Ling 181 Lexical Semantics
Classification of the Semantic Relations in Noun Compounds
Barbara Rosario
([email protected])
Introduction
This paper is about the classification of noun-compounds (NCs) into classes that reflect the
semantic relationships between the nouns. I am ultimately interested in the automatic classification
of NCs with machine learning techniques. In order to learn the classification (to eventually classify
previously unseen NCs) these techniques need as input noun compounds that have been previously
classified. I created a collection of NCs extracting them from abstracts of medical journals and I
classified them. I developed the classification schema without any knowledge of the linguistic
literature on the subject. This paper is an attempt to overcome this ignorance. I will review a few
linguistic accounts of the semantics of NCs, and compare in detail the differences of my
classification with the ones proposed by [Levi 1978] and by [Warren 1978]. I am also interested in
making sure that my classification schema can be justified from a linguistic point of view, in
exploring ways to improve it, in seeing how (and if) NCs in the medical domain differ (from the
point of view of the semantic relationships) from NCs of common usage. My analysis will take into
account the fact that my goal is the automatic classification of NCs.
This paper is organized as follows: in the next Section I introduce NCs and in Section 2 I
explain my motivations for studying NCs. In Section 3 I briefly review (some of) the linguistic
literature on semantics of NCs. Section 4 describes [Levi 1978], in Section 5 I compare my
classification schema with Levi’s, in Section 6 Warren’s theory [Warren1978] is briefly introduced
and my classification schema compared with hers in Section 7.
1) Introducing Noun Compounds
In this Section I will (briefly) introduce some facts about NCs .
There are a multitude of possible definitions for NCs. [Lauer 1995] says the most popular are:
1. Noun premodifier: Any constituent can appear before a noun to form a NC: out-in-thewilds cottages is therefore considered a NC
2. Compounds (phonological definition due to Chomsky): words preceding a noun form a
compound if they receive primary stress, Thus blackboard is a compound, while black
board is not.
3. Complex Nominals: Levi (1978) chooses to include certain adjective along with nouns
as possible compounding elements that she calls Complex Nominals. The adjectives
included are non-predicative adjectives like in electrical engineer.
4. Noun Noun Compounds: any sequence of nouns that itself functions as a noun.
5. Noun Noun Compounds: any sequence of nouns at least two words in length that itself
functions as a noun, but which contains no genitive markers and is not a name.
Compounds with hyphens are called dvandva compounds. Levi doesn’t talk about them, and
Warrens says she found only ten dvandva compounds in her corpus (for example, founder-director,
poet-painter, banker-editor). She says that these compounds are different from the others in that the
modifier does not predicate something about the second constituent and it doesn’t modifies it’s
meaning. The motivation for combining poet and painter is different from the motivation of
combining girl and friend in girl friend. In the former case, we want to convey that someone is not
only a poet but also a painter, and we are not trying to identify which painter, or to define what kind
of painter we are talking about. In the latter case, we want make our reference more precise. Poetpainter means “more” that just painter, we have an expansion of reference scope, while for girl
friend we narrow the reference scope and girl friend “means less” than friend.
For my application, I use definition 5 with the difference that I consider also names (like
Mexico City). Warren (1978) includes also orthographically joined morphemes such as gunman.
She doesn’t differentiate the semantics of continuous compounds vs. the semantics of compounds
of separate words. For the moment, I am considering only sequences of separate nouns, assuming
that the words in orthographically joined morphemes are joined because they are sufficient
common to warrant inclusion in the lexicon, and thus do not require dynamic processing. In the
future, I’ll need to address this issues and I also plan of analyzing the semantics of adjective-noun
pairs.
There are (many) other important issues about NCs such as lexicalisation, syntactic analysis
(parsing), prosody, linguistic function, grammar (nominalizations vs. non-verbal-nexus
compounds), ambiguity, but they are beyond the scope of this paper that focuses on the semantic
analysis of NCs.
2) Motivation of (my) Noun Compound research
Understanding NCs is a vital component of any sophisticated natural language system. It has
been shown that their frequency has been increasing over the last two centuries and they are
especially common in technical text.
I am particularly interested in the automatic analysis of NCs in the bio-medical domain. This is
a very important application of NLP. Some project in this domain that are currently going on are,
for example, the automatic extraction of medical concepts from medical record texts and combining
genome sequence information with the published biological record, “reading” the scientific
2
literature and successfully predict gene interactions1. Very ambitiously, in general, analysis of text
can lead to hypotheses about diseases (see [Swanson97]) and of course if we want to formulate
hypothesis, we need to determine the relationships that hold between different concepts; for
example we need to know that in the phrase “stress can lead to loss of magnesium” stress and
magnesium are connected by a cause-effect kind of relationship.
One of the important challenges of biomedical text, along with most other technical text, is the
proliferation of noun compounds. This is a real title of a medical article:
Open-labeled long-term study of the
subcutaneous sumatriptan efficacy and
tolerability in acute migraine treatment
The real concern in analyzing such a title is in determining the relationships that hold between
different concepts: we need to know that for example, that migraine treatment is a treatment for the
disease migraine (purpose kind of relationship). My research is about finding automatic algorithms
that can learn /find these relationships and obviously in order to do this , we first need to know
what are these relationships.
To create a collection of noun compounds, I performed searches from MedLine, which contains
references and abstracts from 4300 biomedical journals. I used several query terms, intended to
span across different sub fields. I retained only the titles and the abstracts of the retrieved
documents. On these titles and abstracts I ran a part-of-speech tagger and a program that extracts
only sequences of units tagged as nouns. I extracted NCs with up to 6 constituents, but for this
paper I consider only NCs with 2 constituents.
In order to develop a classification schema for the NCs, I used a corpus-based methodology,
i.e. after acquiring the NCs from the corpus I used every NC2 for the development of the
classification schema (this approach allows also to measure the relative frequency of different
semantic relations, which ensures that emphasis is given to semantic behaviors that are more likely
to be encountered3). The relations were found by iterative refinement based on looking at 2245
extracted compounds and finding commonalities among them. The work of this paper is a study of
some of theories in the linguistic literature on the subject and a comparison of my classification
with 2 of such theories.
3) Range of semantic theories4
The semantic properties of NCs have been hotly debated in linguistics, with numerous
contradictory views being proposed. At the optimistic end of the scale, it has been suggested that
that there exists a small set of semantic relationships that NCs may imply. Levi (1978) proposes a
1
See http://www.newscientist.com/dailynews/news.jsp?id=ns9999678 for a description of this project.
I actually retained only those NCs for which both words could be found in a medical ontology.
3
Another methodology is example-based which consists in collecting a set of example compounds, most usually
without regard to the source, with the aim of arriving at as large a list of examples as possible. The list that results
contains compounds, but no contexts, so that the theory is based of ‘out of context interpretation’ of the examples. Of
the two theories that I analyze for this paper, Levi (1978) takes this approach, Warren (1978) the corpus-based one.
4
This short review is based on [Lauer 1995]
2
3
theory along these lines. If these theories are correct, they provide a firm basis for computational
interpretation, since they specify a concrete list of possibilities.
In contrast, the most pessimistic views claim that the implicit relationships between head and
modifier are entirely unconstrained. For example, Dowing (1977) performed a series of
psychological experiments to support her argument that the semantics of NC cannot be exhausted
by any finite listing of relationships. However, she does observe that certain types of relationship
are more common than others. If Downing’s hypothesis is true in a strong sense, then noun
compound interpretation requires complete pragmatic context-dependent inference and therefore
any attempt to handle noun compounds will involve detailed manual knowledge coding. However,
noun compound semantics do form certain patterns. There is certainly hope that adopting a more
restrictive theory that relies on these patterns will prove useful, even if there are cases that it is
unable to handle.
