind, however, that Vickery’s
classification is based on topics (such as ‘performance’) and verbs, whereas we are, in this
technique, solely dealing with semantic relations between nouns. This means that the relations
between slot fillers in Vickery’s classification are solely based on troponymy (Fellbaum, 1998),
which can be paraphrased as ‘to PARENT in a certain way’, where PARENT can be substituted by
the parent node. On the other hand, the extended knowledge fragment in figure 9.8 contains more
relations between nouns than between verbs.
Given structured representations of other knowledge sources such as the one in figure 9.6, the
superset of concepts and relations can be semi-automatically integrated into an ever increasing
knowledge fragment. For instance, Vickery’s PURPOSE category can be extended with
‘accompany activity’. Other additions are ‘structured by content’ and ‘structured like writing’, for
which a functional parent (e.g. structure) needs to be chosen manually.
9.2.3 Propagation through EuroWordNet
If we apply the technique described in section 9.1.3 to obtain further conceptual elements for
inclusion into the knowledge fragments such as the one depicted in figure 9.8, we end up with
additional relational information for our frame that is taken from the direct semantic context of
the relevant synsets in EWN. The direct semantic context consists of concepts and relations in the
EWN network that can be reached by a traversal of one step through any wordnet within the
EWN network. Figure 9.9 illustrates the added knowledge for the fragment in figure 9.8.
Hypernymic knowledge from EuroWordNet is, for reasons of illustrational clarity, not included in
the figure.
134
In principle, any number of node traversals through the EWN network can be used to extend our
frame. In a similar fashion, the same method of extension can be applied to the fragment in figure
9.8 on the basis of the relational paths available in WordNet1.6. In order to maintain the
comprehensibility of the figures, the full EWN and WordNet propagation has not been included
in figures 9.8 and 9.9.
It should be noted that some of the concepts extracted from EWN are verbs. compose and play are
members of synsets and are therefore specific verb senses corresponding to concepts. This is in
contrast with the verb relations extracted from the dictionary definitions, because these have not
been disambiguated with respect to their senses and contain as yet intuitive meanings.
Musician
agent
agent
Compose
Music
critic
patient
Play
patient
patient
Music
Critical
review
Caused
by
Figure 9.9: Additional Semantic Information from the Direct Semantic Context in EuroWordNet
9.2.4 Measuring the Strength of Association
Having discovered all these relations between hypernymic pairs, we can now begin to think of
ways to formulate a tentative quantitative definition of the strength of the relationships between
word senses related through regular polysemy, and consequently their level of activation in the
interpretation process (see section 2.7.5).
135
On the basis of the nature of our data, the salience of the regular polysemic link between the
senses involved can be scored along the dimensions of the following parameters:
-
The frequency of the pattern itself in terms of participating words. For instance, if we
compare the patterns plant-food and press-publisher, the respective associated
frequencies are 393 and 4.
-
The informativity of the pattern in terms of the taxonomic level of the hypernyms in the
pair that constitute the pattern. The more specific the hypernym, the more informative the
pattern.
The frequency of the relations between the word senses as they occurred in the glosses was
considered not to be a valid criterion, because of the loss of valuable information involved (see
section 9.1.4).
After some experimentation, the following formula seems to work reasonably well as a first
attempt:
f



log  (1/ lh1 + lh 2) 




‘f’ is the frequency of the pattern, whereas ‘lh1’ and ‘lh2’ stand for the level in the WordNet
hierarchy of hypernym1 and hypernym2 respectively.
The frequency of the pattern is determined by the number of words that participate in the pattern.
This measure favors taxonomically more specific patterns, i.e. pairs of hypernyms that occupy a
lower position in the WordNet hierarchies. Overall, according to this measure, the salience scores
of gold standard patterns, as described in chapter 6, range from 2.9 to 7.9. If we adopt 2.9 as the
lower threshold score for salient association, we can judge the level of salience for each pattern.
136
Consider, for instance, the pair artefact (a man-made object) - writing (reading matter; anything
expressed in letters of the alphabet). This has a score of 6, which indicates a rather high level of
association between the senses of words such as passage, book, draft, composition, diary and
journal. This high score is reflected in the choice of Pustejovsky’s closed dot notation
(repressenting simultaneous activation in memory) for these senses of book, as described in
section 2.7.5. Appendix G lists the normalized activation scores for the gold standard pairs.
9.2.5 Conclusion
The relations between hypernymic pairs that constitute regular polysemic patterns, extracted from
the glosses, yield a recall of 80% and a precision of 22% for music. For discipline the recall is
45.5% and the precision 83%. The difference between these figures makes it difficult to obtain a
reliable probability of success for all concepts that occur in the hypernymic pairs. Averaging the
figures yields an expected 63% recall and a precision of around 52%.
The fact that all regular polysemic patterns consist of pairs of WordNet synsets allows further
traversal of the WordNet relations. This produces an extended knowledge fragment, and the
results show that it does seem possible to extract a considerable amount of information from a
lexical semantic resource such as WordNet. The combination of the semi-automatic and
automatic techniques described in the previous section provides a step toward the incremental
creation of knowledge fragments on the basis of information from WordNet and EWN.
With some manual intervention, concepts and relations between concepts can be obtained from
resource-internal data that make explicit the lexical and encyclopaedic knowledge that is
implicitly present in WordNet.
137
Chapter 10
Conclusions and Future Directions of Work
This chapter discusses the importance and outcome of the research described in this thesis.
Further, ideas are presented on how ontological fragments extracted in the previous chapters can
be used for various applications. Finally, future extensions and refinements of the applied
methods are discussed and how the obtained lexical knowledge can be integrated into an
incrementally growing knowledge base, which will be useful for a variety of NLP applications.
10.1 Original Contribution
In this thesis I have demonstrated the feasibility of obtaining knowledge from a knowledge base
that is only present in implicit form. The described methodology exploits the hierarchical
structure of WordNet, and the glosses that are associated with each concept. The explicit form of
this implicit knowledge, which is the result of the applied methodology, is presented in the form
of systematically related word senses, and semantic relations that link them, which define
relationships between word senses other than the hypernymic, meronymic and antonymic
relations that already are present in WordNet. This additional knowledge can be exploited in the
interpretation process, and assist many language engineering tasks. For instance, in question
answering, performance improvement has been reported by the incorporation of metonymic
knowledge (Stallard, 1993). Further, this knowledge has been exploited for anaphora resolution
(Markert and Hahn, 2002; Harabagiu, 1998) and information retrieval (Chugur, 2000). Beyond
natural language processing, the importance of making these systematic relations between word
senses explicit is evident for a number of related disciplines. It is desirable for lexicography, so
that the editor can make informed and consistent decisions about presenting or indicating
particular relations. It is desirable for linguistics simply because the relations between
138
metonymically related word senses are central to the language system, and reflect aspects of the
ways in which we categorize our world.
10.2 Results
The overall result of the methodologies described in the previous chapters is a resource that
characterizes regular polysemy, as it is encountered in WordNet. The resource that is the result of
this thesis contains 4810 regular polysemic patterns involving 4376 nouns and 10147 noun
senses. The accuracy of this set has been measured in various ways, and described in chapters 6, 7
and 8. For 2801 of these regular polysemic patterns one or more relations were found between
their member nodes with a reliability of 50%. This means that an underspecified relation has been
postulated between the members of 1575 regular polysemic patterns.
10.3 A Growing Knowledge Base
It is a fact that one of the key necessities for natural language understanding is a substantial
knowledge base in order to bootstrap the interpretation process. This knowledge base should
contain enough information to allow the application of inferencing methods for the processing of
coreference, anaphora and bridging expressions. For a text segment such as “The composer
finished his sonata. Music had always been his first love.”, the extended knowledge fragment for
music, illustrated in figure 9.7 in the previous chapter, enables the detection of linking relations
between the nouns (composer, sonata, music). This is achieved by means of the hypernymic
relation between sonata and music, another hypernymic relation between composer and person,
and the ‘make’ relation between person and music.
No available resource is complete in its coverage of semantic knowledge. WordNet therefore
suffers the same fate, and the fact that the characterization is wholly dependent on WordNet’s
sense distinctions and taxonomic structure of course influences the results of the research
139
described in this thesis. Some might argue that WordNet is not a good resource beause of the
nature of its sense distinctions, but it is freely available, comes with taxonomic information, and
is being used as a de facto standard in computational linguistics. Furthermore, the applied
techniques for extraction can be applied to any resource with taxonomic information, and it can
be expected that regular polysemic patterns from different resources might well complement the
WordNet derived data is some ways.
Furthermore, the ideal knowledge base should contain as much lexical semantic information as
possible in order to identify standardized, i.e. lexicalized, uses of words and their relations on the
one hand, and enable the interpretation of non-standard usage on the other. An appropriate model
must contain both particular knowledge about some non-standard interpretation, and reasoning to
make the non-standard interpretation fit the current context. The overall aim of this thesis has
been to show the feasibility of capturing at least part of the knowledge necessary for forming the
appropriate interpretation of a word in a text.
The various extraction processes described in this thesis show the first steps towards an
incremental process of adding and integrating implicit lexical knowledge into an existing explicit
knowledge structure. As we have seen in the previous chapter, this is equivalent to an incremental
resource-driven building of generic knowledge frames. These frames have not been built on the
basis of an assumption of semantic completeness of WordNet, but on the assumption that
WordNet based regular polysemic patterns are indicative of many instances of this phenomenon.
The incrementally growing knowledge base will provide an increasing amount of background
knowledge for the text comprehension process, against which the textual data will be interpreted.
This additional interpretation will be incorporated into the lexicon, and become background for
new texts. This will lead to an incremental reduction of the lexicographic burden necessary to
create detailed, task-specific lexicons for applications such as information extraction. In future,
semantic information from other, heterogeneous resources can be amalgamated into this single
knowledge base, once correspondences between the resources have been established.
140
In addition to the applications mentioned above, the database, incrementally enriched with
additional knowledge fragments, can be used for other purposes. The regular polysemic patterns
can be used as filters for sense disambiguation, because the regular polysemic patterns cluster
senses into one knowledge frame fragment. One area in which this work is particularly useful is
the actual construction of wordnets. As we saw in chapter 8, an evaluative comparison of
WordNet and other wordnets in EuroWordNet shows that these resources, in many cases, fail to
capture regular polysemy. These shortcomings can be remedied by augmenting the sense
distinctions for words both in WordNet as well as in EuroWordNet. Entries can be made more
complete by explicitly listing missing regular polysemic patterns, or by indicating their sense
extension potential.
10.4 Future Work
10.4.1 Refinement of the Pruning Criteria Described in Chapter 6
The application of all six criteria together, and the selection of their intersection as the final set of
regular polysemic patterns extracted from WordNet regards the criteria as equal in pruning
performance (see section 6.4). This has the undesirable effect of cutting into the gold standard set
too much: only 46% of all gold standard pairs are retained. As an item of future work, we can
consider identifying a subset of the criteria that would maximize the number of retained gold
standard patterns.
10.4.2 Refinement of the Extraction Process
In order to establish a maximally consolidated version of the knowledge base, we first need to
deal with a few issues. The low precision levels for the simple extraction rules applied call for
refinement. For instance, a rule can be added that covers nominalizations in addition to subjectverb-(preposition)-object patterns and their passivization. In a construction such as “the study of a
141
body of literature”, it will be useful to identify the pattern ‘discipline – communication’
(instantiated by study and body of literature respectively).
Also, the above example shows the need for the analysis of quantifier and determiner
constructions: Only the semantic head of the construction (“literature”) should be extracted from
“body of literature”. Of course, there are many more fine-grained adaptations to be made (e.g. the
interpretation of prepositional phrases), and one can argue that these tasks have already been
implemented in parsers. A future task will involve tuning a good parser to the WordNet glosses.
Experimental analysis of the WordNet gloss has been attempted so far with the Cass parser
(Abney, 1990) that produces tuples of predicates and syntactic functions. A few examples of the
CASS tuple output mode are listed in table 10.1 below.
landed
throwing
lends
made
work
broke
asked
:on floor
:obj ball
:obj wings
:obj getaway
:obj moment
:obj record
:obj nurse
:subj hat
:to member
:to inspiration
:subj thieves
:in orgasm
:in attainment
:for relief
:subj necessity
:subj realization
:subj he
:from reduction :in intensity
:subj he
Table 10.1: Output from the Cass Parser
The evaluation of the Cass-derived patterns for the set of 140 gold standard patterns revealed a
coverage of 78% (that is, 78% of the gold standard regular polysemic patterns were assigned a
relation through Cass) with a precision of just 19%.
There are various reasons for this bad performance in precision. First, the Brill tagger that
provided the input to the Cass parser may have produced many errors. Furthermore, CASSinternal processes such as the stemmer and the tag matcher (the latter is an algorithm that matches
any Penn Tree Bank-like tag (Marcus et al., 1994) onto an internal representation) may have been
responsible to a certain extent. Lastly, the program was used in its default settings and no tuning
of the CASS grammars was performed. Appendix F lists the Cass tuples associated with the
hypernym combinations from the gold standard set.
142
10.4.3 The Analysis of Text Corpora
Until now the analysis of textual material has concentrated on the WordNet glosses. Restric ting
oneself to text of this particular kind has several advantages. Firstly, the glosses are associated
with word senses and are therefore directly applicable at the sense level. Secondly, their structure
is more constrained in that dictionary definitions adhere to a more formulaic form of language use
than free text.
We can shift our focus of analysis from machine readable lexical resources to text corpora. It can
be expected that valid relations can be extracted from corpora for regular polysemic patterns. We
might encounter the problem that free text is too unrestrained for the creation of association. A
problem that we will certainly encounter is that of ambiguity. In order to analyse only the relevant
spans of text, we need word sense disambiguation (Stevenson and Wilks, 2001). After
disambiguation we can establish which corpus-derived relations are relevant.
10.4.4 The Disambiguation of the Verbal Relations
The verbs that function as the semantic relation between word senses that are related through
regular polysemy can be homonymous or polysemous in themselves.
A possible method to reduce the ambiguity of these verbs is to relate the verbs to the verbal
unique beginners in WordNet (the top conceps of the verbal hie rarchies), and use the la tter as
semantically underspecified relations. For instance, the verb speak has five senses in WordNet.
1. talk, speak, utter, mouth, verbalize -- (express in speech; "She talks a lot of nonsense")
2. talk, speak -- (exchange thoughts; talk with; "We often talk business"; also used metaphorically:
"Actions talk louder than words")
3. speak, talk -- (use language: "the baby talks already"; "the prisoner won't speak"; "they speak a strange
dialect")
4. address, speak -- (give a speech to; "The chairman addressed the board of trustees")
5. speak -- (make a characteristic or natural sound; "The drums spoke")
Sense 1 to 3 have act as the unique beginner of their taxonomic chain, sense 4 is a unique
beginner in itself, and sense 5 has sound as a unique beginner. This reduces the polysemy of
speak, and therefore the ambiguity of the relation expressed by this verb, from 5 to 3. A second
143
possibility is to connect the verbs up to the Levin database (Levin, 1993). This would yield only
one coarse-grained sense, expressed by the Levin semantic class ‘verbs of speaking’. These
coarse-grained semantic characterizations of the semantic relations can then be integrated into the
growing knowledge base.
