Models of metaphor in NLP

Citation

E. Shutova. 2010. Models of Metaphor in NLP. In Proceedings of ACL 2010, Uppsala, Sweden.

Online version

ACL anthology

Introduction

This is a review paper of modeling metaphors in NLP. The author devised it into two main tasks: "metaphor recognition" and "metaphor interpretation".

Metaphor Recognition

Met* System (Fass, 1991)

• First attempt to identify and interpret metaphorical expression
• Using selectional preference and hand-coded knowledge base
• 3-Stage Approaches

1. Detect selectional preference violation
2. If find violations, tested for being a metonymic relation using hand-coded patterns
3. If not metonymy, search the knowledge base for a relevant analogy in order to discriminate metaphorical relations

• Problem

1. Detects any kind of non-literalness in language (metaphors, metonymies and others), and not only metaphors
2. Fail to detect high conventionality of metaphor

Goatly (1997)

• Identify a set of linguistic cues indicate metaphor
  • metaphorically speaking, utterly, completely, so to speak and, surprisingly, literally.

Peters & Peters (2000)

• Detect figurative language in the WordNet
• Search for systematic polysemy, which allows to capture metonymic and metaphorical relations

CorMet System (Mason, 2004)

• The first attempt for source-target domain mapping
• A corpus-based approach to find systematic variations in domain-specific selectional preferences
• Take Master Metaphor List (Lakoff et al., 1991) as baseline, and achieve an accuracy of 77% (judged by human)

TroFi System (Birke & Sarkar, 2006)

• Sentence clustering approach for non-literal language recognition
• Inspired by a similarity-based word sense disambiguation method
• Approach

1. Use a set of seed sentences (human annotated)
2. Compute similarity between (1) the sentence containing the word to be disambiguated and (2) all of the seed sentences
3. Select the sense corresponding to the annotation in the most similar seed sentences

• F1-score = 0.538. But the task is not clearly defined.

Gedigan et al. (2006)

• A survey-like work about literal and metaphorical uses of verbs
• Method

1. Extract the lexical items whose frames are related to MOTION and CURE from FrameNet (Fillmore et al., 2003).
2. Search the PropBank Wall Street Journal corpus (Kingsbury and Palmer, 2002) for sentences containing such lexical items and annotated them with respect to metaphoricity (by hand).
3. Used PropBank annotation as features to train the maximum entropy classifier

• Accuracy = 95.12%
  • Very close to majority baseline (92.90%) => 92.00% of the verbs of MOTION and CURE in the Wall Street Journal corpus are used metaphorically.

Krishnakumaran & Zhu (2007)

• Deal with not only verbs, but also nouns and adjectives.
• Use hyponymy relation in WordNet and word bigram counts to predict metaphors at sentence level.
• Approach
  • Noun-Noun: Given an IS-A metaphor, verify if the two nouns involved are in hyponymy relation in WordNet, if not, tag as a metaphor
  • Verb-Noun and Adjective-Noun: Calculate bigram probabilities of verb-noun and adj-noun pairs, if the pair has low frequency, tag as a metaphor
• Fail to capture conventionalized metaphors (loose information compared of parsed text)

Metaphor Interpretation

MIDAS System (Martin, 1990)

• A Metaphor Interpretation, Denotation and Acquisition System (MIDAS)
• Capture hierarchical organization of conventional metaphors
• Idea
  • More specific conventional metaphors descend from the general ones
• Approach (modeling how novel metaphors are acquired)

1. Given a metaphor example, search for a corresponding metaphor that would explain the anomaly.
2. If not find: abstract it from the example to more general concepts and repeats search.
3. If find a suitable general metaphor: create a mapping for its descendant (a more specific metaphor) based on this example.

• MIDAS has been integrated with the Unix Consultant (UC), the system that answers questions about Unix.

KARMA System (Narayanan, 1997), ATT-Meta (Barnden and Lee, 2002)

• Performing inferences about entities and events in the source and target domains for metaphor interpretation.
• Metaphor-based reasoning framework
• Approach

1. Reasoning process relies on manually coded knowledge of the source domain.
2. The results are then projected onto the target domain (using the conceptual mapping representation).