Between the two ends of this scale, there exists a range of noun compound establishes
nomenclatures. [Lauer 95] briefly describes four such theories5.
In order of the degree to which they constrain the possible semantics, they are:
•
•
•
•
Levi’s (1978) recoverably deleteable predicates and nominalizations
Leonard’s (1984) typology of compounds in fictional prose
Warren’s (1978) analysis of Brown corpus compounds
Finin’s (1980) role-filling theory, incorporating role nominals.
For this paper I consider Levi’s and Warren’s works. I chose to study [Levi 1978] because she
is cited in virtually every paper on computational analysis on NCs, because her theory that all NCs
can express only a very small set of semantic relations (13) is very attractive and one would like to
believe it and because her theory derives from a theory of language and not from empirical
investigation (like mine) and I am interested in seeing a different approach to this problem. I
decided to study Warren’s thesis for exactly the opposite reason, that is, as she says, her thesis “is
primarily a report of the results of an empirical investigation. In other words, my main interest has
been directed towards finding out facts about the semantic patterns of a particular type of
compound, rather than towards the question of how to account for the compounding process in
accordance with a particular theory of language”. This approach is very similar to mine; she
analyzed 4557 different compounds (while Levi’s collection consists only of about 200 examples)
and classified them according to the covert semantic relation they expressed and her far less
constraining theory seemed, at a first glance, more realistic than Levi’s.
4) Judith N. Levi: The Syntax and Semantics of Complex Nominals
In this Section I briefly review Levi’s theory and in the following one I compare her
classification to mine.
5
I should mention the fact that Lauer’s thesis is about the syntactic analysis of NCs (attachment) and, to a lesser extent,
about the semantic analysis using probabilistic methods.
4
From page 5, “The principal purpose of this book is the exploration of the syntactic and
semantic properties of complex nominals (CNs)6 in English, and the elaboration of detailed
derivations for these forms within a theory of generative semantics. The book also represents an
attempt to incorporate into a grammar of English a model of the productive aspects of complex
nominal formation”. The emphasis is therefore on the formation of NCs. On page 6 she claims:
“One of the most important claims of the present work is that all CNs must be derived by just two
syntactic processes: predicate nominalization and predicate deletion. The analysis of CNs derived
by the latter process will entail the proposal of a set of Recoverably Deletable Predicates (RDPs)
representing the only semantic relations which can underlie CNs. This set whose members are
small in number, specifiable, and probably universal, consists of these nine predicates: CAUSE,
HAVE, MAKE, USE, BE, IN, FOR, FROM and ABOUT.” The analysis of nominalization brings
four types of nominalized verbs (AGENT, PATIENT, PRODUCT and ACT) derived from two
types of underlying structures (Objective and Subjective).
The claims are big: a very small set of (universal)7 relationships, only two syntactic processes.
Moreover, she dismisses the ambiguity as being resolved by semantic, lexical and pragmatic clues:
“The fact that a given CN may be derived by the deletion of any of these predicates, or by the
process of nominalization, means that CNs typically exhibits multiple ambiguity. This ambiguity is,
however, reduced to manageable proportions in actual discourse by semantic, lexical and
pragmatic clues”. Unfortunately, automatic systems cannot take for granted such clues and must
deal with ambiguity (and indeed this is the hardest, unsolved problem of computational linguistics).
4.1) Derivation of CNs by Predicate Deletion
Levi claims that the variety of semantic relationships between head nouns and prenominal
modifiers in NCs is confined within a very limited range of possibilities. A first derivation of CNs
by predicate deletion is proposed to account for the larger part of these possible semantic
relationships for those NCs whose two elements are derived from the two arguments of an
underlying predicate, as in “apple pie” or “malarial mosquitoes”. This small set of specifiable
predicates that are recoverably deletable in the process of CN formation is made up of nine
predicates. These predicates, and only these predicates, may be deleted in the process of
transforming an underlying relative clause construction into the typically ambiguous surface
configuration of the CN.
The NCs derived by predicate deletion are derived by a sequence of transformation in which a
predicated was deleted to form a NC: for example, for viral infection (CAUSE) we have the
following set of transformations
1.
2.
3.
4.
Virus causes infection
Infection is caused by virus
Infection is virus-caused
Infection virus-caused
6
Complex nominals are “expressions with a head noun preceded by a modifying element which in some cases is a
noun, in others what appears to be an adjective” (page 2); “apple cake”, “presidential refusal” and “electric shock” are
examples of CNs. In contrast, I consider only sequences of nouns (for the time being).
7
In Section 6.4 Levi suggests that her principles are likely to reflect linguistic universals that constrain the use and
interpretation of NCs in all human language.
5
5. virus-caused Infection
6. viral infection
Similarly marginal note was derived by note in margin
Examples of CNs derived by deletion of these nine predicates are given in the following table.
RDP8
N1 < direct object of relative
clause
CAUSE1
N1 < subject of relative
clause
CAUSE2
tear gas
disease germ
malarial mosquitoes
traumatic event
mortal blow
drug deaths
birth pains
nicotine fit
viral infection
thermal stress
MAKE (productive;
constitutive, compositional)
MAKE1
MAKE2
honeybee
silkworm
musical clock
sebaceous glands
songbird
daisy chains
snowball
consonantal patterns
molecular chains
stellar configurations
HAVE (possessive/dative)
HAVE 1
HAVE 2
picture book
apple cake
gunboat
musical comedy
industrial area
voice vote
steam iron
manual labor
solar generator
vehicular transportation
soldier ant
target structure
professorial friends
consonantal segment
mammalian vertebrates
government land
lemon peel
student power
reptilian scales
feminine intuition
CAUSE (causative)9
USE (instrumental)
BE (essive/appositional)
IN (locative -spatial or
temporal)
field mouse
morning prayers
marine life
marital sex
autumnal rains
FOR (purposive/benefactive)
horse doctor
arms budget
avian sanctuary
8
9
Recoverably Deletable Predicate
In parenthesis, more traditional terms for the relationships expressed by the RDPs
6
FROM (source/ablative)
ABOUT (topic)
aldermanic salaries
nasal mist
olive oil
test-tube baby
apple seed
rural visitors
solar energy
tax law
price war
abortion vote
criminal policy
linguistic lecture
Table 1
4.2) Derivation of CNs by Predicate Nominalization
The second major group is composed of CNs whose head noun is a nominalized verb, and
whose prenominal modifier is derived from either the underlying subject or the underlying direct
object of this verb.
Levi classifies the CNs first according to categories determined by the head noun, and second,
according to categories determined by the pre nominal modifier. For the first case, Levi proposes
four nominalization types: Act Nominalizations (parental refusal, birth control), Product
Nominalizations10 (musical critiques, royal orders), Agent Nominalizations (city planner, film
cutter) and Patient Nominalizations11 (student inventions, mammalian secretions).
CNs whose head nouns are derived by nominalization (NOM CNs) may also be classified
according to the syntactic source of their prenominal modifier(s).
In this way the data may be divided into two major categories: Subjective NOM CNs, whose
prenominal modifier derives from the underlying subject of the nominalized verb, and Objective
NOM CNs, whose prenominal modifier derives instead from that verb’s underlying direct object.
SUBJECTIVE NOM CNs
Act:
Product:
parental refusal
manager attempts
cell decomposition
judicial betrayal
clerical error
peer judgments
10
Act can be paraphrased by “the act of ..(parents refusing)” while Product Nominalizations seem to represent some
object (usually though not always tangible) that is produced as the result of a specific action or event (outcome) and can
be paraphrased by “ that which is produced by (the act) of criticizing music
11
To clarify the distinction between Product Nominalizations and Patient Nominalizations, Levi presents the following
examples, in which (a) are Products and (b) Patients
a.
The managerial appointments infuriated the sales staff.
b.