10.5 The Separation of Regular Polysemic and Metaphoric Patterns
On the basis of the extraction method described in the previous chapter, it is not possible to
separate cases of metonymy from cases of metaphor in any other way than by inspection. It is
otherwise not feasible to determine pairs of high level concepts in the WordNet hierarchy that
capture a metaphorical mapping between domains. Also, when metonymy and metaphor are both
lexicalized in a resource such as WordNet, it is very difficult to find a way to separate the two
phenomena in a computational way by exploiting WordNet’s hie rarchical structure, given the fact
that the process interacts horribly with the very fraught theoretical issue of the distinction
between metaphor and metonomy (J. Barnden, personal communic ation).
In this work, no attempt has been made to distinguish regular polysemic patterns from metaphoric
patterns. An investigation will need to be launched in order to separate the two phenomena. For
this purpose, we can try several approaches. First, we can map conventionalized metaphors (e.g.
Lakoff and Johnson, 1980) onto WordNet nodes, and investigate if these are present in the data
set resulting from the extraction process, or if they are subsumed by a pair in the data set.
Secondly, we can make use of semantic information other than taxonomic, in the form of
domains associated with word senses. Then we might be able to find clusters of regular polysemic
patterns that are both covered by two domains, which, after inspection, could prove to be
metaphorically related.
144
10.6 Directionality of Metonymic Relations
In this thesis the issue of directionality of the semantic derivational relation between two
metonymically related word senses (base sense versus derived sense) has not been addressed. So
far the only type of directionality covered is the direction of the semantic relation extracted from
the glosses in terms of the syntactic order between subject and (prepositional) object (for
example, ‘person’ – ‘speak’ - ‘language’).
With respect to the semantic directionality of the metonymic derivation for RP patterns, it is my
opinion that we can only look for tendencies and soft rules (see also section 2.7.3). In the case of
established RP patterns,
if the sense combinations are recognized as conventionalized and
lexicalized RP patterns, I contend that there exists a bidirectional activation relation between the
pair members. Addressing the issue of directionality for these patterns then remains interesting
only from an etymological and not from a representational point of view. This, however, only
applies to the closed class of for instance the gold standard set of RP patterns. If we want to add
words to these closed classes, i.e. if we want to capture the productivity of the pattern, in order to
potentially include word candidates that only have one of the senses involved in any RP pattern,
the issue of directionality becomes relevant. As a side issue, the inclusion of new words would
mean an extension of the number of their sense distinctions. This would make WordNet more
complete in its coverage of regular polysemy.
As described in section 2.7.3, there are difficulties with the choice of vehicle and target, i.e.
establishing which sense is the base sense and which the derived sense. Directionality depends on
context and can only be coarsely constrained by epistemological and cultural notions (Radden
and Kovecses, 1999) and conceptual saliency (Langacker, 1993). If we want to approach this
problem in a (semi-)automatic fashion, we need to find quantifiable characteristics of the senses
involved. This is very difficult to achieve with epistemological and cultural notions such as
‘immediate over non-immediate’, because these do not in any way correspond to WordNet
145
concept hierarchies. The notion of conceptual salience in the choice of vehicle and target, such as
Langacker’s ‘human over non-human’ and ‘concrete over abstract’ (see section 2.7.3), gives us
more explicit taxonomic classes to exploit.
The concept ‘abstract’ corresponds to WordNet high level concept ‘abstraction’ (a general
concept formed by extracting common features from specific examples), and ‘concrete’
corresponds to ‘object’ (a tangible and visible entity). Now we can have a closer look at the gold
standard patterns that are subsumed by this example of conceptual saliency. We extract all gold
standard pairs whose pair members are hyponyms of or equal to ‘abstraction’ and ‘object’. The
following pairs are found (see Appendix B for definitions and numbering):
9
22
92
100
110
115
121
134
extremity-line
object - aspect
location – magnitude
location – time period
container- containerful
artefact-shape
object - color
book-record
For these patterns it can be assumed that the first pair member is the base sense, because it is a
hyponym of object, which, according to Langacker, has a higher level of conceptual salience than
the second pair member.
The same method can be applied to Langacker’s ‘human over non-human’. If we correlate
‘human’ to the RP pair member ‘person’, and ‘non-human’ to any other RP pair member but
‘person’, we find 18 instances of this conceptual saliency preference in the gold standard,
amongst which (see Appendix B for definitions and numbering):
6
15
19
21
30
43
61
66
108
person - weaponry
person - social group
person - organization
person - music
person - instrumentality
person - job
person - language
person - deed
person - ability
146
120
person - feeling
Overall, these two principles of conceptual saliency cover 27 patterns in the gold standard
(around 20%).
Future work will involve a further investigation of this issue, and try to define more
computational approaches to capturing directionality.
10.7 From Metonymy to Metaphor
Once the knowledge base has been established sufficiently, the focus of study can be extended
from regular polysemy to the interpretation of metaphor. The contiguity relations captured for
regular polysemy as described in chapters 8 and 9 have only revealed a very limited number of
lexicalized metaphors. It is my claim that the regular polysemic patterns resulting from the
analysis can be exploited and evaluated for metaphor discovery and metaphor-based reasoning.
This work will resemble the results obtained on the basis of a parsed version of LDOCE (Dolan,
1995).
The notion of conventional conceptual metaphor, introduced by Lakoff and Johnson (1980) is
based on the observation that, for some semantic relations, multiple terms from a common source
domain may be used metaphorically to describe multiple corresponding concepts from a common
target domain. Take, for instance, the metaphorical pattern MUSIC IS AN IDEA, which has also
been extracted as a conventionalized metaphor from WordNet (see section 6.2.1). In order to map
these domains onto each other, the contiguity relations for music, as listed in tables 9.2 and 9.6,
can be compared with the contiguity relations of idea. During the mapping process we discover
isomorphic contiguities (as was to be expected from a conventionalized metaphor). For instance,
we see CAUSAL AGENT-ACCOMPLISHES-MUSIC on the one hand, and CAUSAL AGENTACCOMPLISHES-IDEA on the other. Correspondences of this sort form the links between the
147
two domains. In general, we could use the number of commonalities and the salience of the
relations as a determiner of the strength of the metaphor.
Further, once mappings have been made between domains, as in the case of IDEA and MUSIC,
the knowledge acquired for each domain can be altered and extended on the basis of these
mappings. For instance, as described in section 9.2.1, no valid relation between the hypernymic
pair music – message could be extracted, and the hypothetical relation ‘is structured by’ was
manually added. The data set from chapter 6 reveals that message is also a hypernymic collocate
of idea, and that for this pair the relational tuples ‘MESSAGE-BE ABOUT-IDEA’ and
‘MESSAGE-DESCRIBE-IDEA’ have been extracted. Through the projection of the IDEA
domain onto the MUSIC domain, the manually added hypothetical relation ‘is structured by’ in
the MUSIC domain can now be replaced by these two relations obtained from the IDEA domain.
In this way, components from different domains can, through their links, enrich and complement
each other.
148
Appendix A: Corelex Basic Types
Basic
Word
Type
abs
abstraction
Gloss
a general concept formed by extracting common features from specific
examples
act
agt
anm
act
causal agent
animal
something that people do or cause to happen
any entity that causes events to happen
a living organism characterized by voluntary movement
art
atr
cel
artifact
attribute
cell
chm
chemical
element
chm
compound
com
con
communication
consequence
ent
entity
a man-made object
an abstraction belonging to or characteristic of an entity
the basic structural and functional unit of all organisms; cells may exist
as independent units of life (as in monads) or may form colonies or
tissues as in higher plants and animals
any of the more than 100 known substances (of which 93 occur
naturally) that cannot be separated into simpler substances and that
singly or in combination constitute all matter
(chemistry) a substance formed by chemical union of two or more
elements or ingredients in definite proportion by weight
something that is communicated between people or groups
a phenomenon that follows and is caused by some previous
phenomenon; "the magnetic effect was greater when the rod was
lengthwise"; "his decision had depressing consequences for business"
anything having existence (living or nonliving)
evt
fod
event
food
something that happens at a given place and time
any substance that can be metabolized by an organism to give energy
and build tissue
frm
shape
grb
biological group
the spatial arrangement of something as distinct from its substance;
"geometry is the mathematical science of shape"
a group of plants or animals
grp
grs
group
people
grs
hum
social group
person
any number of entities (members) considered as a unit
(plural) any group of human beings (men or women or children)
collectively; "old people"; "there were at least 200 people in the
audience"
people sharing some social relation
a human being; "there was too much for one person to do"
lfr
lme
loc
life form
linear measure
location
any living entity
a measure of length
a point or extent in space
log
region
mea
measure
a large indefinite location on the surface of the Earth; "penguins inhabit
the polar regions"
how much there is of something that you can measure
mic
nat
microorganism
body of water
any organism (animal or plant) of microscopic size
the part of the earth''s surface covered with water; "they invaded our
territorial waters"
nat
land
the solid part of the earth''s surface; "the plane turned away from the
sea and moved back over land"; "the earth shook for several minutes";
"he dropped the logs on the ground"
149
Basic
Type
Word
Gloss
nat
phm
natural object
phenomenon
an object occurring naturally; not made by man
any state or process known through the senses rather than by intuition
or reasoning
pho
plt
pos
object
plant
possession
a physical (tangible and visible) entity; "it was full of rackets
a living organism lacking the power of locomotion
anything owned or possessed
pro
process
a sustained phenomenon or one marked by gradual changes; "events
now in process"; "the process of calcification begins later for boys than
for girls"
prt
part
psy
rel
psychological
feature
definite quantity
indefinite
quantity
relation
a portion of a natural object; "they analyzed the river into three parts";
"he needed a piece of granite"
a feature of the mental life of a living organism
spc
space
an empty area (usually bounded in some way between things); "the
architect left space in front of the building"; "they stopped at an open
space in the jungle"; "the space between his teeth"
sta
state
sub
substance
tme
time
the way something is with respect to its main attributes; "the current
state of knowledge"; "his state of health"; "in a weak financial state"
that which has mass and occupies space; "an atom is the smallest
indivisible unit of matter"
the continuum of experience in which events pass from the future
through the present to the past
tme
time period
tme
time unit
qud
qui
a specific measure of amount
an estimated quantity
an abstraction belonging to or characteristic of two entities or parts
together
an indefinite length of time; "a time period of 30 years"; "hastened the
period of his recovery"
a unit for measuring time periods
150
APPENDIX B: Gold Standard Regular Polysemy Patterns
1
profession
sciences)
discipline
an occupation requiring special education (especially in the liberal arts or
a branch of knowledge; "in what discipline is his doctorate?"; "teachers should be
well trained in their subject"; "anthropology is the study of human beings"
Example words: architecture, law, literature, politics
2
game
a contest with rules to determine a winner; "you need four people to play this
game"
equipment
an artifact needed for an undertaking or to perform a service
Example words: baseball, basketball, football, handball
3
life form
any living entity
consumer goods
goods (as food or clothing) intended for direct use or consumption
Example words: banian, banyan, beaver, boa
4
act
something that people do or cause to happen
social event
an event characteristic of persons forming groups
Example words: celebration, ceremony, commencement, competition
5
act
sum
something that people do or cause to happen
a quantity of money; "he borrowed a large sum"; "the amount he had in cash was
insufficient"
Example words: accumulation, allowance, cleanup, contribution
6
person
a human being; "there was too much for one person to do"
weaponry
weapons considered collectively
Example words: blade, dummy, flak, gun
7
person
position
a human being; "there was too much for one person to do"
(in team sports) the role assigned to an individual player; "what position does he
play?"
Example words: catcher, center, end, field general
8
belief
any cognitive content held as true
institution
an organization founded and united for a specific purpose
Example words: Brahmanism, Brahminism, Buddhism, Christianity
9
extremity
the outermost or farthest region or point
151
line
a length (straight or curved) without breadth or thickness; the trace of a moving
point
Example words: boundary, brink, crown, peak
10
person
a human being; "there was too much for one person to do"
quality
an essential and distinguishing attribute of something or someone
Example words: attraction, authority, beauty, image
11
group action
action taken by a group of people
social group
people sharing some social relation
Example words: administration, alliance, assemblage, assembly
12
flavorer
herb
something added to food primarily for the savor it imparts
a plant lacking a permanent woody stem; many are flowering garden plants or
potherbs;
some having medicinal properties; some are pests
Example words: angelica, anise, basil, borage
13
facility
enormous
structure
something created to provide a particular service; "the assembly plant is an
facility"
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"; "she wore her hair in an amazing construction of whirls and
ribbons"
Example words: armory, armoury, arsenal, field house
14
work
activity directed toward making or doing something; "she checked several points
needing further work"
social group
people sharing some social relation
Example words: administration, alignment, business, enterprise
15
person
a human being; "there was too much for one person to do"
social group
people sharing some social relation
Example words: appointment, armorer, armourer, authority
16
decoration
structure
something used to beautify
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"; "she wore her hair in an amazing construction of whirls and
ribbons"
Example words: architrave, brass, cornice, festoon
17
creation
the human act of creating
commercial enterprise the activity of providing goods and services involving financial and
commercial and industrial aspects; "computers are now widely used in
152
business"
Example words: building, construction, drilling, excavation
18
activity
any specific activity or pursuit; "they avoided all recreational activity"
place of business
an establishment where business is conducted
Example words: cabaret, casino, center, marketplace
19
person
a human being; "there was too much for one person to do"
organization
a group of people who work together
Example words: armorer, armourer, authority, bodyguard
20
feeling
the psychological feature of experiencing affective and emotional states; "he had
a feeling of euphoria"
property
a basic or essential attribute shared by all members of a class; "a study of the
physical properties of atomic particles"
Example words: bitterness, buoyancy, coolness, covetousness
21
person
music
a human being; "there was too much for one person to do"
an artistic form of auditory communication incorporating instrumental or vocal
tones in a structured and continuous manner
Example words: Bach, bass, Beethoven, Brahms
22
group any number of entities (members) considered as a unit
belief any cognitive content held as true
Example words: absolutism, Brahmanism, Brahminism, Buddhism
23
message
what a communication that is about something is about
speech act
the use of language to perform some act
Example words: account, affirmation, agreement, assertion
24
act
something that people do or cause to happen
ability possession of the qualities (especially mental qualities) required to do something or get
something done.
