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)

Almost simultaneously with the work of Fass (1991), Martin (1990) presents a Metaphor Interpretation, Denotation and Acquisition System (MIDAS). In this work Martin captures hierarchical organisation of conventional metaphors. The idea behind this is that the more specific conventional metaphors descend from the general ones. Given an example of a metaphorical expression, MIDAS searches its database for a corresponding metaphor that would explain the anomaly. If it does not find any, it abstracts from the example to more general concepts and repeats the search. If it finds a suitable general metaphor, it creates a mapping for its descendant, a more specific metaphor, based on this example. This is also how novel metaphors are acquired. MIDAS has been integrated with the Unix Consultant (UC), the system that answers users questions about Unix. The UC first tries to find a literal answer to the question. If it is not able to, it calls MIDAS which detects metaphorical expressions via selectional preference violation and searches its database for a metaphor explaining the anomaly in the question.

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

Another cohort of approaches relies on performing inferences about entities and events in the source and target domains for metaphor interpretation. These include the KARMA system (Narayanan, 1997; Narayanan, 1999; Feldman and Narayanan, 2004) and the ATT-Meta project (Barnden and Lee, 2002; Agerri et al., 2007). Within both systems the authors developed a metaphor-based reasoning framework in accordance with the theory of conceptual metaphor. The reasoning process relies on manually coded knowledge about the world and operates mainly in the source domain. The results are then projected onto the target domain using the conceptual mapping representation. The ATT-Meta project concerns metaphorical and metonymic description of mental states and reasoning about mental states using first order logic. Their system, however, does not take natural language sentences as input, but logical expressions that are representations of small discourse fragments. KARMA in turn deals with a broad range of abstract actions and events and takes parsed text as input.

Talking Points (Veale and Hao, 2008)

• Fluid knowledge representation for metaphor interpretation and generation

Talking Points are a set of characteristics of concepts belonging to source and target domains and related facts about the world which the authors acquire automatically from WordNet and from the web.

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. This work builds on the idea of slippage in knowledge representation for understanding analogies in abstract domains (Hofstadter and Mitchell, 1994; Hofstadter, 1995). Below is an example demonstrating how slippage operates to explain the metaphor Make-up is a Western burqa. Make-up => � typically worn by women � expected to be worn by women � must be worn by women � must be worn by Muslim women Burqa <= By doing insertions and substitutions the system arrives from the definition typically worn by women to that of must be worn by Muslim women, and thus establishes a link between the concepts of make-up and burqa. Veale and Hao (2008), however, did not evaluate to which extent their knowledge base of Talking Points and the associated reasoning framework are useful to interpret metaphorical expressions occurring in text.

Shutova (2010)

Shutova (2010) defines metaphor interpretation as a paraphrasing task and presents a method for deriving literal paraphrases for metaphorical expressions from the BNC. For example, for the metaphors in “All of this stirred an unfathomable excitement in her” or “a carelessly leaked report” their system produces interpretations “All of this provoked an unfathomable excitement in her” and “a carelessly disclosed report” respectively. They first apply a probabilistic model to rank all possible paraphrases for the metaphorical expression given the context; and then use automatically induced selectional preferences to discriminate between figurative and literal paraphrases. The selectional preference distribution is defined in terms of selectional association measure introduced by Resnik (1993) over the noun classes automatically produced by Sun and Korhonen (2009). Shutova (2010) tested their system only on metaphors expressed by a verb and report a paraphrasing accuracy of 0.81.

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.