The managerial appointees infuriated the sales staff.
a.
We expect to receive the senatorial nominations shortly.
b.
We expect to receive the senatorial nominees shortly.
7
Agent:
Patient:
faculty decisions
papal appeals
royal gifts
student inventions
presidential appointees
city employees
OBJECTIVE NOM CNs
Act:
Product:
Agent:
Patient:
birth control
heart massage
subject deletion
musical criticism
oceanic studies
chromatic analyses
stage designs
tuition subsidies
draft dodger
urban planner
acoustic amplifier
electrical conductor
-
Table 2
This classification is based on the meaning and source of the head noun alone and on the
syntactic source of their prenominal modifier; it does not depend on the semantic relationships
between the 2 nouns.
5) Classification Comparison
In this section, I compare the classification I developed for the medical NCs with Levi’s. I first
map to each of my classes, one (or more?) of hers and then I do the opposite. The first analysis will
allow me to see whether I can describe all of my classes under Levi’s classification (and if not,
when and, possibly, why) and if (and if so, in which cases) I have a one to many mappings. This
will also (hopefully) help me to detect problems in my classification.
The second analysis will allow me to see how broad her classes are and how specific (too?)
mine are and what are the classes of my classification that fall under the same ones in hers.
8
My 40 relations were found by iterative refinement based on looking at the 2245 extracted
compounds and finding commonalities among them. I developed a taxonomy of implicit semantic
relations with 2 levels of abstraction. The frequencies of the categories are very different12.
I have a definition (or paraphrase) for each relation and it was suggested that these definitions
should be different, for example, instead of N2 is such-and-such it might be better to say that the
whole NC is such and such. I agree that this formulation is more appropriate, however, how I will
define these relations will mainly depend on their use in the automatic task of text interpretation. I
still have to address this and for the moment I kept the definitions I already had.
TOP
CLASS
(when
applicable)
Name
Description
and Order
Lexicalized
(240)14
Examples
Levi’s
classifica
tion13
cd8 mab
crh neurons
vitamin k
genotype b
(not
included
in the
analysis)
As lexicalized NCs do not represent regular grammatical process and must be learned on an
item-by-item basis, Levi does not include them in her analysis. Her study is restricted to
endocentric complex nominals and does not consider, for example, metaphorical, lexicalized or
idiomatic meanings. Chapter 7 of [Levi 1978] deals with exception classes, i.e. types of NCs that
are exceptions to her theory and therefore not included (for example, CNs whose modifiers denote
units of measure, and CNs whose adjectival modifiers must be derived from adverbs rather than
nouns15).
12
Classes with very few examples are problematic for the classification algorithm and it would probably be convenient
to incorporate them into other more prolific categories. For this paper, however, I consider all the classes regardless of
the number of examples.
13
RDP stands for Recoverably Deletable Predicate, NON for Nominalization
14
The numbers in parenthesis are the numbers of NCs classified under this class
15
Other examples whose meanings (and hence, derivations) her theory does not predict are:
a.
b.
c.
d.
e.
f.
g.
h.
i.
coffee/leg/breast/pipe/baseball man
cat/morning/tennis person
volleyball/poetry/alfalfa/fitness/sitar freak (fiend, bug)
vodka/shopping/studying/yogurt/sweater binge
lunar/solar/fiscal /academic year
dental/medical/psychiatric/gynecological appointment
iron/bronze/industrial/space age
salad/hatcheck/coffee/copy girl
soap/diamond/brewery/textile heir(ess)
Her claims is that these NCs are formed not by the syntactic processes of predicate deletion or
nominalization but rather by a morphological process equivalent to suffixation
9
predicate
Another problematic category consists of those NCs for which a well-defined relationship could
not be determined16 (they are simply subtypes of one another like guinea pig17).
Subtype
(187)
(NC is
subtype of–
kind of head)
headaches migraine
weight peptides
albino mouse
gypsy woman
guinea pig
RDP: BE
This class can correspond to Levi’s BE. She identifies 3 major semantic subtypes in BE:
compositional (snowball, glass cup, bronze statue), “genus-species” (pine tree, cactus plant, winter
season, murder charge) and metaphorical (mother church, queen bee, finger lakes). My Subtype18
class roughly corresponds to her “genus-species” subtype, the metaphorical one being included in
Lexicalized and the “compositional” in another class defined later (Made of).
ACTIVITY: Consists of 4 subclasses: Activity/Physical process, Ending/reduction, Beginning,
Change, Produce (on a genetic level)
ACTIVITY
Change
(24)
n2 is a change
in/of n1
headache transformation
disease development
tissue reinforcement
tumor development
ventricle enlargement
cell growth
NOM:
Subjective
Act
tissue atrophy
tract stenoses
tissue tropism
?
My classification does not depend on whether the nouns are simply nouns or derived from a
verb by nominalization, rather, only the semantic relationship between the nouns is taken into
account. For this reason, I classify under the same class “disease development” and “tissue
atrophy”. For the nominalized nouns there is a corresponding class under Levi’s schema, namely,
NOM Subjective Act, but I could not find an appropriate RDP class for the other NCs. Levi’s RDP
schema cannot describe NCs such as “tissue atrophy”. The closest classes could be HAVE 2 (“the
tissue has a atrophy” ?) or IN (“atrophy in the tissue”) but this feels unsatisfactory to me. The first
noun is the subjective of the relative clause, but in the RDP classification the only cases in which
this happens are for CAUSE2, MAKE2, HAVE 2 (see Table1) neither of which is applicable here.
These NCs could be well described by the proposition “OF” that Levi does not include.
16
These two categories, Lexicalized and Subtype are very broad and, not surprisingly, they were the worst performers
for the algorithm.
17
This will prevent the system to classify “guinea pig” as a pig that lives in Guinea.
18
I indicate a class of my classification schema underlying the name
10
The same considerations apply to the following class:
Activity/Physical
process (63)
NC is an
activity of N2
type in which
N1 is involved
bile secretion
virus reproduction
infection transmission
bowel movement
headache activity
brain function
bile drainage
headache signs
cell necrosis
NOM:
Subjective or
Objective Act
?
Note that here the first noun can be either the object or the subject of the nominalized verb. For
example, the bile is secreted and the infection is transmitted (Objective Act) whereas the brain
functions and the viruses reproduce (Subjective Act). In the future I’ll probably need to divide this
class accordingly to this distinction.
Ending/reduction
(8)
n2 is ending/
reduction of n1
migraine relief
headache decrease
headache improvement
headache reduction
headache resolution
NOM:
Subjective
Product
Beginning of (1)
n2 is beginning
of n1
headache onset
?
I have only one example for this class but again I would need an OF RDP.
Produce (on a
genetic level) (40)
n2 produces n1
actin mrna
cmv dna
receptor mrnas
tyrosinase gene
chemokine genes
RDP:
MAKE1
CAUSE: Consists of 2 subclasses, Cause12 (where the first noun causes the second one) and
Cause21 (opposite direction). There is a good overlap with Levi’s CAUSE1 and CAUSE2
CAUSE
Cause 12 (102)
n1 is cause of n2
11
tension headache
aids death
RDP:
CAUSE2
Cause 21 (22)
n1 is effect on
body of n2 (n2 is
cause of n1)
automobile accident
heat shock
flu virus
influenza infection
immunodeficiency virus
diarrhoea virus
CAUSE2
RDP:
CAUSE1
PROPERTIES: Consists of 4 subclasses, Characteristic of, Physical property, Damage, Physical
make up
PROPERTIES
Characteristic of
n2 is property
of n1
Physical property
n2 is physical
property of n1
Defect (hindrance,
damage)
n2 defect/
hindrance/typ
e of damage
of n1
n2 is physical
make up of n1
Physical make up
cell immunity
disease prevalence
drug toxicity
headache persistence
gene polymorphism
artery diameter
artery pressure
artery hypoplasia
artery aneurysm
tract abnormalities
blood plasma
bile vomit
RDP:
HAVE2
(OF?)