Example words: absorption, accomplishment, accounting, acquisition
25
object a physical (tangible and visible) entity.
aspect a distinct feature or element in a problem; "he studied every facet of the question"
Example words: department, hand, sector, side
26
location
a point or extent in space
job
the occupation for which you are paid; "he is looking for a job"
Example words: caliphate, emirate, position, prefecture
153
27
action
something done (usually as opposed to something said); "there were stories of
murders and other unnatural actions"
time period
an indefinite length of time; "a time period of 30 years"
Example words: beginning, climax, commencement, conclusion
28
change
the act of changing something; "the change of government had no impact on the
economy";
interpretation an explanation resulting from interpreting something
Example words: approximation, clarification, enlargement, expansion
29
control
the activity of managing or exerting control over something; "the control of the
mob by the police was admirable"
trait
a distinguishing feature of one's personal nature
Example words: abstinence, inhibition, restraint, self-control
30
person
a human being; "there was too much for one person to do"
instrumentality an artifact (or system of artifacts) that is instrumental in accomplishing some end
Example words: accumulator, adapter, applier, appointment
31
edible fruit
edible reproductive body of a seed plant especially one having sweet flesh
fruit tree
tree bearing edible fruit
Example words: akee, amarelle, anchovy pear, apple
32
press
printed matter in the form of newspapers or magazines
publisher
a firm in the publishing business
Example words: magazine, newspaper, paper,
33
religious ceremony
structure
a ceremony having religious meaning
a thing constructed; a comple x construction or entity; "the structure
consisted of a series of arches"
Example words: chapel, church, office,
34
person
a human being; "there was too much for one person to do"
position
a job in an organization or hierarchy; "he ocupied a post in the treasury"
Example words: Agriculture Secretary, Chief Executive, Commerce Secretary, Defense Secretary
35
group
property
any number of entities (members) considered as a unit
a basic or essential attribute shared by all members of a class; "a study of the
physical properties of atomic particles"
Example words: ancestry, bloodline, body, complement
154
36
beverage
any liquid suitable for drinking: "may I take your beverage order?"
plant part
any part of a plant or fungus
Example words: chicory, chicory root, coffee
37
artifact
a man-made object
increase
the act of increasing something; "he gave me an increase in salary"
Example words: deposit, elevation, enlargement, extension
38
belief
any cognitive content held as true
message
what a communication that is about something is about
Example words: conclusion, gospel, opinion, precept
39
artifact
a man-made object
creation
the human act of creating
Example words: arrangement, art, assembly, avenue
40
substance
that which has mass and occupies space; "an atom is the smallest indivisible unit
of matter"
color
a visual attribute of things that results from the light they emit or transmit or
reflect
Example words: alabaster, apricot, aquamarine, burnt sienna
41
fabric
something made by weaving or felting or knitting or crocheting natural or
synthetic fibers
covering
a natural object that covers or envelops; "the fox was flushed from its cover"
Example words: fleece, hair, tapa, tappa
42
object a physical (tangible and visible) entity; "it was full of rackets
shape the spatial arrangement of something as distinct from its substance; "geometry is the
mathematical science of shape"
Example words: annulus, aperture, asterism, ball
43
person a human being; "there was too much for one person to do"
job
the occupation for which you are paid; "he is looking for a job"; "a lot of people are out
of work"
Example words: Agriculture Secretary, appointment, cabinet minister, Chief Executive
44
writing
message
reading matter; anything expressed in letters of the alphabet (especially when
considered from the point of view of style and effect); "the writing in her novels
is excellent"
what a communication that is about something is about
155
Example words: account, accounting, acrostic, assize
45
activity
any specific activity or pursuit; "they avoided all recreational activity"
trait
a distinguishing feature of one''s personal nature
Example words: abstinence, acceptance, aggression, animation
46
plant
a living organism lacking the power of locomotion
plant material material derived from plants
Example words: abaca, alder, amboyna, ash
47
product
the amount of an artifact that has been produced by someone or some process;
"they improve their product every year"; "they export most of their agricultural
production"
firm
members of a business organization; "he worked for a brokerage house"
Example words: magazine, newspaper, paper,
48
creation
the human act of creating
creation
something that has been brought into existence by someone
Example words: arrangement, art, carving, casting
49
life form
any living entity
garment
an article of clothing; "garments of the finest silk"
Example words: banian, banyan, boa, capuchin
50
plant
a living organism lacking the power of locomotion
edible fruit
edible reproductive body of a seed plant especially one having sweet flesh
Example words: acerola, akee, amarelle, anchovy pear
51
person
a human being; "there was too much for one person to do"
occupation
the principal activity in your life; "he''s not in my line of business"
Example words: Agriculture Secretary, appointment, cabinet minister, Chief Executive
52
speech act
the use of language to perform some act
social group
people sharing some social relation
Example words: colloquium, commission, defence, defense
53
food
any substance that can be metabolized by an organism to give energy and build
tissue
change of state the act of changing something into something different in essential characteristics
Example words: coloring, colouring, cookery, cooking
54
156
content
situation
the sum or range of what has been perceived
the general state of things; the combination of circumstances at a given time; "the
present international situation is dangerous"
Example words: area, background, crossroads, domain
55
possession
anything owned or possessed
distribution
the act of distributing or spreading or apportioning
Example words: allocation, allotment, assignment, contribution
56
organization
a group of people who work together
worker
a person who works at a specific job; "he is a good worker"
Example words: armorer, armourer, carrier, commando
57
combat
an engagement fought between two military forces
emotion
any strong feeling
Example words: aggression, belligerency, hostility,
58
beverage
any liquid suitable for drinking: "may I take your beverage order?"
vascular plant green plant having a vascular system: ferns
Example words: chicory, coffee, mate, sarsaparilla
59
ship
a vessel that carries passengers or freight
skilled worker a worker who has acquired special skills
Example words: Flying Dutchman, privateer, tanker, whaler
60
building
a structure that has a roof and walls and stands more or less permanently in one
place
organization
a group of people who work together
Example words: bucket shop, clinic, conservatory, day school
61
person
language
a human being; "there was too much for one person to do"
a systematic means of communicating by the use of sounds or conventional
symbols; "he taught foreign languages"; "the language introduced is standard
throughout the text"
Example words: Abnaki, Achomawi, Akwa'ala, Alabama
62
animal
a living organism characterized by voluntary movement
animal skin
the tissue forming the outer covering of a fur-bearing animal
Example words: broadtail, buckskin, calf, chamois
63
work
activity directed toward making or doing something; "she checked several points
needing further work"
157
structure
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"
Example words: bath, bodywork, office, shower bath
64
occupation
discipline
the principal activity in your life; "he''s not in my line of business"
a branch of knowledge; "in what discipline is his doctorate?"; "teachers should be
well trained in their subject"; "anthropology is the study of human beings"
Example words: architecture, interior design, law, literature
65
plant
root
a living organism lacking the power of locomotion
the usually underground organ that lacks buds or leaves or nodes; absorbs water and
mineral salts; usually it anchors the pla nt to the ground
Example words: calamus, carrot, cassava, chicory
66
person a human being; "there was too much for one person to do"
deed a notable achievement: "the book was her finest effort"
Example words: appointment, relief, contact, gun
67
passage
structure
a path or channel through or along which someone or something may pass
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"
Example words: aisle, arcade, column, flue
68
life form
any living entity
drug
something that is used as a medicine or narcotic
Example words: absinthe, arnica, belladonna, cannabis
69
act
something that people do or cause to happen
information
a message received and understood that reduces the recipient's uncertainty
Example words: account, bill, certification, confirmation
70
plant a living organism lacking the power of locomotion
seed a small hard fruit
Example words: algarroba, almond, amaranth, bean
71
musical composition
a musical work that has been created; "the composition is written in four
movements"
musical organization
an organization of musicians who perform together
Example words: duet, duette, duo, octet
72
person
conveyance
a human being; "there was too much for one person to do"
something that serves as a means of transportation
158
Example words: camper, carrier, cat, commuter
73
activity
any specific activity or pursuit; "they avoided all recreational activity"
knowledge domain
the content of a particular domain or field of knowledge
Example words: aeronautical engineering, analysis, anatomy, architecture
74
change of integrity
the act of changing the unity or wholeness of something
instrumentality
an artifact (or system of artifacts) that is instrumental in accomplishing
some end
Example words: attachment, chip, connection, connexion
75
action
something done (usually as opposed to something said); "there were stories of
murders and other unnatural actions"
movement
a natural event that involves a change in the position or location of something
Example words: ascent, commotion, deflection, deflexion
76
group
any number of entities (members) considered as a unit
political orientation
an orientation that characterizes the thinking of a group or nation
Example words: absolutism, autocracy, collectivism, communism
77
person
a human being; "there was too much for one person to do"
role
normal or customary activity; "what is your role on the team?"
Example words: catcher, center, end, field general
78
natural object an object occurring naturally; not made by man
location
a point or extent in space
Example words: aerie, Aquarius, Aries, axis
79
writing
reading matter; anything expressed in letters of the alphabet (especially when
considered from the point of view of style and effect); "the writing in her novels
is excellent"
speech act
the use of language to perform some act
Example words: version, words, acceptance, account
80
grain cereal grain suitable as food for human beings
cereal grass whose starchy grains are used as food: wheat; rice; rye; oats; maize; buckwheat;
millet
Example words: barley, corn, green corn, oat
81
dancing
music
taking a series of rhythmical steps (and movements) in time to music
an artistic form of auditory communication incorporating instrumental or vocal
tones in a structured and continuous manner
159
Example words: ballet, bolero, gavotte, hornpipe
82
communication
something that is communicated between people or groups
journal
a record book as a physical object
Example words: book, daybook, diary, ledger
83
animal
a living organism characterized by voluntary movement
food
any substance that can be metabolized by an organism to give energy and build
tissue
Example words: Alaska crab, Alaska king crab, Alaskan king crab, albacore
84
belief
any cognitive content held as true
social group
people sharing some social relation
Example words: absolutism, Brahmanism, Brahminism, Buddhism
85
life form
color
any living entity
a visual attribute of things that results from the light they emit or transmit or
reflect; "white is the coolest summer color"
Example words: apricot, canary, cardinal, cherry
86
attribute
an abstraction belonging to or characteristic of an entity
perception
the process of perceiving
Example words: acidity, acridity, aroma, astringency
87
plant
a living organism lacking the power of locomotion
beverage
any liquid suitable for drinking: "may I take your beverage order?"
Example words: chicory, coffee, mate, sarsaparilla
88
location
people
a point or extent in space
(plural) any group of human beings (men or women or children) collectively;
"old people"; "there were at least 200 people in the audience"
Example words: center, commons, country, domain
89
life form
any living entity
fabric
something made by weaving or felting or knitting or crocheting natural or
synthetic fibers
Example words: alpaca, bunting, chinchilla, cotton
90
object
a physical (tangible and visible) entity
gathering
a group of persons together in one place
Example words: hamlet, house, meeting, metropolis
160
91
life form
any living entity
produce
fresh fruits and vegetable grown for the market
Example words: acerola, acorn squash, akee, amarelle
92
location
magnitude
a point or extent in space
the property of relative size or extent; "they tried to predict the magnitude of the
explosion"
Example words: area, boundary, bounds, brink
93
sign of the zodiac
one of 12 equal areas into which the zodiac is divided
constellation
a configuration of stars as seen from the earth
Example words: Aquarius, Aries, Cancer, Capricorn
94
knowledge domain
record
the content of a particular domain or field of knowledge
anything (such as a document or a phonograph record or a photograph)
providing permanent evidence of or information about past events; "the
film provided a valuable record of stage techniques"
Example words: chronology, etymology, history
95
activity
ceremony
any specific activity or pursuit; "they avoided all recreational activity"
a formal event performed on a special occasion; "a ceremony commemorating
Pearl Harbor"
Example words: burial, graduation, military ceremony, wedding
96
communication
something that is communicated between people or groups
belief
any cognitive content held as true
Example words: conclusion, gospel, mimesis, opinion
97
artifact
beginning
a man-made object
the act of starting something; "he was responsible for the beginning of
negotiations"
Example words: creation, entry, establishment, first base
98
plant
a living organism lacking the power of locomotion
plant product a product made from plant material
Example words: abaca, bowstring hemp, coca, deccan hemp
99
unrighteousness
attitude
failure to adhere to moral principles
a complex mental orientation involving beliefs and feelings and values
and dispositions to act in certain ways; "he had the attitude that work was
fun"
Example words: godlessness, inequity, unfairness
161
100
location
a point or extent in space
time period
an indefinite length of time; "a time period of 30 years"
Example words: apogee, beginning, peak, terminus
101
feeling the psychological feature of experiencing affective and emotional states; "he had a
feeling of euphoria"
state the way something is with respect to its main attributes; "the current state of knowledge";
"his state of health"; "in a weak financial state"
Example words: addiction, agitation, antagonism, anticipation
102
language
a systematic means of communicating by the use of sounds or conventional
symbols; "he taught foreign languages"; "the language introduced is standard
throughout the text"
speech act
the use of language to perform some act
Example words: construction, interrogation, misconstruction, response
103
improvement
the act of improving something: "Their improvements increased the value of the
property"
motion
the act of changing your location from one place to another; "police controlled
the motion of the crowd”
Example words: advancement, airing, service,
104
social group
region
people sharing some social relation
a large indefinite location on the surface of the Earth; "penguins inhabit the polar
regions"
Example words: borough, circus, city, commonwealth
105
person
natural language
a human being; "there was too much for one person to do"
a human written or spoken language used by a community; opposed to
e.g. a computer language
Example words: Abnaki, Achomawi, Akwa''ala, Alabama
106
action something done (usually as opposed to something said); "there were stories of murders
and other unnatural actions"
way any road or path affording passage from one place to another; "he said he was looking for
the way out"
Example words: access, back door, concourse, crossing
107
measure
how much there is of something that you can measure
container
