Figuring Metaphorical Dimensions of Meaning in Political Conflict

Figuring Metaphorical Dimensions of Meaning in
Political Conflict Discourse
Andrew Gargett & John Barnden
School of Computer Science
University of Birmingham
United Kingdom
International Cognitive Linguistics Conference
Newcastle
Summer 2015
Outline
Background
Understanding Metaphor
PoliCon corpus
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
GenMeta project
I
http://www.cs.bham.ac.uk/~gargetad/genmeta-index.html
Gargett
Generating Metaphor
ICLC13 Summer 2015
3 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
The ubiquity of metaphor
I
We can use metaphor to “ground” what we say in the world around us
“from cradle to grave”
“getting into a relationship”
“no strings attached”
“surfing the internet”
I
So metaphor allows us to link relatively abstract concepts to more
common and everyday meanings, and this is pervasive in everyday
communication
Gargett
Generating Metaphor
ICLC13 Summer 2015
4 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
The ubiquity of metaphor
I
We can use metaphor to “ground” what we say in the world around us
“from cradle to grave”
“getting into a relationship”
“no strings attached”
“surfing the internet”
⇒
⇒
⇒
⇒
Life AS Journey
Relationship AS Container
Obligations AS Bonds
Internet AS Ocean [??]
I
So metaphor allows us to link relatively abstract concepts to more
common and everyday meanings, and this is pervasive in everyday
communication
I
Conceptual metaphor theory (e.g. Lakoff & Johnson 1980) directly
models this ubiquity of metaphor
Gargett
Generating Metaphor
ICLC13 Summer 2015
4 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
GenMeta project aims: generate “natural” metaphors
I
Our ultimate ambition is to automatically generate metaphorical
expressions which are more “natural” relative to specific illness
domains
1. What type of metaphor should be used?
eg viewing ILLNESS:
i More common: as an ATTACK → “an asthma attack”
ii More novel: as a RIDE → “the diabetes rollercoaster”
2. How conventional should the expression be?
eg viewing ILLNESS as a POSSESSION:
i More conventional: ”I have a cold”
ii Less conventional: ”I possess a cold”
3. What degree and polarity of sentiment to express?
eg viewing ILLNESS as an ATTACK:
i Relatively negative NP: ”asthma attack”
ii Relatively positive VP: ”preventing asthma attack”
Gargett
Generating Metaphor
ICLC13 Summer 2015
5 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
But generating “natural” metaphors requires examples
As a basis for generating metaphor, we also use Natural Language
Understanding (NLU), to automatically detect and interpret [harder]
metaphors
I NLU can be done by “training” a computer to find features of
metaphor
eg Conceptual features:
I
1. Concreteness: how perceivable & real is the meaning of this word?
[=non-abstract]
2. Imageability: how easily does this word evoke an image?
3. Affective meanings [details below]
eg Linguistic features:
1. Part-of-speech: e.g. nouns, verbs, adjectives, prepositions
2. Dependency relations: asymmetric relations between words in a clause,
typically head & their dependents (e.g. subject-verb, verb-object,
noun-modifier of noun)
Gargett
Generating Metaphor
ICLC13 Summer 2015
6 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
Knowing what to look for
I
Metaphorical meaning has a clear conceptual dimension
More concrete/imageable
concept
“source” domain
egs
“getting into unix”
“getting into government”
Gargett
Generating Metaphor
7−→
⇒
⇒
more abstract
concept
“target” domain
Operating system AS container
Government AS container
ICLC13 Summer 2015
7 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
Knowing what to look for
I
Metaphor is not only conceptual, but also linguistic
∗ Metaphor interacts with part-of-speech (POS)
eg While content words (e.g. nouns, verbs) often seen as prototypically
metaphorical, for particular genres (e.g. academic texts) prepositions
are the most frequent POS for fig. lang. (Steen et al. 2010)
∗ Also: there is interaction between metaphorical expressions and
syntactically defined contexts (e.g. phrase, clause, sentence)
eg Krishnakumaran & Zhu (2007) studied three types of syntactically
defined contexts for the occurrence of metaphors:
I: subject to be object (e.g. “EU is a family”)
II: subject verb object (e.g. “This issue absorbed all his waking
moments”)
III: adjective noun (e.g. “sweet thought”)
Gargett
Generating Metaphor
ICLC13 Summer 2015
7 / 26
Background NLU PoliCon
Motivations Producing meta4 Finding meta4 Definitions
Knowing what to look for
I
Metaphor also seems to have a strong affective dimension
∗ Metaphor has important role in expressing emotional content in
natural language (Kovecses 2003, Meier & Robinson 2005)
∗ Speakers may consistently & reliably rate individual words for levels of
affective meaning/sentiment (e.g. Warriner et al.)