RDP:
HAVE2
(OF?)
RDP:
HAVE2
(OF?)
RDP:
MAKE2
PURPOSE: Consists of 2 subclasses, Purpose and Function
PURPOSE
Purpose
(65)
Function
(2)
n1 is target of n2
n1 is the function of n2
anti-migraine drug
influenza immunization
flu vaccine
asthma therapy
RDP: FOR
headache medication
asthma treatment
tumor treatment
NOM:
Objective
Act
RDP: FOR
growth hormone
Under Levi’s classification “tumor treatment” and “asthma therapy” would be treated differently.
This would not be appropriate for the system I am developing for the automatic semantic analysis
of text.
12
PERSON AFFLICTED BY DISEASE: Person afflicted and Demographic attributes (of person
afflicted)
PERSON AFFLICTED
BY DISEASE
Person afflicted
(52)
n2 is person
afflicted
with n1
Demographic
attributes (19)
n1 is
demographic
attributes of
person
afflicted
with n2
migraine patient
bmt children
hiv populations
headache society
headache sufferer
aids patient
childhood migraine
childhood headache
infant infection
women migraineur
infant irritability
RDP:
HAVE1
RDP: IN
(or also
HAVE1)
TREAT/STUDY/ RESEARCH: Person/center who treats, Research on, Attribute of study,
Procedure
TREAT STUDY
RESEARCH
Person/center
who treats (19)
n2 is person,
center who
treats n1
Research on (8)
n2 is
research on
n1
Attribute of –
clinical- study
(103)
Procedure (60)
n2 is
procedure
for
d t ti /
13
headache specialist
headache center
diseases physicians
disease associations
headache unit
asthma nurse
asthma researchers
headache study
health research
RDP:
FOR
headache parameter
attack study
headache interview
biology analyses
biology laboratory
cell biology
headache pattern
bowel biopsy
genotype diagnosis
blood culture
headache diagnostics
RDP:
ABOUT
RDP:
ABOUT
RDP: ?
(OF?)
detecting/
studying n1
gallbladder pathology
disease pathophysiology
LOCATION: Located in, Part of, Occupied by
LOCATION
Located in
(274)
n2 is located in
n1
Is part of (5)
n2 is part of n1
Occupied by
(3)
n2 is occupied by
n1
liver abnormalities
brain abscess
breast carcinoma
lip mucosa
liver tumor
artery obstruction
dna element
artery segments
artery territory
brain region
RDP: IN
headache interval
attack frequency
football season
headache period
influenza season
fibrosis stage
morning headache
hour headache
weekend migraine
RDP: ?
(OF?)
RDP: IN
RDP:
HAVE 1
TIME: Time of (12), Frequency/duration/ time of
TIME
Frequency/durati
on/ time of 21
(35)
n2 is
frequency,
time of n1
Time of (4)
n1 is time of
n2 (opposite
of previous
class)
RDP: IN
Here again, I couldn’t find an appropriate RDP. OF would have been appropriate.
MEASURE: Measure of, Standard
MEASURE
Measure of (70)
n2 is a measure
of n1
Standard (5)
n2 is a standard
for n1
14
relief rate
asthma mortality
abortion rate
attack duration
hospital survival
headache criteria
society criteria
gold standard
society standard
hygiene standards
RDP: ?
(OF?)
RDP:
ABOUT
INSTRUMENT: Instrument (12), Instrument (21), Instrument (1)
INSTRUMENT
n1 is used
in/for n2
Instrument 12
(102)
Instrument 21 (4)
Instrument 1 (8)
n2 is used
for n1
n1 is used
injection method
laser surgeries
aciclovir therapy
vibration massage
lamivudine therapy
chloroquine treatment
ginseng treatment
laser treatment
biopsy needle
drug use
heroin use
internet use
tobacco use
drug utilization
RDP:
USE
NOM:
Subjectiv
e Act
RDP:
FOR
NOM:
Subjectiv
e Act
Here, like for “tumor treatment” and “asthma therapy” in Purpose (classified respectively under
Objective Act and FOR) “laser treatment” and “laser surgeries” (or “aciclovir therapy”) would be
classified differently (Subjective Act and USE) under Levi’s classification.
OBJECTIVE: Objective, Misuse
OBJECTIVE
Objective (33)
n1 is acted on by
n2 (n2 affects n2,
n2 acts on n1)
bowel transplantation
kidney transplant
liver transplantation
cell death
liver metastases
cell apoptosis
infant death
Misuse (11)
n2 is misuse(r) of
n1
drug abuse
substance abuse
ergotamine abuser
NOM:
Objective
Act
RDP: ?
(OF?)
NOM:
Objective
ACT
acetaminophen overdose
RDP: ?
(OF?)
15
Here again the RDP relations cannot describe, in my view, “infant death” nor “acetaminophen
overdose” for which OF would be a good description.
Other classes, with no superordinates.
Topic (46)
Modal (14)
Material (Made
of) (17)
Subjective (19)
Binds (4)
Activator 12 (6)
Activator 21 (4)
Inhibitor (13)
n1 is the
topic of n2
(n2 is
concerned
with n1)
n1 is how
n2 is done
n2 is
composed/
made by n1
n1 is subject
of n2
n2 binds to
n1
n1 activates
n2
(n1 is
activator of
n2)
n2 activates
n1
(n2 is
activator of
n1)
time visualization
headache questionnaire
tobacco history
vaccination registries
health education
RDP: ABOUT
emergency surgery
emergency nephrectomy
vacuum aspiration
trauma method
agarose gel
aloe gel
gelatin powder
gel capsule
latex gloves
calcium stones
RDP: BE
textile dust
formaldehyde vapors
headache presentation
headache response
FROM19
bile metabolism
drug metabolism
heat transfer
headache response
receptor ligand
carbohydrate ligand
receptor agonist
pain signal
food allergy
food infection20
receptor agonist
receptor antibodies
headache precipitant
headache trigger
receptor antagonist
19
RDP: HAVE 1
NOM:
Subjective Act
RDP: ? (OF)
RDP: FOR
RDP: ? (OF)
NOM:
Subjective
Product
RDP: FOR
NOM
Objective
Agent
RDP: ? (OF)
These nouns compounds could actually be divided into a FROM class and a MADE OF. For example, agarose gel,
aloe gel, and gelatin powder can be MADE OF and textile dust and formaldehyde vapors can be FROM.
20
“food infection” and “food allergy” are also Cause12. Many of my NCs have been classified with 2 (or more) labels.
It is in fact important to associate with a given NC all the possible meanings. For example, “eyelid abnormalities” is
both Located in and Defect, (while “growth abnormalities” is only Defect) and “tissue atrophy” is Subjective, Change,
Located in and Defect.
16
adrenoreceptor blockers
receptor blockers
virus inhibition
influenza prevention
NOM:
Objective
Agent
Food infection and food allergy are also Cause12. Many of my NCs have been classified with 2
(or more) labels. It is in fact important to associate with a given NC all the possible meanings. For
example, “eyelid abnormalities” is both Located in and Defect, (while “growth abnormalities” is
only Defect) and “tissue atrophy” is Subjective, Change, Located in and Defect.
Warren includes an interesting discussion about ambiguities. She says that ambiguity arises
because the semantic features of the constituent scan satisfy more than one semantic pattern and
she distinguishes three different kinds of ambiguous compounds (see page 67 of Warren 78).
In my application, I assume that all the different readings are possible, postponing the
disambiguation to the analysis of the whole text.
This analysis shows that the NCs (of my collection) belonging to the same semantic class
correspond to only one of the Levi relations21. I would hope that this is an indication of the internal
homogeneity of the classes I defined.