something that holds things
Example words: barrel, basin, basket, bath
162
108
person a human being; "there was too much for one person to do"
ability possession of the qualities (especially mental qualities) required to do something or get
something done; "danger heightened his powers of discrimination"
Example words: genius, hand, intellect, sensation
109
language
a systematic means of communicating by the use of sounds or
conventional symbols; "he taught foreign languages"; "the language
introduced is standard throughout the text"
administrative district a district defined for administrative purposes
Example words: Alabama, Cheyenne, Delaware, Illinois
110
container
something that holds things
containerful
the quantity that a container will hold
Example words: barrel, basin, basket, bowl
111
social control control exerted (actively or passively) by group action
social group
people sharing some social relation
Example words: administration, authority, delegation, deputation
112
substance
that which has mass and occupies space; "an atom is the smallest indivisible unit
of matter"
drug
something that is used as a medicine or narcotic
Example words: alcohol, alum, bromide, cider
113
instrumentality an artifact (or system of artifacts) that is instrumental in accomplishing some end
worker
a person who works at a specific job; "he is a good worker"
Example words: carrier, cat''s-paw, cleaner, cutter
114
work
activity directed toward making or doing something; "she checked several points needing
further work"
content the sum or range of what has been perceived
Example words: analogy, analysis, anatomy, attention
115
artifact a man-made object
shape the spatial arrangement of something as distinct from its substance; "geometry is the
mathematical science of shape"
Example words: aperture, ball, barrel, base
116
communication something that is communicated between people or groups
speech act
the use of language to perform some act
Example words: acceptance, account, affirmation, agreement
163
117
written communication communication by means of written symbols
message
what a communication that is about something is about
Example words: account, accounting, acrostic, application
118
structure
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"
organization
a group of people who work together
Example words: academy, bastion, battery, bucket shop
119
measure
structure
how much there is of something that you can measure
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches"
Example words: agora, bar, barn, barrow
120
person a human being; "there was too much for one person to do"
feeling the psychological feature of experiencing affective and emotional states; "he had a
feeling of euphoria"
Example words: relief, sensation, shadow, terror
121
object a physical (tangible and visible) entity; "it was full of rackets
color a visual attribute of things that results from the light they emit or transmit or reflect;
"white is the coolest summer color"
Example words: alabaster, apricot, aquamarine, buff
122
jewelry
an adornment (as a bracelet or ring or necklace) made of precious metals and set
with gems (or imitation gems)
crystal
a solid formed by the solidification of a chemical and having a highly regular
atomic structure
Example words: diamond, emerald, gem,
123
property
a basic or essential attribute shared by all members of a class; "a study of the
physical properties of atomic particles"
perception
the process of perceiving
Example words: acidity, acridity, aroma, bitterness
124
act
something that people do or cause to happen
appearance
outward or visible aspect of a person or thing
Example words: beguilement, coating, discoloration, discolouration
125
religion
a strong belief in a supernatural power or powers that control human destiny; "he
lost his faith but not his morality"
164
organization
a group of people who work together
Example words: Brahmanism, Brahminism, Buddhism, Christian Science
126
relation
hypothesis
an abstraction belonging to or characteristic of two entities or parts together
a concept that is not yet verified but that if true would explain certain facts or
phenomena
Example words: assumption, basis, conclusion, foundation
127
plant a living organism lacking the power of locomotion
food any substance that can be metabolized by an organism to give energy and build tissue
Example words: acerola, acorn squash, akee, alfalfa
128
plant a living organism lacking the power of locomotion
artifact a man-made object
Example words: absinthe, arnica, bay, belladonna
129
person a human being; "there was too much for one person to do"
device an instrumentality invented for a particular purpose; "the device is small enough to wear
on your wrist"; "a device intended to conserve water"
Example words: accumulator, adapter, applier, automaton
130
ingredient
herb
food that is a component of a mixture in cooking
a plant lacking a permanent woody stem; many are flowering garden plants or
potherbs; some having medicinal properties; some are pests
Example words: angelica, anise, basil, borage
131
communication
product
something that is communicated between people or groups
the amount of an artifact that has been produced by someone or some
process; "they improve their product every year"
Example words: annals, bible, blue book, book
132
content
the sum or range of what has been perceived
time period
an indefinite length of time; "a time period of 30 years"
Example words: beginning, conclusion, Immaculate Conception, life
133
structure
a thing constructed; a complex construction or entity; "the structure consisted of
a series of arches”
institution
an organization founded and united for a specific purpose
Example words: academy, bucket shop, clinic, college
134
book a book as a physical object: a number of pages bound together; "he used a large book as a
doorstop"
165
record a document that can serve as legal evidence of a transaction; "they could find no record
of the purchase"
Example words: book, daybook, journal, ledger
135
food
any substance that can be metabolized by an organism to give energy and build
tissue
artifact
a man-made object
Example words: banger, barbecue, bass, bear claw
136
creation
writing
something that has been brought into existence by someone
reading matter; anything expressed in letters of the alphabet (especially when
considered from the point of vie w of style and effect); "the writing in her novels
is excellent"
Example words: ballad, bible, book, canon
137
gathering
a group of persons together in one place
organization
a group of people who work together
Example words: army, business, chapter, Christianity
138
organization
region
a group of people who work together
a large indefinite location on the surface of the Earth; "penguins inhabit the polar
regions"
Example words: circus, commonwealth, community, country
166
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
act
something that
people do or
cause to
happen
ability
act
something that
people do or
cause to
happen
social
event
act
something that
people do or
cause to
happen
social
event
act
something that
people do or
cause to
happen
sum
activity
any specific
activity or
pursuit; "they
avoided all
recreational
activity"
any specific
activity or
pursuit; "they
avoided all
recreational
activity"
any specific
activity or
pursuit; "they
avoided all
recreational
activity
activity
Number
of
Total
Words
Number of
Hyper2 gloss Relation
for
Participating
which
Words
Relation
Holds
possession of achieve
23
41
the qualities
(especially
mental
qualities)
required to do
something or
get something
done; "danger
heightened his
powers of
discrimination"
an event
accompany
32
32
characteristic
of persons
forming groups
an event
require
characteristic
of persons
forming groups
Percentage of
Successful
Relations
between Word
Senses
0.56
1
32
32
1
4
8
0.5
20
22
0.9109
knowledge the content of deal with
domain
a particular
domain or field
of knowledge
20
22
0.9109
trait
57
57
1
a quantity of
change
money; "he
borrowed a
large sum";
"the amount
he had in cash
was
insufficient"
knowledge the content of concerned
domain
a particular
with
domain or field
of knowledge
a
perform
distinguishing
feature of
one''s personal
nature
167
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
Number
of
Total
Words
Number of
Relation
for
Participating
which
Words
Relation
Holds
Percentage of
Successful
Relations
between Word
Senses
activity"
artifact
a man-made
object
beginning
artifact
a man-made
object
a man-made
object
creation
artifact
artifact
beverage
container
creation
a man-made
creation
object
any liquid
plant part
suitable for
drinking: "may I
take your
beverage
order?"
something that containerful
holds things
decoration
something used structure
to beautify
edible fruit
edible
fruit tree
reproductive
body of a seed
plant especially
one having
sweet flesh
the act of
form
starting
something; "he
was
responsible for
the beginning
of
negotiations"
the human act construct
of creating
the human act create
of creating
8
10
0.8
35
40
0.875
35
40
0.875
the human act make
of creating
any part of a
used as
plant or fungus
35
40
0.875
3
3
1
the quantity
contain
that a
container will
hold
a thing
made of
constructed; a
complex
construction or
entity; "the
structure
consisted of a
series of
arches"; "she
wore her hair
in an amazing
construction of
whirls and
ribbons"
tree bearing
bear
edible fruit
19
19
1
4
9
0.4444
88
88
1
168
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
edible fruit
Hyper1 gloss
Hyper2
edible
reproductive
body of a seed
plant especially
one having
sweet flesh
edible
reproductive
body of a seed
plant especially
one having
sweet flesh
edible
reproductive
body of a seed
plant especially
one having
sweet flesh
a contest with
rules to
determine a
winner; "you
need four
people to play
this game"
cereal grain
suitable as food
for human
beings
fruit tree
consumer
goods
goods (as food
or clothing)
intended for
direct use or
consumption
life form
life form
any living entity drug
garment
an article of
clothing;
"garments of
edible fruit
edible fruit
game
grain
Number
of
Total
Words
Number of
Hyper2 gloss Relation
for
Participating
which
Words
Relation
Holds
tree bearing
cultivated
88
88
edible fruit
for
Percentage of
Successful
Relations
between Word
Senses
1
fruit tree
tree bearing
edible fruit
have
88
88
1
fruit tree
tree bearing
edible fruit
produce
88
88
1
equipment
an artifact
used in
needed for an
undertaking or
to perform a
service
6
6
1
cereal
grass whose
grown for
starchy grains
are used as
food: wheat;
rice; rye; oats;
maize;
buckwheat;
millet
any living
made of
entity
8
9
0.89
6
12
0.5
something that be
is used as a
medicine or
narcotic
any living
made of
entity
20
31
0.65
8
8
1
life form
169
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
Number
of
Total
Words
Number of
Relation
for
Participating
which
Words
Relation
Holds
Percentage of
Successful
Relations
between Word
Senses
the finest silk"
measure
message
message
object
object
how much there container
is of something
that you can
measure
what a
speech act
communication
that is about
something is
about
something that contained
holds things
in
24
24
1
the use of
language to
perform some
act
52
52
1
52
52
1
40
40
1
40
40
1
8
11
0.73
12
12
1
convey
what a
speech act the use of
make
communication
language to
that is about
perform some
something is
act
about
a physical
color
a visual
be
(tangible and
attribute of
visible) entity; "it
things that
was full of
results from
rackets
the light they
emit or
transmit or
reflect; "white
is the coolest
summer color"
a physical
color
a visual
have
(tangible and
attribute of
visible) entity; "it
things that
was full of
results from
rackets
the light they
emit or
transmit or
reflect; "white
is the coolest
summer color"
person
a human being; deed
"there was too
much for one
person to do"
person
a human being; position
"there was too
much for one
person to do"
a notable
accomplish
achievement:
"the book was
her finest
effort"
(in team
play
sports) the role
assigned to an
individual
player; "what
170
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
plant
a living
organism
lacking the
power of
locomotion
beverage
plant
a living
organism
lacking the
power of
locomotion
beverage
plant material
material derived plant
from plants
plant product
a product made plant
from plant
material
seed
a small hard
fruit
skilled worker
worker who has ship
acquired special
skills
social control
control exerted social
(actively or
group
passively) by
group action
that which has color
mass and
occupies space;
"an atom is the
smallest
indivisible unit
of matter"
substance
plant
Hyper2 gloss
position does
he play?"
any liquid
suitable for
drinking: "may
I take your
beverage
order?"
any liquid
suitable for
drinking: "may
I take your
beverage
order?"
a living
organism
lacking the
power of
locomotion
a living
organism
lacking the
power of
locomotion
a living
organism
lacking the
power of
locomotion
Number
of
Total
Words
Number of
Relation
for
Participating
which
Words
Relation
Holds
be
5
5
1
make by
steeping
5
5
1
76
76
1
obtained
from
9
9
1
produced
by
34
34
1
4
4
1
13
13
1
33
33
1
obtain from
a vessel that
work on
carries
passengers or
freight
people sharing be
some social
relation
a visual
attribute of
things that
results from
the light they
emit or
transmit or
reflect; "white
Percentage of
Successful
Relations
between Word
Senses
be
171
Appendix C: Success Rate of Relations Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
substance
that which has color
mass and
occupies space;
"an atom is the
smallest
indivisible unit
of matter"
substance
that which has drug
mass and
occupies space;
"an atom is the
smallest
indivisible unit
of matter"
something that substance
is used as a
medicine or
narcotic
drug
written
communication message
communication by means of
written symbols
writing
Hyper2 gloss
Number
of
Total
Words
Number of
Relation
for
Participating
which
Words
Relation
Holds
is the coolest
summer color"
a visual
have
attribute of
things that
results from
the light they
emit or
transmit or
reflect; "white
is the coolest
summer color"
something that be
is used as a
medicine or
narcotic
that which has made from
mass and
occupies
space; "an
atom is the
smallest
indivisible unit
of matter"
what a
express
communication
that is about
something is
about
reading matter; speech act the use of
anything
language to
expressed in
perform some
letters of the
act
alphabet
(especially
when
considered from
the point of view
of style and
effect); "the
writing in her
novels is
excellent"
perform
Percentage of
Successful
Relations
between Word
Senses
33
33
1
25
25
1
25
25
1
39
43
0.96
16
18
0.89
172
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
achieve
convey
18
accompany
32
require
32
Relation
act
something that
people do or
cause to happen
ability
act
something that
people do or
cause to happen
information
act
something that
people do or
cause to happen
social event
act
something that
people do or
cause to happen
social event
act
something that
people do or
cause to happen
sum
a quantity of
change
money; "he
borrowed a large
sum"; "the amount
he had in cash was
insufficient"
8
movement
a natural event
that involves a
change in the
position or
location of
something
action
29
activity
any specific
activity or pursuit;
"they avoided all
recreational
activity"
any specific
activity or pursuit;
"they avoided all
recreational
activity"
knowledge
domain
something done
involve
(usually as
opposed to
something said);
"there were stories
of murders and
other unnatural
actions"
the content of a
concerned
particular domain or with
field of knowledge
the content of a
deal with
particular domain or
field of knowledge
22
activity
knowledge
domain
possession of the
qualities (especially
mental qualities)
required to do
something or get
something done;
"danger heightened
his powers of
discrimination"
a message
received and
understood that
reduces the
recipient''s
uncertainty
an event
characteristic of
persons forming
groups
an event
characteristic of
persons forming
groups
Total
Number of
Participating
Words
41
22
173
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
trait
a distinguishing
feature of one''s
personal nature
activity
animal
a living organism
characterized by
voluntary
movement
food
artifact
a man-made
object
beginning
creation
the human act of
creating
the human act of
creating
the human act of
creating
a man-made
object
creation
creation
artifact
plant part
Hyper2 gloss
Relation
any specific activity perform
or pursuit; "they
avoided all
recreational
activity"
any substance that be
can be metabolized
by an organism to
give energy and
build tissue
Total
Number of
Participating
Words
57
172
10
artifact
the act of starting
form
something; "he was
responsible for the
beginning of
negotiations"
a man-made object construct
artifact
a man-made object create
40
artifact
a man-made object make
40
shape
the spatial
arrangement of
something as
distinct from its
substance;
"geometry is the
mathematical
science of shape"
resemble
73
used as
3
made by
steeping
5
convey
77
any part of a plant beverage
or fungus
any liquid suitable
for drinking: "may I
take your beverage
order?"
beverage
any liquid suitable vascular plant green plant having
for drinking: "may
a vascular system:
I take your
ferns
beverage order?"