but: what about larger text (e.g. phrases, etc)
eg: “the journey of life”, “strings attached”
Gargett
Generating Metaphor
ICLC13 Summer 2015
7 / 26
Outline
Background
Understanding Metaphor
PoliCon corpus
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
The Amsterdam Metaphor Corpus (VUAMC)
I
Vrije University Amsterdam Metaphor Corpus (VUAMC):
∗ approx. 188K words
∗ from the British National Corpus-Baby (BNC-Baby)
∗ annotated for metaphor using the Metaphor Identification Procedure
(MIP) (Steen et al., 2010)
∗ four registers [44K – 50K words each]: academic texts, news texts,
fiction, and conversations
http://ota.oucs.ox.ac.uk/headers/1054.xml
Gargett
Generating Metaphor
ICLC13 Summer 2015
9 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
The Amsterdam Metaphor Corpus (VUAMC)
I
Metaphor Identification Procedure (MIP):
1. For each lexical unit (LU), establish its meaning in context
2. For each LU, does it have a more basic meaning? Basic meaning:
∗
∗
∗
∗
More concrete
Related to bodily action
More precise (vs. vague)
Historically older
[Not necessarily the most frequent meanings]
3. If yes to 2., does current meaning contrast with the basic meaning, yet
can be understood in comparison with it?
4. If yes to all of the above, label LU as metaphorical
Gargett
Generating Metaphor
ICLC13 Summer 2015
9 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
The Amsterdam Metaphor Corpus (VUAMC)
I
Augmenting the VUAMC
∗ Aim: detect metaphor, using features of our target texts
∗ VUAMC is for us a training and testing resource
∗ VUAMC words are marked as metaphorical or unmarked, and we will
add to each word:
1. Conceptual information about:
o “measurements” of concreteness of each word
o “measurements” of imageability of each word
o “measurements” of affective meanings of each word
2. Linguistic information about:
o part-of-speech of each word
o the syntactic context in which the word occurs
Gargett
Generating Metaphor
ICLC13 Summer 2015
9 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
“Concreteness” & “Imageability” vs. Affective meanings
I
Using scores from MRC Psycholinguistic Database (Wilson, 1988)
∗ dictionary of 150,837 words
∗ different subsets of these words having been rated by human subjects
in a variety of psycholinguistic experiments [26 psycho/linguistic
attributes]
∗ includes 8,228 words rated with degrees of concreteness
also 9,240 words rated for degrees of imageability
∗ ratings range from 100 to 700 ⇒ a higher score indicating greater
concreteness/imageability [∼ scale of 1 to 7]
Gargett
Generating Metaphor
ICLC13 Summer 2015
10 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
“Concreteness” & “Imageability” vs. Affective meanings
I
It has been long recognised that concreteness & imageability scores
are highly correlated (Paivio et al., 1968)
∗ But: Dellantonio et al., (2014) show interesting differences between
these scores, based on interaction with sentiment
I
MRC scores have been used extensively in metaphor detection (e.g.