• Problems

1. ATT-Meta concerns metaphorical and metonymic description of mental states and reasoning about mental states using first order logic. But they can not take natural language sentences as input.
2. KARMA doesn't have this problem, it takes parsed text as input.

Talking Points (Veale and Hao, 2008)

• Fluid knowledge representation for metaphor interpretation and generation
• Approach

1. From WordNet and from the web, extract a set of characteristics of (1) concepts belonging to source and target domains and (2) related facts about the world.
2. Talking Points are then organized in Slipnet, a framework that allows for a number of insertions, deletions and substitutions in definitions of such characteristics in order to establish a connection between the target and the source concepts.

• Problem
  Not evaluate to which extent their knowledge base are useful.

Shutova (2010)

• Defines metaphor interpretation as a paraphrasing task
• Derive literal paraphrases for metaphorical expression
• Approach

1. Given a context, apply a probabilistic model to rank all possible paraphrases for the metaphorical expression
2. Use automatically induced selectional preferences to discriminate between figurative and literal paraphrases

• Tested on metaphors expressed by a verb, and achieve an accuracy of 0.81.

Metaphor Resources

• Metaphor is a knowledge-hungry phenomenon.
• Two type of resources are needed

1. Extensive manually-created knowledge-base (easier, most work is of this type)
2. Robust knowledge acquisition system for interpretation of metaphorical expressions (much harder)

• Although hand-coded knowledge is useful for metaphor interpretation, it has a very limited coverage.

Master Metaphor List (Lakoff et al., 1991)

• Includes a classification of metaphorical mappings (mainly those related to mind, feelings and emotions) with the corresponding examples of language use.
• Problem: lack of clear structuring principles of the mapping ontology

1. The taxonomical levels are often confused
2. The same classes are referred to by different class labels
3. (Consquence) Not suitable for computational use.

• The idea and its actual mappings ontology inspired a lot of other work

MetaBank (Martin, 1994) & Mental Metaphor Databank

• Created in the framework of the ATT-meta project
• A knowledge-base of English metaphorical conventions
• Represented in the form of metaphor maps (Martin, 1988) which contains detailed information about source-target concept mappings

Other Language

• There has been a wealth of research on metaphor in Spanish, Chinese, Russian, German, French and Italian.
• German and French: Hamburg Metaphor Database (L¨onneker, 2004; Reining and L¨onneker-Rodman,2007)
  • Contain examples of metaphorical expressions which are mapped to senses from EuroWordNet and annotated with source–target domain mappings (taken from the Master Metaphor List.)
• Italian: Alonge and Castelli (2003)
  • Discuss how metaphors can be represented in ItalWordNet

Metaphor Annotation in Corpora

• Two distinct aspects of the phenomenon, two stages of metaphor annotation

1. Identify metaphorical senses in text (akin word sense disambiguation)
2. Annotating source–target domain mappings of it

• Traditional approaches to metaphor annotation is manually search for:

1. Lexical items used metaphorically
2. Source and target domain vocabulary
3. Linguistic markers of metaphor

Metaphor & Polysemy

• Two kinds of metaphorical language

1. Novel (or poetic) metaphors (surprise our imagination)
2. Conventionalized metaphors (become a part of an ordinary discourse)

• Over time some of the aspects of the target domain are added to the meaning of a term in a source domain, resulting in a (metaphorical) sense extension of this term.
• It is often unclear that if a metaphorical instance is:

1. A case of broadening of the sense in context due to general vagueness, or
2. It manifests a formation of a new distinct metaphorical sense

Metaphor Identification

Pragglejaz Procedure (2007)

• A metaphor identification procedure (MIP)
• Metaphor annotation at the word level
• Annotators guidelines

1. For each verb establish its meaning in context and try to imagine a more basic meaning of this verb on other contexts. Basic meanings normally are: (1) more concrete; (2) related to bodily action; (3) more precise (as opposed to vague); (4) historically older.
2. If you can establish the basic meaning that is distinct from the meaning of the verb in this context, the verb is likely to be used metaphorically.