communication something that is speech act
the use of language
communicated
to perform some
between people
act
or groups
communication something that is speech act
communicated
between people
or groups
container
something that
containerful
holds things
40
the use of language make
to perform some
act
77
the quantity that a
container will hold
19
contain
174
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
Relation
creation
the human act of
creating
creation
something that has create
been brought into
existence by
someone
decoration
something used
to beautify
structure
fruit tree
tree bearing
edible fruit
edible fruit
fruit tree
tree bearing
edible fruit
edible fruit
fruit tree
tree bearing
edible fruit
edible fruit
a thing constructed;
a complex
construction or
entity; "the structure
consisted of a
series of arches";
"she wore her hair
in an amazing
construction of
whirls and ribbons"
edible reproductive
body of a seed
plant especially one
having sweet flesh
edible reproductive
body of a seed
plant especially one
having sweet flesh
edible reproductive
body of a seed
plant especially one
having sweet flesh
fruit tree
tree bearing
edible fruit
edible fruit
covering
a natural object
fabric
that covers or
envelops; "the fox
was flushed from
its cover"
equipment
an artifact needed game
for an undertaking
or to perform a
service
grain
cereal grain
cereal
suitable as food
for human beings
group
any number of
property
Total
Number of
Participating
Words
24
made of
9
bear
88
cultivated
for
88
have
88
edible reproductive produce
body of a seed
plant especially one
having sweet flesh
something made by made from
weaving or felting
or knitting or
crocheting natural
or synthetic fibers
88
a contest with rules used in
to determine a
winner; "you need
four people to play
this game"
grass whose
grown for
starchy grains are
used as food:
wheat; rice; rye;
oats; maize;
buckwheat; millet
a basic or essential have
6
5
9
22
175
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
entities
(members)
considered as a
unit
Hyper2 gloss
Relation
attribute shared by
all members of a
class; "a study of
the physical
properties of atomic
particles"
any living entity
made of
Total
Number of
Participating
Words
consumer
goods
goods (as food or life form
clothing) intended
for direct use or
consumption
life form
any living entity
drug
something that is
be
used as a medicine
or narcotic
31
garment
life form
any living entity
8
life form
an article of
clothing;
"garments of the
finest silk"
any living entity
produce
314
life form
any living entity
produce
fresh fruits and
be
vegetable grown for
the market
fresh fruits and
bear
vegetable grown for
the market
life form
any living entity
produce
fresh fruits and
cultivateed
vegetable grown for for
the market
314
life form
any living entity
produce
fresh fruits and
produce
vegetable grown for
the market
314
life form
any living entity
produce
fresh fruits and
yield
vegetable grown for
the market
314
measure
how much there
is of something
that you can
measure
what a
communication
that is about
something is
about
container
something that
holds things
speech act
the use of language convey
to perform some
act
52
what a
communication
that is about
something is
about
a physical
speech act
the use of language make
to perform some
act
52
color
a visual attribute of
40
message
message
object
made of
contained
in
be
12
314
24
176
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
(tangible and
visible) entity; "it
was full of rackets
object
a physical
color
(tangible and
visible) entity; "it
was full of rackets
object
a physical
shape
(tangible and
visible) entity; "it
was full of rackets
object
a physical
shape
(tangible and
visible) entity; "it
was full of rackets
occupation
the principal
activity in your
life; "he''s not in
my line of
business"
passage
a path or channel structure
through or along
which someone
or something may
pass
discipline
Hyper2 gloss
Relation
things that results
from the light they
emit or transmit or
reflect; "white is the
coolest summer
color"
a visual attribute of have
things that results
from the light they
emit or transmit or
reflect; "white is the
coolest summer
color"
the spatial
arrangement of
something as
distinct from its
substance;
"geometry is the
mathematical
science of shape"
the spatial
arrangement of
something as
distinct from its
substance;
"geometry is the
mathematical
science of shape"
Total
Number of
Participating
Words
40
consist of
101
resemble
101
a branch of
concerned
knowledge; "in what with
discipline is his
doctorate?";
"teachers should be
well trained in their
subject";
"anthropology is the
study of human
beings"
a thing constructed; used in
a complex
construction or
entity; "the structure
consisted of a
series of arches";
"she wore her hair
in an amazing
construction of
whirls and ribbons"
6
8
177
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
person
a human being;
"there was too
much for one
person to do"
deed
person
a human being;
"there was too
much for one
person to do"
use
person
a human being;
"there was too
much for one
person to do"
person
a human being;
"there was too
much for one
person to do"
instrumentality an artifact (or
system of artifacts)
that is instrumental
in accomplishing
some end
job
the occupation for
which you are paid;
"he is looking for a
job"; "a lot of people
are out of work"
job
the occupation for
which you are paid;
"he is looking for a
job"; "a lot of people
are out of work"
person
a human being;
"there was too
much for one
person to do"
language
speak
257
person
a human being;
"there was too
much for one
person to do"
natural
language
speak
252
person
a human being;
"there was too
much for one
person to do"
a human being;
"there was too
occupation
person
occupation
a notable
achievement: "the
book was her finest
effort"
Total
Number of
Relation
Participating
Words
accomplish
11
a systematic means
of communicating
by the use of
sounds or
conventional
symbols; "he taught
foreign languages";
"the language
introduced is
standard
throughout the
text"; "the speed
with which a
program can be
executed depends
on the language in
which it
a human written or
spoken language
used by a
community;
opposed to e.g. a
computer language
the principal activity
in your life; "he''s
not in my line of
business"
the principal activity
in your life; "he''s
159
be
appointed
27
hold
27
be
appointed
28
hold
28
178
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
person
person
Hyper1 gloss
much for one
person to do"
a human being;
"there was too
much for one
person to do"
a human being;
"there was too
much for one
person to do"
Hyper2
organization
34
a job in an
hold
organization or
hierarchy; "he
ocupied a post in
the treasury"
(in team sports) the play
role assigned to an
individual player;
"what position does
he play?"
23
normal or
play
customary activity;
"what is your role
on the team?"
people sharing
enroll in
some social relation
12
social group
people sharing
united in
some social relation
43
beverage
any liquid suitable
for drinking: "may I
take your beverage
order?"
be
5
any liquid suitable
for drinking: "may I
take your beverage
order?"
edible reproductive
body of a seed
plant especially one
having sweet flesh
edible reproductive
body of a seed
plant especially one
having sweet flesh
make by
steeping
5
position
a human being;
"there was too
much for one
person to do"
position
person
a human being;
"there was too
much for one
person to do"
a human being;
"there was too
much for one
person to do"
a human being;
"there was too
much for one
person to do"
a living organism
lacking the power
of locomotion
role
person
plant
not in my line of
business"
a group of people
who work together
Relation
build
person
person
Hyper2 gloss
Total
Number of
Participating
Words
social group
plant
a living organism beverage
lacking the power
of locomotion
plant
a living organism edible fruit
lacking the power
of locomotion
plant
a living organism edible fruit
lacking the power
of locomotion
plant
a living organism edible fruit
lacking the power
of locomotion
12
43
bear
159
cultivated
for
159
edible reproductive have
body of a seed
plant especially one
159
179
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
Hyper2 gloss
Relation
Total
Number of
Participating
Words
having sweet flesh
plant
a living organism edible fruit
lacking the power
of locomotion
plant
a living organism edible fruit
lacking the power
of locomotion
plant
a living organism food
lacking the power
of locomotion
plant
a living organism food
lacking the power
of locomotion
plant
a living organism food
lacking the power
of locomotion
plant
a living organism food
lacking the power
of locomotion
plant
a living organism food
lacking the power
of locomotion
plant material
material derived
from plants
plant
plant product
a product made
from plant
material
a small hard fruit
plant
an occupation
requiring special
education
(especially in the
discipline
seed
profession
plant
edible reproductive
body of a seed
plant especially one
having sweet flesh
edible reproductive
body of a seed
plant especially one
having sweet flesh
any substance that
can be metabolized
by an organism to
give energy and
build tissue
any substance that
can be metabolized
by an organism to
give energy and
build tissue
produce
159
yield
159
be
393
bear
393
any substance that cultivate for
can be metabolized
by an organism to
give energy and
build tissue
any substance that produce
can be metabolized
by an organism to
give energy and
build tissue
393
any substance that
can be metabolized
by an organism to
give energy and
build tissue
a living organism
lacking the power
of locomotion
a living organism
lacking the power
of locomotion
a living organism
lacking the power
of locomotion
a branch of
knowledge; "in what
discipline is his
doctorate?";
393
yield
393
obtained
from
76
obtained
from
9
produced
by
34
concerned
with
5
180
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
liberal arts or
sciences)
skilled worker
social control
substance
substance
substance
drug
writing
a worker who has ship
acquired special
skills
control exerted
social group
(actively or
passively) by
group action
that which has
mass and
occupies space;
"an atom is the
smallest
indivisible unit of
matter"
that which has
mass and
occupies space;
"an atom is the
smallest
indivisible unit of
matter"
that which has
mass and
occupies space;
"an atom is the
smallest
indivisible unit of
matter"
something that is
used as a
medicine or
narcotic
color
reading matter;
anything
expressed in
letters of the
alphabet
(especially when
considered from
the point of view
message
color
drug
substance
Hyper2 gloss
Relation
"teachers should be
well trained in their
subject";
"anthropology is the
study of human
beings"
a vessel that
work on
carries passengers
or freight
people sharing
be
some social relation
Total
Number of
Participating
Words
4
13
a visual attribute of be
things that results
from the light they
emit or transmit or
reflect; "white is the
coolest summer
color"
a visual attribute of have
things that results
from the light they
emit or transmit or
reflect; "white is the
coolest summer
color"
something that is
be
used as a medicine
or narcotic
33
that which has
made from
mass and occupies
space; "an atom is
the smallest
indivisible unit of
matter"
what a
express
communication that
is about something
is about
25
33
25
36
181
Appendix D: Relations between Word Senses Extracted in Section 9.1
Hyper1
Hyper1 gloss
Hyper2
of style and
effect); "the
writing in her
novels is
excellent"
written
communication by message
communication means of written
symbols
Hyper2 gloss
Relation
what a
express
communication that
is about something
is about
Total
Number of
Participating
Words
43
182
Appendix E: Words Subsumed by the Hypernym Pairs in Table 9.3
word
sense_number1 gloss1
sense_number2
accompaniment 2
a subordinate musical
3
part; provides
background for more
important parts
Bach
1
(1685-1750) German
composer and organist
2
ballet
1
bass
2
a theatrical
2
representation of a story
performed to music by
ballet dancers
the lowest part in
3
polyphonic music
Beethoven
1
(1770-1827) German
composer
2
bolero
1
music written in the
rhythm of the bolero
dance
3
Brahms
1
(1833-1897) German
composer
2
Chopin
1
gavotte
1
the music of Chopin; "he 2
practiced Chopin day and
night"
an old formal French
2
dance in quadruple time
Handel
1
Hare Krishna
1
harmonization
1
Haydn
hornpipe
1
1
landler
1
(1685-1759) a prolific
2
German composer noted
for his oratorios
worshipper of Krishna
3
a piece of harmonized
music
1732-1809
a British solo dance
performed by sailors
music in triple time for
dancing the landler
2
2
2
2
gloss2
the act of
accompanying
someone or
something in order
to protect them
the music of Bach;
"he played Bach on
the organ"
music written for a
ballet
an adult male singer
with the lowest
voice
the music of
Beethoven; "he
enjoyed Beethoven
most of all
a Spanish dance in
triple time;
accompanied by
guitar and castanets
the music of
Brahms; "Brahms
was included in the
program"
(1810-1849)
composer and
pianist
music composed in
quadruple time for
dancing the gavotte
the music of Handel
a chant to the Hindu
god Krishna
singing in harmony
the music of Haydn
music for dancing
the hornpipe
a moderately slow
Austrian country
183
Appendix E: Words Subsumed by the Hypernym Pairs in Table 9.3
word
sense_number1 gloss1
sense_number2 gloss2
dance in triple time;
involves spinning
and clapping
line
19
the principal activity in
26
a succession of
your life; "he''s not in my
notes forming a
line of business"
distinctive sequence;
"she was humming
an air from
Beethoven"
mazurka
1
music composed for
2
a Polish national
dancing the mazurka
dance in triple time
minuet
1
a stately court dance in 2
a stately piece of
the 17th century
music composed for
dancing the minuet;
often incorporated
into a sonata or suite
Mozart
1
1756-1791
2
the music of Mozart;
"the concert was
mostly Mozart"
part
8
an actor''s portrayal of 11
the melody carried
someone in a play; "she
by a particular voice
played the part of
or instrument in
Desdemona"
polyphonic music;
"he tried to sing the
tenor part"
part
10
any one of a number of 11
the melody carried
individual efforts in a
by a particular voice
common endeavor: "I am
or instrument in
proud of my contribution
polyphonic music;
to the team''s success";
"he tried to sing the
"they all did their share
tenor part"
of the work"
part
6
the actions and activities 11
the melody carried
assigned to or required or
by a particular voice
expected of a person or
or instrument in
group: "the function of a
polyphonic music;
teacher"; "the
"he tried to sing the
government must do its
tenor part"
part" or "play its role" or
"do its duty"
paso doble
1
a ballroom dance in fast 2
music in march time
duple time
composed for
dancing the paso
doble; often played
a bull fights
pavane
1
music composed for
2
a stately court dance
184
Appendix E: Words Subsumed by the Hypernym Pairs in Table 9.3
word
sense_number1 gloss1
sense_number2 gloss2
dancing the pavane
of the 16th and 17th
centuries
polka
1
music performed for
2
a Bohemian dance
dancing the polka
with 3 steps and a
hop in fast time
quadrille
1
music for dancing the
2
a square dance of 5
quadrille
or more figures for 4
or more couples
rumba
1
syncopated music in
2
a rhythmic
duple time for dancing
syncopated Cuban
the rumba
folk dance in duple
time
rumba
1
syncopated music in
3
a ballroom dance
duple time for dancing
based on the Cuban
the rumba
folk dance
samba
2
music composed for
4
a form of canasta
dancing the samba
using three decks
and six jokers
samba
2
music composed for
3
a lively ballroom
dancing the samba
dance from Brazil
saraband
1
music composed for
2
a stately court dance
dancing the saraband
of the 17th and 18th
centuries; in slow
time
schottische
1
music performed for
2
a German round
dancing the schottische
dance resembling a
slow polka
Stravinsky
1
1882-1971
2
the music of
Stravinsky;
"Stravinsky no
longer causes riots
in the streets"
tarantella
1
music composed in six- 2
a lively whirling
eight time for dancing
Italian dance for two
the tarantella
persons
variation
2
an activity that varies
3
a repetition of a
from a norm or standard;
musical theme in
"any variation in his
which it is modified
routine was immediately
or embellished
reported"
voice
8
an advocate who
11
the melody carried
represents someone
by a particular voice
else''s policy or purpose;
or instrument in
"the meeting was
polyphonic music;
attended by
"he tried to sing the
185
Appendix E: Words Subsumed by the Hypernym Pairs in Table 9.3
word
sense_number1 gloss1
sense_number2
spokespersons for all the
major organs of
government"
voice
7
a singer; "he wanted to 11
hear trained voices sing
it"
Wagner
1
(1813-1883) German
composer
2
gloss2
tenor part"
the melody carried
by a particular voice
or instrument in
polyphonic music;
"he tried to sing the
tenor part"
the music of
Wagner; "they say
that Hitler listened
only to Wagner"
186
Appendix F: Evaluation of CASS Derived Relations Between Gold Standard Hypernym
Pairs
Hypernym1
Hypernym2
Relation Frequency
Valid/Invalid
artifact
person
plant
plant
life form
person
creation