5.1) Exclusion of “OF”
I was surprised by the exclusion of the preposition “OF” from the set of recoverably deletable
predicates (especially because I wasn’t able to classify about 250 of my NCs).
On page 97, Levi lists some examples of CNs derived by FOR deletion (“bird sanctuary” from
“sanctuary FOR birds”, “nose drops” from “drops FOR nose”) and on page 95, examples of CNs
derived by IN deletion (“office friendships” from “friendships IN the office”, “marginal note”
from “note IN the margin”). It's not clear to me why she excludes CNs derived by OF deletion:
“headache activity” from “activity OF headache”, “brain function” from “function OF the brain”,
“cell growth” from “growth OF the cell”. On page 161, she discusses about the membership
requirements for the RDP set but she mainly defends her choice for the nine predicates rather than
explaining the reasons for the exclusion of others22. She claims that the predicates she proposes
“seem to embody some of the most rock-bottom-basic semantic relationships expressible in human
language; assuming that there are such things as semantic primes, these nine are surely
outstanding candidates”. This seems quite vague to me. She goes on saying that all the nine
predicates manifest “surface invisibility” and “grammaticality”. She defines “surface invisibility”
as a characterization of those predicate “that can be either deleted or incorporated before they
reach the surface without leaving a surface trace”. So for example the sign “employees only”
means “FOR employees only”, “Adar is as talented as Ariella” means “Adar is as talented as
Ariella IS” and “Max wants a lollipop” means “Max wants (TO HAVE) a lollipop”. I am not sure if
21
Except for the cases in which the different derivations of the nouns determines the classification as both NOM and
RDF (as we saw for Activity/Physical process, Change). I don’t deem this important for the semantic analysis I am
interested in.
22
She never explicitly mentions the preposition “OF”.
17
this applies to the preposition OF, but one example comes to mind: in a medical laboratory with
plates containing biopsies, a label “bowel” would mean “biopsy OF bowel”, and the label “brain”
would stand for “biopsy OF brain”.
The notion of grammaticality means that these predicates “are often expressed not by
independent lexical items but by bound grammatical morphemes”. For example, some predicates in
the RDP set are grammatized into marking of nouns, i.e. into case endings, and for example, in the
Proto-Indo-European system, CAUSE and FROM surface as ablative, FOR as dative, USE as
instrumental, IN as locative, MAKE as accusative, BE as nominative and HAVE as possessive
genitive. It is important to note here that the genitive case has been described only as possessive
genitive that can indeed be paraphrased by HAVE23. However, the genitive is not used only to show
possession but can also be indefinite, objective, partitive, descriptive24 in which cases the
preposition is OF25.
For these reasons, I would argue that the exclusion of “OF” is not well motivated.
I now compare the two classifications in the other direction.
Levi’s
classification
CAUSE1
CAUSE2
MAKE1
MAKE2
HAVE1
HAVE2
USE
BE
IN
My classification
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Cause 12
Cause 21
Produce (on a genetic level) (Activity)
Physical make up (Properties)
Person afflicted (Person afflicted by disease)
Occupied by (Location)
Material, made of
Characteristic of (Properties)
Physical property (Properties)
Defect (hindrance, damage) (Properties)
Instrument 12 (Instrument)
Subtype of
Modal
Demographic attributes (Person afflicted by
23
And this was in fact the case for the classes Characteristic of, Physical property, Defect for which both HAVE2 and
OF are good descriptions.
24
(Example taken from http://www.nd.edu/~archives/gen.htm: Latin Dictionary and Grammar Aid)
genitive charge.He is guilty OF MURDER.
genitive. indefinite. A box OF [some] WEIGHT.
genitive. description. He was a man OF NO CHARACTER.
genitive. material. A statue made OF SILVER.
genitive. objective. He had no fear OF DEATH.
genitive. partitive. One out OF A MILLION.
25
For example, I would define “bowel biopsy” and “acetaminophen overdose” as a objective genitives and “asthma
mortality” as description genitive.
18
FOR
FROM
ABOUT
NOM: Subjective
Act
NOM: Subjective
Product
NOM: Subjective
Patient:
NOM: Objective
Act
NOM: Objective
Product
NOM: Objective
Agent
•
•
•
•
•
•
•
•
•
•
•
disease)
Located in (Location)
Is part of (Location)
Time of (Time)
Function (Purpose)
Purpose (Purpose)
Person/center who treats
(Treat/study/research)
Instrument 21 (Instrument)
Binds
Activator 21
Material, made of
Research on (Treat/study/research)
Attribute of clinical study
(Treat/study/research)
Procedure (Treat/study/research)
Topic
Standard (Measure)
Misuse (Objective)
Activity/physical process (Activity)
Change (Activity)
Instrument 12 (Instrument)
Instrument 1 (Instrument)
Subjective
Ending/reduction (Activity)
Activator 12
•
•
•
Purpose (Purpose)
Misuse (Objective)
Objective (Objective)
•
•
Activator 21
Inhibitor
•
•
•
•
•
•
•
•
•
•
•
•
An obvious observation is that to most of Levi’s classes correspond many categories of my
classification schema (and very different ones). For example Demographic attributes, Located in, Is
part of and Time of all correspond to the RDP IN. This is hardly surprising; the purposes of the two
classification schemas are very different and Levi justifies this vagueness (as she puts it on page
82) “on three major grounds. First, the identification of the RDP set which they constitute allows us
to make highly accurate predictions about which semantic structures can underlie CNs and which
19
can not; second, the specification of general predicates serves to include a number of more specific
relations that might otherwise have to be enumerated individually with a concomitant loss of
generality; third, this analysis creates no more confusion or vagueness than is observable in
periphrastic equivalents with overt lexicalizations of the RDPs (e.g., x has y, x is in y, x is for y),
where some fundamental distinctions are made but where a great deal is typically left unspecified.”
She lists as an example, some of the possible more specific relationships for IN: inhabits, grow-in,
according-to, during, found-in and occur-in. She claims that: “The positing of six or more separate
groups, however, would obscure the generalization that is captured in the present theory by
positing a single predicate IN for both spatial and temporal, both concrete and abstract location.
[...] we must recognize that (a) it cannot be accidental that all of these paraphrases must in any
case include the locative preposition IN, and (b) the more specific verbs are wholly predictable on
the basis of the semantic characteristics of the surface components of the CN. For example, almost
any CN whose head noun denotes an animate creature and whose modifier is a place noun will be
a suitable member for the INHABIT group. [...] Since, however, these differences not only are
predictable on independent grounds but also seem to have no effect whatsoever on the formation of
the corresponding CNs, our description need not include this of specificity.”
The important point is that the purpose of her work is the formation of CNs, while mine is the
semantic analysis of CN by a computer. In order for the computer to predict the specific
relationships on the basis of the semantics of the nouns, I would need to define the specific readings
anyway and to add rules such as:
I.
II.
If the (high level) relationship is IN
If the first noun belongs to the class A (to be defined..) and the second noun
to the class B (also to be defined..)
THEN Ÿ IN is INHABITS
In real life (unfortunately?) rules are always flawed and also it would be difficult to define the
memberships; I think an approach in which NCs are directly mapped to more specific categories is
undoubtedly superior.
An appropriate question would be then, why consider Levi’s work if the goals are so different?
As I said, I wanted to analyze a linguistic theory on this subject and, as already mentioned, most of
the papers on computational linguistics for the analysis of NCs cite and, more importantly, some of
them justify their classification schemas as based on Levi’s (for example [Vanderwende 1994]).
Based on this analysis, I argue that this justification doesn’t hold and different classification
schemas are needed.