Neuman et al., 2013; Turney et al., 2011; Tsvetkov et al., 2013)
∗ We are interested in the use of such resources for investigating the
cognitive aspects of figurative meanings, as well as the computational
applications
Gargett
Generating Metaphor
ICLC13 Summer 2015
10 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
“Concreteness” & “Imageability” vs. Affective meanings
I
I
Using scores from Affective Norms for English Words (ANEW,
Warriner et al. 2013, Bradley & Lang 1999)
Dictionary of 13,915 English content words, rated for:
1. Valence [V]: how pleasant is the concept denoted by this word?
[“relaxing” vs. “murder”]
2. Arousal [A]: what is the emotional intensity of the concept denoted
by this word? [“rampage” vs. “calm”]
3. Dominance [D]: what is the level of control associated with the
concept denoted by this word? [“self” vs. “earthquake”]
∗ For each of category, we have statistics (mean, standard deviation)
∗ We record these for:
1. heads
2. avgs over dependents of heads [avg deps]
Gargett
Generating Metaphor
ICLC13 Summer 2015
10 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
Machine learning studies
I
Which combination of which features is optimal for detecting
metaphor?
⇒ Building the models:
Full:
Minimal:
MRC:
ANEW:
Base:
POS + heads[MRC + ANEW] + avgd deps[MRC + ANEW]
POS + heads[MRC] + avgd deps[MRC] + avgd deps[D.SD
POS + heads[MRC] + avgd deps[MRC]
POS + heads[ANEW] + avgd deps[ANEW]
heads[MRC]
& D.RAT]
⇒ Results (F measure):
Gargett
Full
Minimal
MRC
ANEW
Base
0.7813
0.7362
0.7275
0.7144
0.6906
Generating Metaphor
ICLC13 Summer 2015
11 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
Overview of approach to interpretation (1)
(1) Adding labels for semantic domains to items in the corpus
* Using the tagger for the UCREL Semantic Analysis System (USAS,
Rayson et al. 2004)
* USAS employs semantic field taxonomy, 21 major discourse fields
splitting into 232 semantic tags
* The tagger assigns tags to words and multiword expressions (MWEs)
to represent a coarse-grained word sense
* USAS uses lexical (through a lexicon knowledge-base), sentential
(through contextual features) and discourse-level (via domain
selection) features to assist its contextual disambiguation
⇒ http://ucrel.lancs.ac.uk/usas/
Gargett
Generating Metaphor
ICLC13 Summer 2015
12 / 26
Background NLU PoliCon
Data Metaphor detection Metaphor interpretation
Overview of approach to interpretation (2)
(2) Inferring a metaphorical relationship, from linguistic and other
features of the surrounding context
I
For example, based on syntactically defined metaphor “types”
(following Krishnakumuran & Zhu)
(3) This “surrounding context” could be anything from relatively close
linguistic context, up to much wider discourse context
⇒ “Other” features of the surrounding context could include:
* Concreteness and/or Imageability
* Affective meaning (e.g. valence, arousal, dominance)
Gargett
Generating Metaphor
ICLC13 Summer 2015
13 / 26
Outline
Background
Understanding Metaphor
PoliCon corpus
Background NLU PoliCon
Design of corpus
I
Texts from online opinion pieces, on news broadcaster website
(Australian Broadcasting Corporation):
(1) Texts consist of target article + comments from a open readership
(2) The comments have an embedded structure, with comments in
response to other comments in response to the target article, etc
(3) The items in the corpus are automatically annotated with [details given
earlier for all of these]:
*
*
*
*
Gargett
parts-of-speech
dependency relations
semantic domains
metaphors
Generating Metaphor
ICLC13 Summer 2015
15 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
PoliCon corpus eg
Gargett
Generating Metaphor
ICLC13 Summer 2015
16 / 26
Background NLU PoliCon
Metaphor annotation project – manual track
I
Supported by Ramsay Fund (School of CS, Uni. Birmingham)
I
Aim:
* Collect 5000+ total metaphorically used words, i.e. annotation of
around 45000+ words in total (assuming 1 in 8 words are metaphorical)
* Collect texts in 3 topics: Immigration, Climate, Security
* So: approx. 1600+ metaphors in each category
I
2 phases of project, collection (6 annotators) followed by correction (2
annotators)
I
Release date of manually annotated data: early 2016
Gargett
Generating Metaphor
ICLC13 Summer 2015
17 / 26
Background NLU PoliCon
Metaphor annotation project – automatic track
I
With the manual data in hand, we then turn to using the metaphor
detection and interptation tools, to automatically find and tag
metaphors in a lot more text
I
We will do evaluation studies of the automatic track, as follow-up to
the manual annotation project [i.e. with the same annotators]
I
Release date of first batch of automatically annotated data: early
2016
Gargett
Generating Metaphor
ICLC13 Summer 2015
18 / 26
Thanks!