artifact
person
property
attribute
person
person
person
person
plant
activity
passage
person
person
act
act
person
food
person
person
instrumentality
person
person
person
life form
location
life form
action
life form
language
location
object
person
object
person
shape
instrumentality
edible fruit
food
produce
occupation
creation
creation
occupation
perception
perception
job
job
job
occupation
food
knowledge domain
structure
position
role
social event
social event
occupation
artifact
position
position
worker
job
position
position
produce
magnitude
color
way
drug
administrative district
magnitude
aspect
ability
shape
deed
be
speak
be
be
be
catch
show
show
unethical
do
do
unethical
drive
catch
drive
ripen
receive
reduce
unethical
unethical
have
be
have
available
catch
drive
available
have
unethical
have
ripen
happen
happen
hit
assimilate
make
extend
happen
strive
happen
unethical
28
28
28
28
28
26
26
26
26
25
25
25
25
25
25
24
23
23
23
23
22
22
21
21
21
21
21
21
21
21
20
17
17
17
16
16
15
14
14
14
14
0
0
0
1
1
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
1
0
0
1
0
1
0
0
0
0
1
0
0
0
0
0
0
187
natural object
plant
object
extremity
plant
natural object
plant
written communication
writing
work
beverage
object
control
profession
gathering
person
plant
person
person
measure
object
act
natural object
writing
message
message
speech act
natural object
social group
person
organization
extremity
person
person
content
animal
ship
location
location
container
creation
location
edible fruit
shape
line
plant material
location
food
message
message
content
vascular plant
shape
trait
discipline
organization
device
beverage
quality
social group
structure
shape
social event
location
speech act
speech act
speech act
social group
location
region
organization
region
line
organization
social group
time period
food
skilled worker
magnitude
time period
containerful
commercial enterprise
happen
spur
assume
starboard
spur
be
spur
have
have
receive
assimilate
blunt
receive
receive
confront
be
assimilate
have
have
be
point
take
point
have
control
have
separate
cut
sleep
have
sleep
blunt
issue
issue
relate
find
upset
go
have
hold
hilarious
60
60
61
60
60
60
60
51
51
50
50
50
50
50
50
50
50
51
50
50
50
50
50
41
40
41
40
40
40
40
40
40
40
40
40
41
40
40
40
41
40
188
plant
person
activity
life
form
act
person
activity
person
communication
person
plant
animal
content
plant
person
animal
activity
plant
act
person
activity
person
communication
person
plant
content
life
form
personfruit
edible
activityfruit
edible
act
plant
activity
plant
communication
animal
plant
content
flavorer
ingredient
object
instrumentality
object
person
ingredient
object
flavorer
substance
substance
plant
animal
animal
substance
ingredient
flavorer
food
quality assimilate issue
450 1 4
produce ceremony
assimilate be
353 1 4
instrumentality
appearancennnnn
upset
cause
266 0 4
language ceremony
make
have
252 1 4
natural language
speech make
act
control
246 1 4
food
food assimilate leave
225 1 4
edible fruittime period
assimilate hold
185 0 4
food
quality buy
issue
149 0 4
food
ceremony
yield
be
133 1 4
instrumentality
appearance
available cause
129 0 4
device ceremony
upset
have
129 0 4
device speech available
act
control
129 0 4
edible fruitfood yield
leave
118 1 4
produce time period
yield
hold
118 1 4
fruit tree quality yield
issue
98 0 4
fruit tree ceremony
assimilate be
98 0 4
plant material
appearance
yield
cause
76 1 4
plant material
ceremony
pinnate have
75 0 4
food
speech carp
act
control
64 0 4
food
food erect
leave
60 0 4
herb
time period
erect
hold
58 0 4
herb
erect
58 0
shape
be
52 0
worker
upset
48 0
color
happen 47 0
conveyance
upset
40 0
herb
assimilate 38 0
color
have
38 1
herb
assimilate 38 0
color
have
36 1
color
happen 36 0
food
rush
34 0
food
aquatic
33 0
food
quail
32 0
drug
mind-alter 29 0
herb
rush
29 0
herb
rush
29 0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
189
APPENDIX G: Salience Scores for Gold Standard Patterns
Original score\normalized score from section 9.2.4
creation
ability
the human act of creating
possession of the qualities (especially mental qualities) required to do something or
get something done; "danger heightened his powers of discrimination"
Activation Score: 2.89\0.07
person
feeling
a human being; "there was too much for one person to do"
the psychological feature of experiencing affective and emotional states; "he had a
feeling of euphoria"
Activation Score: 2.99\0.09
location
a point or extent in space
time period
an indefinite length of time; "a time period of 30 years"; "hastened the period of his
recovery"
Activation Score: 3.17\0.12
combat
an engagement fought between two military forces
emotion
any strong feeling
Activation Score: 3.17\0.12
religious ceremony
a ceremony having religious meaning
structure
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches";
"she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 3.17\0.12
unrighteousness
attitude
failure to adhere to moral principles
a complex mental orientation involving beliefs and feelings and values and
dispositions to act in certain ways; "he had the attitude that work was fun"
Activation Score: 3.17\0.12
speech act
the use of language to perform some act
social group
people sharing some social relation
Activation Score: 3.17\0.12
improvement
the act of improving something: "Their improvements increased the value of the
property"
motion
the act of changing your location from one place to another; "police controlled the
motion of the crowd"; "the movement of people from the farms to the cities"; "his move
put him directly in my path"
Activation Score: 3.29\0.14
beverage
any liquid suitable for drinking: "may I take your beverage order?"
plant part
any part of a plant or fungus
Activation Score: 3.29\0.14
life form
any living entity
fabric
something made by weaving or felting or knitting or crocheting natural or synthetic fibers
Activation Score: 3.40\0.16
act
sum
something that people do or cause to happen
a quantity of money; "he borrowed a large sum"; "the amount he had in cash was
insufficient"
190
Activation Score: 3.46\0.17
creation
commercial enterprise
the human act of creating
the activity of providing goods and services involving financial and
commercial and industrial aspects; "computers are now widely used in business"
Activation Score: 3.46\0.17
jewelry
an adornment (as a bracelet or ring or necklace) made of precious metals and set with
gems (or imitation gems)
crystal
a solid formed by the solidification of a chemical and having a highly regular atomic
structure
Activation Score: 3.49\0.18
product
the amount of an artifact that has been produced by someone or some process; "they
improve their product every year"; "they export most of their agricultural production"
firm
members of a business organization; "he worked for a brokerage house"
Activation Score: 3.49\0.18
location
people
a point or extent in space
(plural) any group of human beings (men or women or children) collectively; "old
people"; "there were at least 200 people in the audience"
Activation Score: 3.55\0.19
activity
ceremony
any specific activity or pursuit; "they avoided all recreational activity"
a formal event performed on a special occasion; "a ceremony
commemorating Pearl Harbor"
Activation Score: 3.58\0.19
person
ability
a human being; "there was too much for one person to do"
possession of the qualities (especially mental qualities) required to do something or
get something
done; "danger heightened his powers of discrimination"
Activation Score: 3.58\0.19
knowledge domain
record
the content of a particular domain or field of knowledge
anything (such as a document or a phonograph record or a photograph)
providing permanent evidence of or information about past events; "the film
provided a valuable record of stage techniques"
Activation Score: 3.58\0.19
extremity
the outermost or farthest region or point
line
a length (straight or curved) without breadth or thickness; the trace of a moving point
Activation Score: 3.58\0.19
fabric
something made by weaving or felting or knitting or crocheting natural or synthetic
fibers
covering
a natural object that covers or envelops; "the fox was flushed from its cover"
Activation Score: 3.68\0.21
plant
a living organism lacking the power of locomotion
beverage
any liquid suitable for drinking: "may I take your beverage order?"
Activation Score: 3.68\0.21
location
job
a point or extent in space
the occupation for which you are paid; "he is looking for a job"; "a lot of people are
out of work"
191
Activation Score: 3.73\0.22
possession
anything owned or possessed
distribution
the act of distributing or spreading or apportioning
Activation Score: 3.73\0.22
control
the activity of managing or exerting control over something; "the control of the mob
by the police
was admirable"
trait
a distinguishing feature of one''s personal nature
Activation Score: 3.73\0.22
content
time period
the sum or range of what has been perceived
an indefinite length of time; "a time period of 30 years"; "hastened the period of his
recovery"
Activation Score: 3.73\0.22
language
a systematic means of communicating by the use of sounds or conventional symbols;
"he taught foreign languages"; "the language introduced is standard throughout the text
speech act
the use of language to perform some act
Activation Score: 3.73\0.22
communication something that is communicated between people or groups
journal
a record book as a physical object
Activation Score: 3.78\0.23
profession
discipline
an occupation requiring special education (especially in the liberal arts or sciences)
a branch of knowledge; "in what discipline is his doctorate?"; "teachers
should be well trained in their subject"; "anthropology is the study of human beings"
Activation Score: 3.80\0.23
press
printed matter in the form of newspapers or magazines
publisher
a firm in the publishing business
Activation Score: 3.80\0.23
beverage
any liquid suitable for drinking: "may I take your beverage order?"
vascular plant
green plant having a vascular system: ferns
Activation Score: 3.80\0.23
relation
an abstraction belonging to or characteristic of two entities or parts together
hypothesis
a concept that is not yet verified but that if true would explain certain facts or
phenomena; "he
proposed a fresh theory of alkalis that later was accepted in chemical practices"
Activation Score: 3.87\0.24
change
the act of changing something; "the change of government had no impact on the
economy"; "his
change on abortion cost him the election"
interpretation
an explanation resulting from interpreting something
Activation Score: 3.87\0.24
game
a contest with rules to determine a winner; "you need four people to play this game"
equipment
an artifact needed for an undertaking or to perform a service
Activation Score: 3.87\0.24
occupation
the principal activity in your life; "he''s not in my line of business"
192
discipline
a branch of knowledge; "in what discipline is his doctorate?"; "teachers should be
well trained in their subject"; "anthropology is the study of human beings"
Activation Score: 3.87\0.24
book
a book as a physical object: a number of pages bound together; "he used a large book
as a doorstop"
record
a document that can serve as legal evidence of a transaction; "they could find no
record of the purchase"
Activation Score: 3.87\0.24
work
activity directed toward making or doing something; "she checked several points
needing further work"
social group
people sharing some social relation
Activation Score: 3.91\0.25
work
activity directed toward making or doing something; "she checked several points
needing further work"
content
the sum or range of what has been perceived
Activation Score: 3.98\0.27
belief
any cognitive content held as true
message
what a communication that is about something is about
Activation Score: 3.98\0.27
activity
any specific activity or pursuit; "they avoided all recreational activity"
place of business
an establishment where business is conducted
Activation Score: 4.02\0.27
person
a human being; "there was too much for one person to do"
weaponry
weapons considered collectively
Activation Score: 4.02\0.27
ship
a vessel that carries passengers or freight
skilled worker
a worker who has acquired special skills
Activation Score: 4.02\0.27
artifact
a man-made object
increase
the act of increasing something; "he gave me an increase in salary"
Activation Score: 4.02\0.27
act
something that people do or cause to happen
appearance
outward or visible aspect of a person or thing
Activation Score: 4.09\0.28
group
any number of entities (members) considered as a unit
political orientation
an orientation that characterizes the thinking of a group or nation
Activation Score: 4.09\0.28
object
a physical (tangible and visible) entity; "it was full of rackets
aspect
a distinct feature or element in a problem; "he studied every facet of the question"
Activation Score: 4.09\0.28
work
further
structure
activity directed toward making or doing something; "she checked several points needing
work"
a thing constructed; a complex construction or entity; "the structure consisted of a
193
series of arches";
"she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 4.14\0.29
life form
any living entity
garment
an article of clothing; "garments of the finest silk"
Activation Score: 4.15\0.30
food
any substance that can be metabolized by an organism to give energy and build tissue
change of state the act of changing something into something different in essential characteristics
Activation Score: 4.15\0.30
action
something done (usually as opposed to something said); "there were stories of
murders and other unnatural actions"
time period
an indefinite length of time; "a time period of 30 years"; "hastened the period of his
recovery"
Activation Score: 4.17\0.30
social control
control exerted (actively or passively) by group action
social group
people sharing some social relation
Activation Score: 4.17\0.30
content
situation
the sum or range of what has been perceived
the general state of things; the combination of circumstances at a given time; "the
present international situation is dangerous"; "wondered how such a state of affairs had
come about"
Activation Score: 4.17\0.30
action
something done (usually as opposed to something said); "there were stories of
murders and other unnatural actions"
way
any road or path affording passage from one place to another; "he said he was
looking for the way out"
Activation Score: 4.18\0.30
person
quality
a human being; "there was too much for one person to do"
an essential and distinguishing attribute of something or someone; "the quality of
mercy is not strained"--Shakespeare
Activation Score: 4.18\0.30
organization
a group of people who work together
worker
a person who works at a specific job; "he is a good worker"
Activation Score: 4.24\0.31
life form
any living entity
consumer goods goods (as food or clothing) intended for direct use or consumption
Activation Score: 4.27\0.32
person
a human being; "there was too much for one person to do"
role
normal or customary activity; "what is your role on the team?"
Activation Score: 4.27\0.32
decoration
structure
something used to beautify
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches";
"she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 4.27\0.32
194
passage
structure
of arches";
a path or channel through or along which someone or something may pass
a thing constructed; a complex construction or entity; "the structure consisted of a series
"she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 4.27\0.32
location
a point or extent in space
magnitude
the property of relative size or extent; "they tried to predict the magnitude of the
explosion"
Activation Score: 4.34\0.33
person
a human being; "there was too much for one person to do"
deed
a notable achievement: "the book was her finest effort"
Activation Score: 4.34\0.33
communication something that is communicated between people or groups
belief
any cognitive content held as true
Activation Score: 4.38\0.34
change of integrity
instrumentality
the act of changing the unity or wholeness of something
an artifact (or system of artifacts) that is instrumental in accomplishing
some end
Activation Score: 4.38\0.34
building
a structure that has a roof and walls and stands more or less permanently in one
place; "there was a three-story building on the corner"; "it was an imposing edifice"
organization
a group of people who work together
Activation Score: 4.38\0.34
facility
facility"
structure
something created to provide a particular service; "the assembly plant is an enormous
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches";
"she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 4.38\0.34
artifact
a man-made object
beginning
the act of starting something; "he was responsible for the beginning of negotiations"
Activation Score: 4.38\0.34
plant
a living organism lacking the power of locomotion
plant product
a product made from plant material
Activation Score: 4.39\0.34
person
position
a human being; "there was too much for one person to do"
(in team sports) the role assigned to an individual player; "what position does he
play?"