6) Warren, Beatrice. Semantic Patterns of Noun-Noun Compounds
Warren’s theory [Warren1978] is far less constraining. Her purpose was to determine whether
there exists a limited number of semantic relations between the constituents of NCs, and to
determine the nature of these relations. As I said in the introduction, [Warren1978] is primarily a
report of the results of an empirical investigation. She made a comprehensive study of compound
20
nouns in 360 thousand words of the Brown corpus26 and she manually extracted 4557 different
compounds; she then developed a taxonomy of implicit semantic relations, with four hierarchic
levels of abstraction. Before getting into the semantic relations, I’d like first to say a few words
about some other aspects of NCs that Warren describes in her thesis. Chapter I is a brief but fairly
comprehensive account of what is known about the structure of NCs in general. Warren enumerates
4 different structures for NCs:
1.
2.
3.
4.
5.
Morphological Structure
Syntactic Structure
Phonological Structure
Information Structure
Semantic Structure
1. Morphological Structure
Warren says that the most important point is that, as a rule, the first noun does not take inflectional
endings, such as plural or genitive suffices, but then she lists exceptions to this rule: plural firstwords (sports center, narcotics law) and first-words with genitive –s (driver’s seat, women’s
colleges)27.
2. Syntactic Structure
Warren discusses several ways of parsing NCs with more that 2 constituents28 (that she calls
compounds-within-compounds) and shows the frequency and distributions of various
combinations.
3. Phonological Structure
Warren talks about the stress pattern and identifies 2 fundamental stress patterns: fore-stress (or
first-stress) and double-stress. Fore-stress involves primary stress on the first noun and secondary
stress on the second noun (cówbòy). Double-stress involves heavy stress on both the first and
second constituents: stóne wàll. She also observes that certain semantic relations are connected
with certain stress patterns and that stress may be used to signal syntactic and semantic differences:
for example, compounds expressing material-artifact are usually pronounced with double stress
(bronze screws, brass wire) and compounds expressing Purpose or Origin have been assigned forestress.
4. Information Structure
Elements in a sentence are not equally important. Some elements are, communicatively, more
important than others and have therefore a different degree of communicative dynamism (CD).
Elements that have a low degree of CD within a sentence are called thematic, and those that carry a
high degree of CD, are called rhematic29. Warren suggests that the same analysis is possible for
NCs and that the two constituents of a NC do not normally have equal informative force and that in
26
In particular, her collection comes from “The standard Corpus of Present-Day Edited American English” assembled
at Brown University during 1963 and 1964.
27
Warren includes a short discussion about the controversy regarding such constructs and she identifies the semantic
relations of these compounds as either Purpose (driver’s seat) or Possession (pastor’s cap).
28
For example spider-leg pedestal table is left-branching within a left-branching structure, whereas home workshop
tool is right-branching within a left-branching structure
29
Warren is referring here to the Prague school of linguistics (Fibras)
21
NCs the first position is reserved to the element with the highest degree of CD; the “information
structure” of a compound is therefore that the rhematic elements tends to precede the thematic
element30.
5. Semantic Structure
Comment-topic structure: NCs can be described as having a bi-partite structure, in that they consis
of a topic and a comment. The topic is what we are talking about (represented by the head of the
NC) and the comment is what we say about the comment. The function of the comment is to make
the reference of the topic more precise. It restricts the reference scope of the topic by either
classifying what kind of topic we have in mind, or by identifying which of a number of known
topics we are referring to. The first case refers to NCs in which the comment has classifying
function (paraphrasable by: “that kind of –topics- that (verb) –comment-”)31, and the second one, to
NCs in which comment has identifying function (paraphrasable by: “that -topics- that (verb) (some
specific) –comment-”).32 Warren hypothesizes that compounds with identifying comments don’t
allow generic reference, in contrast to compounds with classifying comments. Within the semantic
structure, Warren lists also Idiomatization, Established and Novel Compounds33 and finally the
Partecipant Roles: it is obvious that the constituent of the NC are semantically related and that the
semantic relations are not the same: it is indeed the purpose of [Warren 1978] to established
whether there is a limited number of possible types of semantic relations between the constituents
of compounds and what these semantic relations are. Accordingly to Warren’s results, there are six
major types of semantic relations than can be expressed in compounds:
1) A is something that wholly constitutes B34, or vice versa: this is the CONSTITUTE
class divided into Source-Result, Result-Source and Copula classes. Some examples:
metal coupon, paste wax, student group.
2) A is something of which B is a part or a feature or vice versa: POSSESSION class
divided into Part-Whole, Whole-Part and Size-Whole classes. Examples: board
member apple pie.
3) A is the location or origin of B in time or space: LOCATION. Participant roles are
Place-OBJ, Time-OBJ, Origin-OBJ. Examples: coast road, Sunday school, seafood.
4) The comment indicates the PURPOSE of B. Participant roles: Goal-Instrumental.
Examples: pie tin, tablecloth, football.
5) The comment indicates the activity or interest with which B is habitually concerned.
Participant roles: ACTIVITY-ACTOR. (cowboy, fire department)
6) RESEMBLANCE. A indicates something that B resembles. Participant roles are:
Comparant - Compared. Examples: egghead, bullet head.
The semantic relations can often be paraphrased35.
30
My results for the automatic classification seem to contradict this hypothesis (see Rosario 2001).
Like bilge water
32
Like: (the) bathroom door, the Podger cat (where Podger is a family name)
33
Established compounds are those combination that have become the generally accepted word for their referent (like
toothpaste instead of mouth hygiene paste) and novel compounds are combination that have been created for the
occasion
34
In Warren’s notation, A indicates the first noun, B the second one.
35
Paraphrasing does not work with idiomatization.
31
22
MAJOR CLASS
CONTITUTE
POSSESSION
LOCATION
PURPOSE
ACTIVITY-ACTOR
RESEMBLANCE
SUB CLASSES
Source-Result
Result-Source
Copula
Part-Whole
Whole-Part
Place-OBJ
Time-OBJ
Origin-OBJ
Goal-Instrumental
Activity-Obj
Comparant- Compared
PARAPHRASE
OF
IN
OF
WITH
IN, AT, ON
IN, AT, ON
FROM
FOR
LIKE
EXAMPLES
metal sheet
sheet metal
Girl friend
eggshell
armchair
coast road
Sunday school
seafood
pie tin
cowboy
cherry bomb
These classes are further subdivided. I won’t describe the classes in detail (there is a whole
chapter for each class) but in the following pictures I report the hierarchical structures for these
semantic classes and in the next section I’ll describe some of the classes in greater detail.
CONSTITUTE
Figure 1
23
POSSESSION
Figure 2
LOCATION
Figure 3
24
PURPOSE and ACTIVITY-ACTOR
Figure 4
Warren does not include NCs in which one noun is deverbal (the nominalizations in Levi’s
terminology). Her motivation for doing so is that a main goal of her work is to find out what
semantic relations between two nouns can be left unexpressed, or what verb connecting the two
nouns may be discarded. In compound containing a deverbal noun there is no verb deleted and the
type of semantic relation that exists between the constituents is explicit while this relation is
implicit in non-verbal compounds. An important consequence of this distinction is that
nominalizations do NOT appear to be restricted as to the type of relations that may exists between
the constituents (at least not in parallel fashion). Indeed, there is no need for such a restriction since
the type of relation between the nouns is explicitly indicated by the deverbal noun. In contrast,
there is a limited number of relations that are left unexpressed for non-verbal NCs.
As I already said, I don’t distinguish between verbal and non-verbal NCs (although some of the
classes contain mainly one type or the other) and as we’ll see for the classes with mainly verbal
NCs, I don’t find a correspondent class in Warren’s classification.