Acknowledgments:
∗ Some work reported here is joint with Josef Ruppenhofer (Uni.
Hildesheim, Germany); Paul Rayson, Zsofia Demjen & Elena Semino
(Uni. Lancaster); Sarah Turner, Susan Hunston & Jeannette
Littlemore (Uni. Birmingham, UK)
∗ We acknowledge financial support through a Marie Curie International
Incoming Fellowship (project 330569) (Gargett as fellow, Barnden as
P.I.)
∗ We also acknowledge the support of the Ramsay Fund, School of
Computer Science, Uni. of Birmingham, which provided funds for
annotating the PoliCon corpus
∗ Finally, thanks to the Institute of Advanced Studies, Uni. of
Birmingham, which for funding a visit by Ruppenhofer to UoB
through their Distinguished Visiting Fellowship Scheme
Background NLU PoliCon
The Natural language generation “pipeline”
Gargett
Generating Metaphor
ICLC13 Summer 2015
20 / 26
Background NLU PoliCon
GenMeta aims
I
Provide a natural-language interface for John Barnden’s system for
processing metaphor, ATT-Meta
I
To develop a system using deep reasoning to model how speaker’s
choose metaphorical views, and how such views gain meaning from
everyday experiences. Eg. viewing ILLNESS:
1. More common: as an ATTACK 7→ “an asthma attack”
2. More novel: as a RIDE 7→ “the diabetes rollercoaster”
I
To express such views using generated metaphorical expressions that
are comprehensible and natural-seeming to an ordinary person
⇒ Special focus on varying formulaic expressions as necessary to capture
the information to be conveyed
I
To cover metaphor within the domains of illness and political conflict
discourse.
Gargett
Generating Metaphor
ICLC13 Summer 2015
21 / 26
Background NLU PoliCon
Strategic vs. tactical generation
I
A useful distinction:
∗ Strategic generation: working out what you want to say
∗ Tactical generation: working out how to say it
⇒ Metaphor generation involves both
∗ Within NLG research, strategic generation relatively under-researched
I
Our primary focus is tactical generation
∗ But we also have a secondary focus on strategic generation
Gargett
Generating Metaphor
ICLC13 Summer 2015
22 / 26
Background NLU PoliCon
Conventionality of metaphorical expressions
I
Having parsed the VUAMC using a dependency parser, we are
pursuing various approaches to model conventionality, including:
1. Detecting patterns in the VUAMC:
I
I
Based on Pattern Grammar [Hunston] (in progress)
Using the dependency labels, infer patterns
2. We consult a frequency database for patterns in the VUAMC, whereby
we can “look up” a potential conventional expression for a particular
metaphorical expression
3. (in progress)
Gargett
Generating Metaphor
ICLC13 Summer 2015
23 / 26
Background NLU PoliCon
Sentiment of metaphorical expressions
I
We have integrated sentiment scores into the VUAMC corpus
extended with dependency info:
⇒ We can use sentence-level info about metaphors, to manipulate degree
& polarity of sentiment in different metaphors
eg for ARGUMENT-AS-ATTACK:
1. Relatively positive NP: “a rebuttal”
2. Rel. negative VP: “failed to make a rebuttal”
3. Rel. pos. complex VP: “never failed to make a rebuttal”
Gargett
Generating Metaphor
ICLC13 Summer 2015
24 / 26
Background NLU PoliCon
Refining our ML studies: specialising classifiers
I
Boost results by constructing models for specific POS & text types
⇒ Training several random forests classifiers, specialised for:
POS
Genre:
Nouns vs. Verbs vs. Prepositions