Activation Score: 4.43\0.35
organization
region
a group of people who work together
a large indefinite location on the surface of the Earth; "penguins inhabit the polar
regions"
Activation Score: 4.43\0.35
group
any number of entities (members) considered as a unit
195
property
a basic or essential attribute shared by all members of a class; "a study of the physical
properties of atomic particles"
Activation Score: 4.47\0.35
state
the way something is with respect to its main attributes; "the current state of
knowledge"; "his state of
health"; "in a weak financial state"
emotion
any strong feeling
Activation Score: 4.56\0.37
act
something that people do or cause to happen
social event
an event characteristic of persons forming groups
Activation Score: 4.56\0.37
structure
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches";
"she wore her hair in an amazing construction of whirls and ribbons"
institution
an organization founded and united for a specific purpose
Activation Score: 4.56\0.37
measure
structure
how much there is of something that you can measure
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches"; "she wore her hair in an amazing construction of whirls and ribbons"
Activation Score: 4.68\0.39
person
music
a human being; "there was too much for one person to do"
an artistic form of auditory communication incorporating instrumental or vocal tones in a
structured and continuous manner
Activation Score: 4.68\0.39
sign of the zodiac
one of 12 equal areas into which the zodiac is divided
constellation
a configuration of stars as seen from the earth
Activation Score: 4.68\0.39
life form
color
"white is the
any living entity
a visual attribute of things that results from the light they emit or transmit or reflect;
coolest summer color"
Activation Score: 4.77\0.41
feeling
the psychological feature of experiencing affective and emotional states; "he had a
feeling of euphoria"
property
a basic or essential attribute shared by all members of a class; "a study of the physical
properties of atomic particles"
Activation Score: 4.82\0.42
act
something that people do or cause to happen
information
a message received and understood that reduces the recipient''s uncertainty
Activation Score: 4.83\0.42
grain
cereal
cereal grain suitable as food for human beings
grass whose starchy grains are used as food: wheat; rice; rye; oats; maize;
buckwheat; millet
Activation Score: 4.83\0.42
activity
any specific activity or pursuit; "they avoided all recreational activity"
196
knowledge domain
the content of a particular domain or field of knowledge
Activation Score: 4.88\0.43
musical composition
a musical work that has been created; "the composition is written in four
movements"
musical organization
an organization of musicians who perform together
Activation Score: 4.90\0.43
plant
root
a living organism lacking the power of locomotion
the usually underground organ that lacks buds or leaves or nodes; absorbs water and
mineral salts; usually it anchors the plant to the ground
Activation Score: 4.96\0.44
writing
reading matter; anything expressed in letters of the alphabet (especially when
considered from the point of view of style and effect); "the writing in her novels is
excellent"
speech act
the use of language to perform some act
Activation Score: 4.96\0.44
action
something done (usually as opposed to something said); "there were stories of
murders and other unnatural actions"
movement
a natural event that involves a change in the position or location of something
Activation Score: 4.97\0.45
animal
a living organism characterized by voluntary movement
animal skin
the tissue forming the outer covering of a fur-bearing animal
Activation Score: 5.01\0.45
dancing
music
taking a series of rhythmical steps (and mo vements) in time to music
an artistic form of auditory communication incorporating instrumental or vocal tones
in a structured and continuous manner
Activation Score: 5.07\0.46
act
ability
something that people do or cause to happen
possession o f the qualities (especially mental qualities) required to do something or
get something done; "danger heightened his powers of discrimination"
Activation Score: 5.09\0.47
measure
how much there is of something that you can measure
container
something that holds things
Activation Score: 5.12\0.47
creation
the human act of creating
creation
something that has been brought into existence by someone
Activation Score: 5.12\0.47
person
a human being; "there was too much for one person to do"
occupation
the principal activity in your life; "he''s not in my line of business"
Activation Score: 5.12\0.47
Container
something that holds things
containerful
the quantity that a container will hold
Activation Score: 5.14\0.48
group
belief
any number of entities (members) considered as a unit
any cognitive content held as true
197
Activation Score: 5.16\0.48
substance
that which has mass and occupies space; "an atom is the smallest indivisible unit of
matter"
drug
something that is used as a medicine or narcotic
Activation Score: 5.16\0.48
group action
action taken by a group of people
social group
people sharing some social relation
Activation Score: 5.17\0.48
language
administrative district
Activation Score: 5.17\0.48
a systematic means of communicating by the use of sounds or conventional
symbols; "he taught foreign languages"; "the language introduced is standard
throughout the text"
a district defined for administrative purposes
social group
region
people sharing some social relation
a large indefinite location on the surface of the Earth; "penguins inhabit the polar
regions"
Activation Score: 5.19\0.48
attribute
an abstraction belonging to or characteristic of an entity
perception
the process of perceiving
Activation Score: 5.20\0.49
person
a human being; "there was too much for one person to do"
position
a job in an organization or hierarchy; "he ocupied a post in the treasury"
Activation Score: 5.21\0.49
property
properties of
a basic or essential attribute shared by all members of a class; "a study of the physical
atomic particles"
perception
the process of perceiving
Activation Score: 5.21\0.49
life form
any living entity
drug
something that is used as a medicine or narcotic
Activation Score: 5.22\0.49
person
job
a human being; "there was too much for one person to do"
the occupation for which you are paid; "he is looking for a job"; "a lot of people are
out of work"
Activation Score: 5.24\0.49
communication
product
something that is communicated between people or groups
the amount of an artifact that has been produced by someone or some
process; "they improve their product every year"; "they export most of their
agricultural production"
Activation Score: 5.28\0.50
natural object
an object occurring naturally; not made by man
location
a point or extent in space
Activation Score: 5.28\0.50
person
a human being; "there was too much for one person to do"
198
organization
a group of people who work together
Activation Score: 5.31\0.51
belief
any cognitive content held as true
social group
people sharing some social relation
Activation Score: 5.31\0.51
person
a human being; "there was too much for one person to do"
conveyance
something that serves as a means of transportation
Activation Score: 5.37\0.52
person
a human being; "there was too much for one person to do"
social group
people sharing some social relation
Activation Score: 5.37\0.52
belief
any cognitive content held as true
institution
an organization founded and united for a specific purpose
Activation Score: 5.44\0.53
gathering
a group of persons together in one place
organization
a group of people who work together
Activation Score: 5.45\0.53
artifact
a man-made object
creation
the human act of creating
Activation Score: 5.48\0.54
structure
a thing constructed; a complex construction or entity; "the structure consisted of a
series of arches"; "she wore her hair in an amazing construction of whirls and ribbons"
organization
a group of people who work together
Activation Score: 5.55\0.55
feeling
the psychological feature of experiencing affective and emotional states; "he had a
feeling of euphoria"
state
the way something is with respect to its main attributes; "the current state of
knowledge"; "his state of health"; "in a weak financial state"
Activation Score: 5.56\0.55
creation
writing
something that has been brought into existence by someone
reading matter; anything expressed in letters of the alphabet (especially when
considered from the point of view of style and effect); "the writing in her novels is
excellent"
Activation Score: 5.56\0.55
substance
that which has mass and occupies space; "an atom is the smallest indivisible unit of
matter"
color
a visual attribute of things that results from the light they emit or transmit or reflect;
"white is the coolest summer color"
Activation Score: 5.57\0.55
religion
a strong belief in a supernatural power or powers that control human destiny; "he lost his
faith but not his morality"
organization
a group of people who work together
Activation Score: 5.63\0.56
object
a physical (tangible and visible) entity; "it was full of rackets
199
color
a visual attribute of things that results from the light they emit or transmit or reflect;
"white is the coolest summer color"
Activation Score: 5.63\0.56
activity
any specific activity or pursuit; "they avoided all recreational activity"
trait
a distinguishing feature of one''s personal nature
Activation Score: 5.65\0.57
instrumentality an artifact (or system of artifacts) that is instrumental in accomplishing some end
worker
a person who works at a specific job; "he is a good worker"
Activation Score: 5.66\0.57
object
a physical (tangible and visible) entity; "it was full of rackets
gathering
a group of persons together in one place
Activation Score: 5.85\0.61
message
what a communication that is about something is about
speech act
the use of language to perform some act
Activation Score: 5.89\0.61
plant
a living organism lacking the power of locomotion
seed
a small hard fruit
Activation Score: 5.92\0.62
writing
from the
reading matter; anything expressed in letters of the alphabet (especially when considered
point of view of style and effect); "the writing in her novels is excellent"
message
what a communication that is about something is about
Activation Score: 5.98\0.63
ingredient
herb
food that is a component of a mixture in cooking
a plant lacking a permanent woody stem; many are flowering garden plants or potherbs;
some having medicinal properties; some are pests
Activation Score: 5.98\0.63
plant
a living organism lacking the power of locomotion
artifact
a man-made object
Activation Score: 5.99\0.63
written communication
message
Activation Score: 6.06\0.64
flavorer
herb
some having
communication by means of written symbols
what a communication that is about something is about
something added to food primarily for the savor it imparts
a plant lacking a permanent woody stem; many are flowering garden plants or potherbs;
medicinal properties; some are pests
Activation Score: 6.06\0.64
artifact
shape
mathematical
a man-made object
the spatial arrangement of something as distinct from its substance; "geometry is the
science of shape"
Activation Score: 6.08\0.65
food
any substance that can be metabolized by an organism to give energy and build tissue
200
artifact
a man-made object
Activation Score: 6.12\0.65
communication
something that is communicated between people or groups
speech act
the use of language to perform some act
Activation Score: 6.13\0.66
object
shape
a physical (tangible and visible) entity; "it was full of rackets
the spatial arrangement of something as distinct from its substance; "geometry is the
mathematical science of shape"
Activation Score: 6.22\0.67
plant
a living organism lacking the power of locomotion
plant material
material derived from plants
Activation Score: 6.41\0.71
person
device
a human being; "there was too much for one person to do"
an instrumentality invented for a particular purpose; "the device is small enough to
wear on your wrist"; "a device intended to conserve water"
Activation Score: 6.51\0.72
person
a human being; "there was too much for one person to do"
instrumentality an artifact (or system of artifacts) that is instrumental in accomplishing some end
Activation Score: 7.01\0.82
animal
a living organism characterized by voluntary movement
food
any substance that can be metabolized by an organism to give energy and build tissue
Activation Score: 7.09\0.83
edible fruit
edible reproductive body of a seed plant especially one having sweet flesh
fruit tree
tree bearing edible fruit
Activation Score: 7.18\0.85
plant
a living organism lacking the power of locomotion
edible fruit
edible reproductive body of a seed plant especially one having sweet flesh
Activation Score: 7.37\0.88
person
language
a human being; "there was too much for one person to do"
a systematic means of communicating by the use of sounds or conventional symbols;
"he taught foreign languages"; "the language introduced is standard throughout the text"
Activation Score: 7.62\0.93
person
a human being; "there was too much for one person to do"
natural language a human written or spoken language used by a community; opposed to e.g. a
computer language
Activation Score: 7.72\0.95
life form
any living entity
produce
fresh fruits and vegetable grown for the market
Activation Score: 7.82\0.96
plant
a living organism lacking the power of locomotion
food
any substance that can be metabolized by an organism to give energy and build tissue
Activation Score: 7.91\0.98
201
Bibliography
Anick, P. and Bergler, S., Lexical Structures for Linguistic Inference,
In: Pustejovsky, J. and Bergler, S. (eds), Lexical Semantics and Knowledge Representation Berlin:
Springer Verlag.
Anonymous (1954), Ad C. Herennium, De Ratione Dicendi (Rhetorica ad Herennium), London:
Heinemann.
Apresjan, J. (1973), Regular Polysemy
In: Linguistics 142, pp. 5-32
Baeza-Yates, Ricardo and Berthier Ribeiro-Neto (1999), Modern Information Retrieval.
ACM Press and Addison Wesley.
Blank, A. (1999), Co-presence and Succession: A Cognitive Typology of Metonymy,
In: Panther, K. and Radden, G. (Eds.), Metonymy in Language and Thought, Amsterdam, John Benjamins,
pp. 169-192
Bowerman, M. (1977), The Acquisition of Word Meaning: An Investigation of Some Current Concepts
In: Johnson-Laird, P. and Wason, P., Thinking. Readings in Cognitive Science
Cambridge: Cambridge University Press, pp. 239-244
Bredenkamp, A., Markantonatou, S. and L. Sadler (1996), Lexical rules: What are they?
In: Proceedings of COLING 96, Copenhagen, pp. 163-168.
Bredin, H., Metonymy
Poetics Today 5: 45-58
Briscoe, T. and Copestake, A. (1999), Lexical Rules in Constraint-bases Grammar
Computational Linguistics 15 (4), pp. 487-526
Buitelaar, P (1998)
Corelex: Systematic Polysemy and Underspecification,
Ph.D., Department of Computer Science, Brandeis University, Boston, U.S.A.
Chugur, I., Gonzalo, J. and Verdejo, F. (2000), Sense Clusters for Information Retrieval: Evidence from
Semcor and the InterLingual Index
Proceedings of the ACL 2000 Workshop on Word Senses and Multilinguality, Hong Kong
CIDE (1995)
Cambridge International Dictionary of English.
CUP, Cambridge, England.
Clark, E. (1993), The Lexicon in Acquisition, Cambridge, Cambridge University Press
Copestake, A. and Briscoe, T. (1996), Semi-productive Polysemy and Sense Extension
In: Pustejovsky, J. and Boguraev, B. (Eds.), Lexical Semantics, the Problem of Polysemy,
Clarendon Press, Oxford
Croft, W. (1993), The Role of Domains in the Interpretation of Metaphors and Metonymies
Cognitive Linguistics 4: 335-370
Cruse, D.A. (1986). Lexical Semantics,
Cambridge University Press, Cambridge, U.K.
202
Cruse, D. A. (2000.), Meaning in language: An introduction to semantics and pragmatics.
New York: Oxford University Press
Daudé J., Padró L. and Rigau G. (2001), A Complete WN1.5 to WN1.6 Mapping,
Proceedings of NAACL Workshop "WordNet and Other Lexical Resources: Applications, Extensions and
Customizations". Pittsburg, PA, United States
Deerwester, S., Dumais, S., Furnes, G., Landauer, T. and Harshman, R. (1990)
Indexiong by Latent Semantic Analysis.
Journal of the American Society for Information Science 41 (16): 391\407.
DHLF (1992),
Dictionnaire Historique de la Langue Française, Alain Rey, Paris: Dictionnaires le Robert
Dik, S. C. (1977), Inductive Generalization in Semantic Change. In: P.J. Hopper (Ed.), Studies in
Descriptive and Historical L inguistics. Festschrift for W.P. Lehmann. Amsterdam and Philadelphia:
Benjamins, pp. 283-300
EAGLES (1999)
Preliminary Recommendations on Semantic Encoding. Technical report, EAGLES Lexicons Interest
Group.
Evens, M.W. (1988), Relational Models of the Lexicon: Representing Knowledge in Semantic Networks,
Cambridge: Cambridge University Press
Fass, Dan C. (1988). An Account of Coherence, Semantic Relations, Metonymy, and Lexical Ambiguity
Resolution.
In Small, S., Cottrell, G. and Tanenhaus, M. (Eds.) Lexical Ambiguity Resolution: Perspectives from
Psycholinguistics, Neuropsychology, and Artificial Intelligence. Los Altos, CA: Morgan Kaufmann, pp.
151-178.
Fass, D. (1991), Metonymy, Case Substitution and Sense Ambiguity,
In Fass, D. et al. (eds.), Proceedings of the IJCAI Workshop on Computational Approaches to Non-Literal
Language: Metaphor, Metonymy, Idiom, Speech Acts and Implicature, 12th International Conferencee on
Artificial intelligence, Sydney
Fass, D. (1997), Processing Metaphor and Metonymy,
Ablex, Stanford, CA
Fellbaum, Christiane (ed.) (1998)
WordNet: An Electronic Lexical Database.
Cambridge, Mass.: MIT Press.
Fillmore, C (1977), Scenes and frames semantics.