If the fact that in verbal NCs the relationships are explicit in the verb is true, in my application,
I could first divide NCs into verbal and non-verbal (with a binary classification system) and then
analyze the two classes separately. This would make the system much more complicated but it
would be useful if the number of possible types of relations in verbal NCs is indeed unconstrained
and if we knew how to infer the relation given the verb (which is another problem in it own!).
[Lapata 00] proposes algorithms for the classification of nominalizations according to whether the
modifier is the subject or the object of the underlying verb expressed by the head noun; my aim is
to have more specific relations, sufficiently general to cover a significant number of noun
compounds, but that can be domain specific enough to be useful in analysis and subject-object is
not specific enough.
7) Classification Comparison
In this section, I compare the classification I developed for the medical NCs with Warren’s.
25
TOP
CLASS
(when
applicable)
Name
Description
and Order
Lexicalized
(240)
Examples
Warren’s
classification
cd8 mab
crh neurons
vitamin k
genotype b
csf culture
Proper names
Warren says that the only NCs in her collection that can be said to lack a semantic relation between
the members are compounds that in some sense are proper names, like for example X-ray, Y-region,
Q-fever. She does not include them in the classification schema
Subtype
(187)
(NC is
subtype of–
kind of head)
headaches migraine
albino mouse
gypsy woman
CONTITUTE
(COPULA)
Copula NCs are those noun compounds in which both members cane be said to be two alternative
“names” for the same referent (girl friend, cypress tree). She further divides the copula class into
Attributive NCs (NCs that represent a cross-classification, like girl friend, houseboat) and
Subsumptive NCs (those that are based on a Species genius relation, like cypress tree, panel board.
(See Figure1). I don’t distinguish between Attributive and Subsumptive.
ACTIVITY
Activity/Physical
process (63)
NC is an
activity of N2
type in which
N1 is involved
bile secretion
virus reproduction
infection transmission
headache activity
brain function
cell necrosis
headache signs
bile drainage
?
I couldn’t find a class for these NC s which in fact are mainly nominalizations that are not covered
by Warren. The closest class could be ACTIVITY – ACTOR (Figure 4) of crime syndacate but
Warren says that for all the NCs in this class, the head refers to a human being, a group of people or
an organization which is not the case in my class Activity/Physical process and perhaps more
importantly, the order is different: the paraphrases she proposes for these NCs is “the head is
habitually concerned with the modifier” while I have just the opposite. Some of these NCs could
also be thought as LOCATION (brain function bile drainage) but in her LOCATION class, the
26
heads (called also topics) can only be objects (See Figure 3) while in my case, the heads are really
activities, not objects. The same comments apply to the following classes
n2 is a change
in/of n1
headache transformation
disease development
tissue reinforcement
tumor development
ventricle enlargement
cell growth
tissue atrophy
tract stenoses
tissue tropism
Ending/reduction
(8)
n2 is ending/
reduction of n1
migraine relief
headache decrease
headache improvement
?
Beginning of (1)
n2 is beginning
of n1
headache onset
?
Change (24)
?
Again, we have here mainly nominalizations.
CAUSE: Consists of 2 subclasses, Cause12 (where the first noun causes the second one) and
Cause21 (opposite direction).
CAUSE
Cause 12 (102)
n1 is cause of n2
Cause 21 (22)
n2 is cause of n1
tension headache
aids death
automobile accident
heat shock
flu virus
influenza infection
immunodeficiency virus
(LOCATION,
Origin-Object)
Causer-Result
(Just mentioned)
It interesting to note here that while for my class Cause 12 there is a corresponding class CauserResult of hay fever (Figure 3), Warren just mentions that “there are some compounds where the
topic can be assigned the role Causer: disease agent, lifeblood” (page189), but she found only two
NCs of this type does not include this class in her classification.
27
PROPERTIES
cell immunity
disease prevalence
drug toxicity
headache
persistence
gene polymorphism
artery diameter
artery pressure
BELONGING
TO (Whole-Part)
(Whole-feature)
artery hypoplasia
artery aneurysm
tract abnormalities
BELONGING
TO (Whole-Part)
(Whole-feature)
Characteristic of
n2 is property
of n1
Physical property
n2 is physical
property of n1
Defect (hindrance,
damage)
n2 defect/
hindrance/typ
e of damage
of n1
Physical make up
n2 is physical blood plasma
make up of n1 bile vomit
bile salt
BELONGING
TO (Whole-Part)
(Whole-feature)
LOCATION
(Origin-Obj)
Place of OriginOBJ
Warren describes the Whole-feature class as the class of NCs in which B represents an inherent
delimitable “part” of A. She further subdivides this class into Object-Quality (room temperature),
Object-Extension (particle size) and Object-Abrstact Shape (crime trend) (see Figure 2). My
classes don’t reflect these subclasses but I think that “Whole-feature” is a good description for
them.
PURPOSE
Purpose
(65)
anti-migraine drug
influenza immunization
flu vaccine
asthma therapy
n1 is target of n2
PURPOSE
(GoalInstrument)
headache medication
asthma treatment
tumor treatment
Here also, I think that Warren’ s class Goal-Instrument is similar to mine, but it’s subclasses are
less so although maybe “primary relation involves Instrumental” can work.
PERSON AFFLICTED BY DISEASE: Person afflicted and Demographic attributes (of person
afflicted)
PERSON AFFLICTED
BY DISEASE
Person afflicted
n2 is person
28
migraine patient
BELONGING
(52)
afflicted
with n1
bmt children
hiv populations
headache society
headache sufferer
aids patient
TO (Part Whole), (PartObj)
Belonging Possessor
Demographic
attributes (19)
n1 is
demographic
attributes of
person
afflicted
with n2
childhood migraine
childhood headache
infant infection
women migraineur
infant irritability
BELONGING
TO (WholePart),
Possessor Belonging
Belonging –Possessor (of gunman) and Possessor –Belonging (family estate) are opposite classes
within the BELONGING TO major class (see Figure 2). Belonging –Possessor is sub-class of Part
–Whole while Possessor –Belonging is sub-class of Whole- Part. This is logical, however, the
results is that these two classes that are the natural opposite of each other are “located” in two
positions in the hierarchy that are quite apart from each other. I found this quite confusing. Just
looking at the hierarchy is not immediately evident that these two classes are so close related.
TREAT/STUDY/ RESEARCH: Person/center who treats, Research on, Attribute of study,
Procedure
TREAT STUDY
RESEARCH
Person/center
who treats (19)
n2 is person,
center who
treats n1
Research on (8)
n2 is
research on
n1
headache specialist
headache center
diseases physicians
disease associations
headache unit
asthma nurse
asthma researchers
headache study
health research
headache parameter
attack study
headache interview
biology analyses
biology laboratory
cell biology
headache pattern
Attribute of –
clinical- study
(103)
ACTIVITY –
ACTOR (A is
activity
performed by B)
CONTITUTE
(Source Results)
SubjectMatter-Whole
CONTITUTE
(Source Results)
SubjectMatter-Whole
NCs in the Subject-Matter-Whole class can be paraphrased as “B that concern A” (gangster story,
tax issue)
29
LOCATION
Located in
(274)
n2 is located in
n1
Is part of (5)
n2 is part of n1
liver abnormalities
brain abscess
breast carcinoma
lip mucosa
liver tumor
artery obstruction
dna element
artery segments
LOCATION
(Place-Object)
(ghetto street,
school friends)
BELONGING TO
(Whole -Part)
Whole -Discrete,
Integral Part
(like spoon handle
and hillside)
For the Whole -Discrete, Integral Part NCs, B is a discrete delimitable part of A.