Source - Target Domain Vocabulary

• Lists of source and target domain vocabulary
• Support the metaphor extraction technique that search for sentences containing lexical items from the source domain, the target domain, or both.
• Martin (2006)
  • A corpus study in order to confirm that metaphorical expressions occur in text in contexts containing such lexical items.
  • Goal
  1. Evaluate predictive ability of contexts containing vocabulary from source domain and target domain
  2. Estimate the likelihood of a metaphorical expression following another metaphorical expression described by the same mapping
  • Data: Wall Street Journal (WSJ) corpus
  • Focus on four conceptual metaphors
  1. NUMERICAL VALUE AS LOCATION
  2. COMMERCIAL ACTIVITY AS CONTAINER
  3. COMMERCIAL ACTIVITY AS PATH FOLLOWING
  4. COMMERCIAL ACTIVITY AS WAR
  • Method
  1. Manually compile the lists of terms characteristic for each domain by examining sampled metaphors of these types
  2. Augment them through the use of thesaurus
  3. Search the WSJ for sentences containing vocabulary from these lists
  4. Checked whether the sentences contain metaphors of the above types
  • Result
    NUMERICAL-VALUE-AS-LOCATION type has the best result, P(Metaphor|Source) = 0.069, P(Metaphor|Target) = 0.677, P(Metaphor|Metaphor) = 0.703

Annotating Source and target Domains

• Wallington et al. (2003)
  • Two teams of annotators:
  1. Team A
     Asked to annotate “interesting stretches”, whereby a phrase was considered interesting if
     1. its significance in the document was non-physical
     2. it could have a physical significance in another context with a similar syntactic frame
     3. this physical significance was related to the abstract one.
  2. Team B
     Annotate phrases according to their own intuitive definition of metaphor
  • Also annotate the source – target domain mappings
    Annotators were given a set of mappings from the Master Metaphor List, and were asked to assign the most suitable ones.
  • Problem
  1. Not report the level of inter-annotator agreement
  2. Not report the coverage of the mappings in the Master Metaphor List
• Shutova and Teufel (2010)
  • Different approach to the annotation of source – target domain mappings
    • Not rely on predefined mappings
    • Derive independent sets of most common source and target categories
  • 2-stage Procedure
  1. Metaphors are first identified using MIP
  2. The source domain and the target domain are selected from the lists of categories
  • Inter-annotator agreement (Kappa) = 0.61

Conclusion

The eighties and nineties provided us with a wealth of ideas on the structure and mechanisms of the phenomenon of metaphor. The approaches formulated back then are still highly influential, although their use of hand-coded knowledge is becoming increasingly less convincing. The last decade witnessed a high technological leap in natural language computation, whereby manually crafted rules gradually give way to more robust corpus-based statistical methods. This is also the case for metaphor research. The latest developments in the lexical acquisition technology will in the near future enable fully automated corpusbased processing of metaphor. However, there is still a clear need in a unified metaphor annotation procedure and creation of a large publicly available metaphor corpus. Given such a resource the computational work on metaphor is likely to proceed along the following lines: (1) automatic acquisition of an extensive set of valid metaphorical associations from linguistic data via statistical pattern matching; (2) using the knowledge of these associations for metaphor recognition in the unseen unrestricted text and, finally, (3) interpretation of the identified metaphorical expressions by deriving the closest literal paraphrase (a representation that can be directly embedded in other NLP applications to enhance their performance). Besides making our thoughts more vivid and filling our communication with richer imagery, metaphors also play an important structural role in our cognition. Thus, one of the long term goals of metaphor research in NLP and AI would be to build a computational intelligence model accounting for the way metaphors organize our conceptual system, in terms of which we think and act.

Related papers

The widely cited Pang et al EMNLP 2002 paper was influenced by this paper - but considers supervised learning techniques. The choice of movie reviews as the domain was suggested by the (relatively) poor performance of Turney's method on movies.

An interesting follow-up paper is Turney and Littman, TOIS 2003 which focuses on evaluation of the technique of using PMI for predicting the semantic orientation of words.