Academic vs. Conversation vs. News vs. Fiction
⇒ Results (Accuracy, F measure):
Genre
POS
n
Acc
F1
All
N
V
P
1468
2214
2292
.734
.800
.811
.733
.814
.828
Academic
N
V
P
500
536
694
.702
.759
.919
.703
.775
.922
Gargett
Generating Metaphor
ICLC13 Summer 2015
25 / 26
Background NLU PoliCon
Refining our ML studies: specialising classifiers
I
Boost results by constructing models for specific POS & text types
⇒ Training several random forests classifiers, specialised for:
POS
Genre:
Nouns vs. Verbs vs. Prepositions
Academic vs. Conversation vs. News vs. Fiction
⇒ Results (Accuracy, F measure):
Genre
POS
n
Acc
F1
All
N
V
P
1468
2214
2292
.734
.800
.811
.733
.814
.828
Conversation
N
V
P
154
406
314
.734
.803
.761
.748
.815
.783
Gargett
Generating Metaphor
ICLC13 Summer 2015
25 / 26
Background NLU PoliCon
Refining our ML studies: specialising classifiers
I
Boost results by constructing models for specific POS & text types
⇒ Training several random forests classifiers, specialised for:
POS
Genre:
Nouns vs. Verbs vs. Prepositions
Academic vs. Conversation vs. News vs. Fiction
⇒ Results (Accuracy, F measure):
Genre
POS
n
Acc
F1
All
N
V
P
1468
2214
2292
.734
.800
.811
.733
.814
.828
News
N
V
P
528
786
804
.731
.752
.792
.744
.775
.808
Gargett
Generating Metaphor
ICLC13 Summer 2015
25 / 26
Background NLU PoliCon
Refining our ML studies: specialising classifiers
I
Boost results by constructing models for specific POS & text types
⇒ Training several random forests classifiers, specialised for:
POS
Genre:
Nouns vs. Verbs vs. Prepositions
Academic vs. Conversation vs. News vs. Fiction
⇒ Results (Accuracy, F measure):
Genre
POS
n
Acc
F1
All
N
V
P
1468
2214
2292
.734
.800
.811
.733
.814
.828
Fiction
N
V
P
286
486
480
.675
.710
.690
.685
.739
.735
Gargett
Generating Metaphor
ICLC13 Summer 2015
25 / 26
Background NLU PoliCon
Refining our ML studies: specialising classifiers
I
Boost results by constructing models for specific POS & text types
⇒ Training several random forests classifiers, specialised for:
POS
Genre:
Nouns vs. Verbs vs. Prepositions
Academic vs. Conversation vs. News vs. Fiction
⇒ Results (Accuracy, F measure):
Genre
POS
n
Acc
F1
All
N
V
P
1468
2214
2292
.734
.800
.811
.733
.814
.828
ACN
N
V
P
1182
1728
1812
.739
.803
.813
.738
.814
.826
Gargett
Generating Metaphor
ICLC13 Summer 2015
25 / 26
Background NLU PoliCon
Cross-domain performance
⇒ Results (Accuracy, F measure):
Training
on
Testing
on
n
Acc
F1
Academic
Carer
Patient
Client
Interview
88
60
148
234
.693
.633
.595
.564
.752
.703
.648
.648
Gargett
Generating Metaphor
ICLC13 Summer 2015
26 / 26
Background NLU PoliCon
Cross-domain performance
⇒ Results (Accuracy, F measure):
Training
on
Testing
on
n
Acc
F1
News
Carer
Patient
Client
Interview
88
60
148
234
.761
.583
.574
.624
.807
.658
.687
.712
Gargett
Generating Metaphor
ICLC13 Summer 2015
26 / 26
Background NLU PoliCon
Cross-domain performance
⇒ Results (Accuracy, F measure):
Training
on
Conversation
Gargett
Testing
on
n
Acc
F1
Carer
Patient
Client
Interview
88
60
148
234
.682
.683
.615
.573
.759
.759
.719
.701
Generating Metaphor
ICLC13 Summer 2015
26 / 26
Background NLU PoliCon
Cross-domain performance
⇒ Results (Accuracy, F measure):
Training
on
Fiction
Gargett
Testing
on
n
Acc
F1
Carer
Patient
Client
Interview
88
60
148
234
.705
.683
.682
.701
.750
.725
.740
.748
Generating Metaphor
ICLC13 Summer 2015
26 / 26