TIME: Time of (12), Frequency/duration/ time of
TIME
Frequency/durati
on/ time of 21
(35)
n2 is
frequency,
time of n1
Time of (4)
n1 is time of
n2 (opposite
of previous
class)
headache interval
attack frequency
football season
headache period
influenza season
fibrosis stage
morning headache
hour headache
weekend migraine
BELONGING TO
(Part- Whole)
OBJ-Time (golf
season)
BELONGING TO
(Size- Whole)
Duration-Whole
(3-day affair)
Also:
LOCATION,
Time-OBJ
(weekend guests,
Sunday paper)
For the “Time of” relation, Warren distinguishes NCs for which the comment indicates when the
topic takes place (morning concert) (LOCATION) with NCs for which the comments indicates the
duration of the topic (75-minute concert) (BELONGING TO -Size- Whole). I agree with this
distinction, however, (like for the Possessor-Belonging case) I find the classification of such close
classes into two different major classes quite confusing.
30
MEASURE: Measure of, Standard
MEASURE
Measure of (70)
n2 is a measure
of n1
relief rate
asthma mortality
abortion rate
attack duration
hospital survival
Standard (5)
n2 is a standard
for n1
headache criteria
society criteria
gold standard
society standard
hygiene standards
BELONGING TO
(Whole-Part)
Whole-Feature.
Could be OBJQuality of room
temperature or
OBJ-Extension of
particle size
BELONGING TO
(Whole-Part)
Whole-Feature
INSTRUMENT: Instrument (12), Instrument (21), Instrument (1)
INSTRUMENT
Instrument 12
(102)
n1 is used
in/for n2
Instrument 21 (4)
n2 is used
for n1
Instrument 1 (8)
n1 is used
injection method
laser surgeries
aciclovir therapy
vibration massage
lamivudine therapy
chloroquine treatment
ginseng treatment
laser treatment
biopsy needle
transport proteins
drug use
heroin use
internet use
tobacco use
drug utilization
COPULA,
Source -Result
PURPOSE
(GoalInstrument)
?
Instrument 12 could correspond to COPULA, Source –Result but looking at the sub-classes of
Source –Result (See Figure 1) I can’t find an appropriate one (the closest one is perhaps MaterialArtefact of clay bird and brass wire).
OBJECTIVE: Objective, Misuse
31
OBJECTIVE
Objective (33)
Misuse (11)
n1 is acted on by
n2 (n2 affects n2,
n2 acts on n1)
n2 is misuse(r) of
n1
bowel transplantation
kidney transplant
liver transplantation
cell death
liver metastases
cell apoptosis
infant death
drug abuse
substance abuse
ergotamine abuser
?
?
acetaminophen overdose
These two classes do not have a correspondence in Warren’s classification, again for the fact that
the NCs are mainly verbal-compounds. Perhaps, “PURPOSE Primary relation involves location”
(see Figure 4) could describe some NCs like kidney transplant, liver transplantation although her
examples for this class are quite different: water bucket, tablecloths.
Topic (46)
n1 is the
topic of n2
(n2 is
concerned
with n1)
time visualization
headache questionnaire
tobacco history
vaccination registries
health education
Modal (14)
n1 is how
n2 is done
Material (Made
of) (17)
n2 is
composed/
made by n1
emergency surgery
vacuum aspiration
trauma method
agarose gel
aloe gel
gelatin powder
gel capsule
latex gloves
calcium stones
textile dust
formaldehyde vapors
CONTITUTE,
Source-Results,
Non Material
Substance-Whole,
Subject-MatterWhole (detective
story)
CONTITUTE,
Source-Results
CONTITUTE,
Source-Results,
Material-Artefact
or also
BELONGING TOPart-Whole, PartObj
For Material, we would have the CONTITUTE for NCs in which what is indicated by one member
is what wholly constitutes the other (clay bird), however if the substance is only part of the whole
we should use the BELONGING TO class.
32
Subjective (19)
Binds (4)
Activator 12 (6)
headache presentation
headache response
bile metabolism
drug metabolism
heat transfer
headache response
receptor ligand
n2 binds to n1
carbohydrate ligand
receptor agonist
n1 activates n2
(n1 is activator of pain signal
food allergy
n1 is subject of
n2
BELONGING
TO (WholePart) Possessor
- Belonging)
?
?
n2)
Activator 21 (4)
Inhibitor (13)
food infection
receptor agonist
receptor antibodies
headache precipitant
headache trigger
receptor antagonist
adrenoreceptor blockers
receptor blockers
virus inhibition
influenza prevention
n2 activates n1
(n2 is activator of
n1)
?
?
The main problem I found in comparing my classification with Warren’s was for those relations
that have mainly verbal compounds that Warren does not include. I stress again the fact that non
including such compounds is not a viable option for my problem. It would be interesting to study
whether it is indeed straightforward to infer the relations from the verbs and what are the
“mechanisms” to do so. However, this would not solve the problem that I want headache
transformation and tissue atrophy to be classified as having the same semantic relation. Warren
acknowledges this fact and notice that her distinction would bring her to include sugar bowl while
excluding sugar container as a verbal combination in spite of the fact that the two NCs express the
same semantic relationships.
Other differences were due to the very different nature of our corpora. I find her classification
very detailed and I especially liked how she defined the sub-classes. I wasn’t always happy with the
organizations of her classes. I already mentioned the fact that she puts Possessor-Belonging and
Belonging-Possessor and Duration-Event (her Duration-whole) and Time-Event (her Time-Obj)
quite apart in the hierarchy, obscuring therefore the similarity of these relations. The fact that
Duration-Event is a size-whole (and therefore a BELONGING TO) kind of relationships was more
relevant in her classification than its similarity with Time-Event. If we don’t use the hierarchy but
only the classes in their own this is not very important, however, if we do want to take advantage of
the hierarchical structure, I think that the fact that 3-day affair ends up being a BELONGING
relation while Saturday-Sunday affair is a LOCATION relation can be problematic.
33
References
[Dowing 1977] Dowing, Pamela Ann, 1977. On the creation and Use of English Compound
Nouns. Language, 53:810-842.
[Finin, 1980] Finin, Tim 1980. The semantic Interpretation of Compound Nominals. PhD
Thesis, University of Illinois, Urbana, IL.
[Humphreys 1998] Humphreys et al. 1998. The unified medical language system: An
informatics research collaboration. Journal of the American Medical Informatics Association
5(1):1--13.
[Lapata, 2000] Maria Lapata, The Automatic Interpretation of Nominalizations, AAAI
Proceedings 2000.
[Lauer, 1995] Mark Lauer. Designing Statistical Language Learners: Experiments on Noun
Compounds. PhD Thesis.
[Leonard 1984] Leonard Rosemary, 1984. The Interpretation of English Noun Sequences on
the Computer. North-Holland,Amsterdam.
[Levi, 1978] Judith Levi, 1978. The Syntax and Semantics of Complex Nominals. Academic
Press, New York.
[Lowe 1994] Lowe and Barnett 1994. Understanding and using the medical subject headings
(MeSH) vocabulary to perform literature searches. Journal of the American Medical Association
(JAMA), 271(4):1103--1108.
[Rosario 2001] Barbara Rosario and Marti Hearst, 2001. “Classifying the Semantic Relations in
Noun Compounds via a Domain-Specific Lexical Hierarchy”. To appear in the Proceeding of
EMNLP 2001 (can be found at: http://www.sims.berkeley.edu/~rosario/papers.html)
[Swanson97] Don R. Swanson and N. R. Smalheiser, 1997. “An interactive system for finding
complementary
literatures:
a
stimulus
to
scientific
discovery”
Artificial Intelligence, 91:183-203.
[Vanderwende 1994] Lucy Vanderwende 1994. Algorithm for automatic interpretation of
noun sequences. Proceedings of COLING-94, pages 782—788.
[Warren, 1978] Warren, Beatrice, 1978. Semantic Patterns of Noun-Noun Compounds.
Gothenburg Studies in English 41, Gothenburg, Actr Universitatis Gothoburgensis.
34