In: Zampolli, A (ed.) Linguistic structures processing.
Benjamins, Amsterdam, The Netherlands, pp. 55-81.
Fillmore, C. & Atkins, B., Describing Polysemy: The Case of ‘Crawl’,
In: Ravin, Y. & Leacock, C. (Eds.) (2000), Polysemy, Theoretical and Computational Approaches, Oxford,
Oxford University Press
Fillmore, C. and Atkins, S. (1992), Towards a Frame-based Organization of the Lexicon: the Semantics of
RISK and its Neighbors.
In: Lehrer, A. and Kittay, E. (Eds.) Frames, Fields and Contrasts: New Essays in Semantics and Lexical
Organization, Hillsdale: Lawrence Erlbaum
203
Firth, Modes of Meaning
In: Firth, J.R., Papers in Linguistics 1934-1951
London: Oxford University Press
Gale, W., Church, K. and Yarowsky, D. (1992), One Sense Per Discourse,
In: Proceedings of the ARPA Workshop on Speech and Natural language Processing
Morgan Kaufmann, San Francisco, U.S.A.
Geeraerts, D. (1997), Diachronic Prototype Semantics: A Contribution to HistoricalLlexicology. Oxford:
Clarendon.
Givon, T. (1967), Transformations of Ellipsis, Sense Development and Rules of Lexical Derivation,
SP2896, Santa Monica, California, Systems Development Corp.
Gove, P.B. (Ed.) (1971), Webster’s Third New International Dictionary of the English Language
Chicago: Merriam Co.
Grefenstette, G. (1994a)
Corpus-derived first-, second- and third-order word affinities.
In Proc. Euralex, Amsterdam, The Netherlands
Grefenstette, G. (1994b)
Explorations in Automatic thesaurus discovery.
Dordrecht: Kluwer.
Grice, H.P. (1975), Logic and Conversation,
In: Cole, P. and Morgan, J., Syntax and Semantics 3: Speech Acts
New York, Academic Press
Hanks, P. (1987), Definitions and Explanations.
In Sinclair, J. (ed.) Looking Up: An account of theCOBUILD Project in lexical computing,
Collins, London and Glasgow.
Hanks, P. (2000), Do Word Meanings Exist?
in Computing and the Humanities, Volume 34
Harabagiu, S. (1998), Deriving Metonymic Coercions from WordNet
Workshop on the Usage of WordNet in Natural language Processing Systems, COLING-ACL, Montreal,
Canada, pp.142-148
Harabagiu, S. and Maiorano, S. (2000), Acquisition of Linguistic Patterns for Knowledge-based
Information Extraction,
In: Proceedings of LREC 2000, Athens, Greece
Ide, N. (1999), Cross-lingual Sense Determination: Can It Work?
Computers and the Humanities, Kluwer Academic Publishers,
Jorgensen, J., (1990). The Psychological Reality of Word Senses,
In: Journal of Psycholinguistic Research 19 (3), pp. 167-190
Kamei, S. and Wakao, T. (1992), Metonymy: Reassessment, Survey of Acceptability, and its Treatment in a
Machine Translation System
In: Proceedings of ACL, Newark, Del., pp. 309-311
204
Kilgarriff, A. (1997),
I Don’t Believe in Word Senses,
Special Issue on Word Sense Disambiguation. Computers and the Humanities
Kilgarriff, A. and Palmer, M. (2000) Guest editors,
Special Issue on SENSEVAL: Evaluating Word Sense Disambiguation Programs. Computers and the
Humanities.
Koch, P. (1999), Frame and Contiguity,
In: Panther, K. and Radden, G. (Eds.), Metonymy in Language and Thought, Amsterdam, John Benjamins,
pp. 140-167
Krovetz, B. (1998), More than One Sense per Discourse,
In: Proceedings of the ACL-SIGLEX Workshop, Trento, Italy
Lakoff, G. (1987), Women, Fire, and Dangerous Things: What Categories Reveal about the Mind.
Chicago: The University of Chicago Press.
Lakoff, George. 1993. The Contemporary Theory of Metaphor
In: Andrew Ortony (ed.), Metaphor and Thought, second ed. New York:
Cambridge University Press, pp. 202-251
Lakoff, G. and Johnson, M. (1980), Metaphors we live by
Chicago, The University of Chicago Press
Lakoff, George (1993), The Contemporary Theory of Metaphor
In: Andrew Ortony (ed.), Metaphor and Thought,
Cambridge University Press, New York , U.S.A., pp. 202-251
Landau, S. (1989), Dictionaries: The Art and Craft of Lexicography
Cambridge, Cambridge University Press
Landes, S. Claudia Leacock, C. and Tengi, R. (1998), Building Semantic Concordances
In: C. Fellbaum (ed.), WordNet: An Electronic Lexical Database.
MIT Press, Cambridge, Mass., U.S.A.
Langacker, R. (1984), Concept, Image, and Symbol: The Cognitive Basis of Grammar.
Berlin: Mouton de Gruyter
Langacker, R. (1987), Foundations of Cognitive Grammar. Volume I: Theoretical Prerequisites, Stanford,
California: Stanford University Press.
Levin, B. (1993)
English Verb Classes and Alternations.
University of Chicago Press
Levinson, S. (1983), Pragmatics
Cambridge University Press, Cambridge U.K.
Lin, Dekang (1998)
Automatic retrieval and clustering of similar words.
In COLING-ACL. Montreal.
Macquarie, (1986). Macquarie Thesaurus.
Edited by J. R. L. Bernard (First published 1984). Sydney.
205
Macquarie, (1997). Macquarie Dictionary, 3rd Edition.
Editor in Chief Arthur Delbridge. Sydney.
Marcus, M., Santorini, B., Markinkiewicz , M. (1994), Building a Large Annotated Corpus of English: The
Penn Treebank
Computational Linguistics, 19(2): 313-330
Markert, K. and Hahn, U. (2002), Understanding Metonymies in Discourse.
Artificial Intelligence, 135(1/2), pp.145–198
Markert, K. and Nissim, M. (2002), Towards a Corpus Annotated for Metonymies: the Case of location
Names
In: Proceedings of LREC2002, Las Palmas, Spain, pp. 1385-1392
Markert, K. and Nissim, M. (2002), Metonymy Resolution as a Classification Task
In: Yarowsky, D., (Ed.), Proceedings of EMNLP2002
Masterman, M., (1957). The Thesaurus in Syntax and Semantics. Mechanical Translation, 4:1–2.
McArthur, T. (1981), Longman Lexicon of Contemporary English
Longman. London and Harlow.
Miller G.A, Beckwidth, R., Fellbaum, C., Gross, D. and Miller, K.J. (1990),
Introduction to WordNet: An On-line Lexical Database,
In: International Journal of Lexicography, Vol. 3, No.4, 235-244.
Minsky, M. (1975), A Framework for Representing Knowledge.
In: Winston, P.H. (Ed.), The Psychology pof Computer Vision, New York: McGraw-Hill, pp. 211-277.
Nehrlich, B. and Clarke, D., A Dynamic Model of Semantic Change,
Journal of Literary Semantics 17, pp. 73-90
Nehrlich, B., Clarke, D. and Todd, Z. (1999), Mummy, I like being a Sandwich: Metonymy in Language
Acquistition
In: Panther, K.U. and Radden, G. (eds.) Metonymy in language and Thought.
John Benjamins, Amsterdam
Nunberg, G. (1978), The Pragmatics of Reference.
Bloomington: The Indiana University Linguistics Club.
Nunberg, G. (1996), Transfers of Meaning,
In: Pustejovsky, J. and Boguraev, B. (Eds.), Lexical Semantics, the Problem of Polysemy,
Clarendon Press, Oxford
Ostler, N. and Atkins, B. (1991), Predictable Meaning Shift: Some Linguistic properties of Lexical
Implication Rules,
In: Pustejovsky, J., Bergler, S. (Eds.), Lexical Semantics and Knowledge Representation, ACL SIGLEX
Workshop Berkeley, California
Peters, W., Peters, I. and Vossen, P. (1998).
Automatic sense clustering in EuroWordNet.
In Proc. First Intnl Conf on Language Resources and Evaluation.
Granada, Spain
Peters, W. and Peters, I. (2000)
Lexicalised Systematic Polysemy in WordNet
206
In Proc. Second Intnational Conference on Language Resources and Evaluation
Athens, Greece
Peters, W. and Kilgarriff, A., Discovering Semantic Regularity in Lexical Resources,
International Journal of Lexicography, 13, 4, pp. 287-312
Peters, W., Vossen, P., Diez-Orzas, P., Adriaens, G. (1998), Cross-linguistic Alignment of
Wordnets with an Inter-Lingual-Index,
In: Ide, N., Greenstein, D.,Vossen, P. (eds), Special Issue on EuroWordNet. Computers and the
Humanities, Volume 32, Nos. 2-3 1998. 221-251.
Peters, W. (2002), Extraction of Implicit Lexical Knowledge from WordNet
In: proceedings of Ontolex workshop, Las Palmas, Spain
Peters, W. and Wilks, Y. (2002), Data-driven Detection of Figurative Language Use in Electronic
language Resources
Metaphor and Symbol Special Issue
Proctor, P. (Ed.) (1978), The Longman Dictionary of Contemporary English ,
Longman, London
Pustejovsky, J. (1995). The Generative Lexicon,
MIT Press, Cambridge MA, U.S.A.
Meister, F. (Ed.) (1886), Quintilianus, M.F., Institutionis Oratoriae, Leipzig: Freitag
Radden and Kovecses (1999), Towards a Theory of Metonymy,
In: Panther, K. and Radden, G. (Eds.), Metonymy in Language and Thought, Amsterdam, John Benjamins,
pp. 17-60
Raible, W. (1981), Von der Allgegenwart des Gegensinns. Strategien zur Einordnung Semantischer
Informationen, Zeitschrift fuer Romanische Philologie: pp. 1-40.
Ravin, Y & Leacock, C. (Eds.) (2000), Polysemy: An Overview,
In: Ravin, Y. & Leacock, C. (Eds.) (2000), Polysemy, Theoretical and Computational Approaches, Oxford,
Oxford University Press
Resnik, P. (1997), Selectional Preferences and Sense Disambiguation,
In: Proceedings of the ACL SIGLEX Workshop on Tagging Text with Lexical Semantics: Why, What and
How?, Washington, D.C., USA
Resnik, P. (1999), Semantic Similarity in a Taxonomy: An Information-Based Measure and its Application
to Problems of Ambiguity in Natural Language
Journal of Artificial Intelligence Research 11, pp. 95-130
Resnik, P. and Yarowsky, D. (1997). A perspective on word sense disambiguation methods and their
evaluation
ACL-SIGLEX Workshop Tagging Text with Lexical Semantics: Why, What, and How? April 4-5, 1997,
Washington, D.C., 79-86.
Roget, P.M., Thesaurus of English Words and Phrases,
Oxford, 1852
Rosch, E. (1973), On the Internal Structure of Perceptual and Semantic Categories,
In: Moore, Th.E. (Ed.), Cognitive Development and the Acquisition of Language.
207
New York, Academic Press
Ross, W.D. (ed.) (1924), The Works of Aristotle,
Oxford Clarendon Press
Schank, R. and Abelson, R. (1975), Scripts, Plans and Knowledge
In: Advanced Papers 4th Internatioinal Joint Conference on Artificial Intelligence
Schütze, H. (1998)
Automatic word sense discrimination.
Computational Linguistics, 24(1):97–124.
Seto, Ken-ichi (1996), On the Cognitive Triangle: the Relation of Metaphor, Metonymy and Synecdoche
In: A. Burkhardt & N. Norrick, eds., Tropic Truth, Berlin-NewYork: De Gruyter, 1996.
Seto, K. (1999), Distinguishing Metonymy from Synecdoche,
In: Panther, K. and Radden, G. (Eds.), Metonymy in Language and Thought, Amsterdam, John Benjamins,
pp. 91-120
Sparck Jones, K., Synonymy and Semantic Classification (thesis, 1964), Edinburgh: Edinburgh University
Press, 1986.
Sperber, D. and Wilson, D. (1986), Relevance: Communication and Cognition
Harvard University Press, Cambridge, MA
Stallard, D. (1993), Two Kinds of Metonymy,
Proceedings of ACL, Columbus, Ohio, pp.87-94
Stock, P. (1983), Polysemy.
In Proc. Exeter Lexicography Conference.
Summers, D. (1988), The role of dictionaries in language learning
In R. A. Carter and M. McCarthy (eds.), Vocabulary and Language Teaching.
London: Longman, pages 111–125.
Summers, D. (1995), The Longman Language Activator.
Harlow: Longman.
Taylor, J. R. (1995), Linguistic categorization: Prototypes in Linguistic Theory, Oxford, Oxford University
Press
Traugott, E. and Koenig, E., The Semantics/Pragmatics of Grammaticalization Revisited,
In: Traugott, E. and Heine, B. (Eds.), Approaches to Grammaticalization, Amsterdam and Philadelphia,
Benjamins, pp. 189-218.
Turner, M. (1993), Reading Minds,
Princeton University Press, Princeton, U.S.A.
Ullmann, S. (1962), Semantics: An Introduction to the Science of Meaning
Oxford: Blackwell
Véronis, J. (2001). Sense tagging: does it make sense?
In: Proceedings of Corpus Linguistics 2001, Lancaster, U.K.
Verspoor, C. (1997), Conventionality-governed logical metonymy
In Bunt, H., Kievit, L, Muskens, R. (Eds.), Proceedings of the 2nd International Workshop on
Computational Semantics, Tilburg, The Netherlands, pp. 300-312
208
Vickery, B.C. (1975), Classification and Indexing in Science,
Butterworth and Co., London
Vossen, P. (1998), EuroWordNet: Building a Multilingual Database with Wordnets for European
Languages
In: Choukri, K., Fry, D., Nilsson, M. (eds), The ELRA Newsletter, Vol 3
Vossen, P., Introduction to EuroWordNet.
In: Nancy Ide, N., Greenstein, D. and Vossen, P. (eds), Special Issue on EuroWordNet. Computers and the
Humanities, Volume 32, Nos. 2-3 1998. 73-89.
Vossen, P., Peters, W. and Díez-Orzas, P. (1997), The Multilingual design of the EuroWordNet
Database
In: Mahesh, K. (ed.) Ontologies and multilingual NLP, Proceedings of IJCAI-97 workshop,
Nagoya, Japan, August 23-29.
Warren, B. (1999), Aspects of Referential Metonymy
In: Ravin, Y. & Leacock, C. (Eds.) (2000), Polysemy, Theoretical and Computational Approaches, Oxford,
Oxford University Press
Wilks, Y. (1975), A Preferential, Pattern-Seeking Semantics for Natural Language Inference
Artificial Intelligence, 6:53-74.
Wilks, Y., Slator, B.M. and Guthrie, L. (1996)
Electric words: dictionaries, computers and meanings.
Cambridge
Wittgenstein, L., (1953), Philosophical Investigations
Oxford: Blackwell
209