Computational Modeling Of Metaphor In Discourse

List of Figures2.12.2Procedure and explication of the MIP (Pragglejaz-Group, 2007) . . . . . . . . . 13Predicates from LAB that select for liquids are transferred to FINANCE andselect for money. On the other hand, predicates from FINANCE that select formoney are transferred to LAB and do not select for liquids (Mason, 2004) . . . . 183.1Correspondence between MTurkers and trained annotators. X-axis: the numberof MTuckers annotating a word as metaphor. Y-axis: the relative percentage ofeach type. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304.14.2Sample excerpt from interviews (Gratch et al., 2014). . . . . . . . . . . . . . . . 35Distribution of metaphor use; The x-axis represents the number of metaphors aperson used, and the y-axis represents the number of people. . . . . . . . . . . . 395.1Distribution of journey metaphor centered around diagnosis event (x-axis: monthsfrom event, y-axis: average frequency of metaphor usage) . . . . . . . . . . . . . 49Distribution of warrior metaphor centered around diagnosis event (x-axis: monthsfrom event, y-axis: average frequency of metaphor usage) . . . . . . . . . . . . . 505.26.1Graph representation depicting lexical cohesion among words in a given text.Edges represent lexical relatedness between a topic and a word or between words.For example, w1 is directly related to the topic of discussion, whereas w7 is onlyindirectly related to the topic through w2 . . . . . . . . . . . . . . . . . . . . . . 597.1Proportions of topics assigned to target sentences, when target words were usedmetaphorically vs. literally. The proportions of metaphorical and literal casesare different with statistical significance of p 0.01 by Pearson’s chi-square test. 72Proportions of the topics of the sentences that are nearest to the target sentenceand have a different topic from the target sentence. The proportions of metaphorical and literal cases are different with statistical significance of p 0.01 byPearson’s chi-square test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73Proportions of target sentences whose topic is different from that of the previous/next sentence, when target words were used metaphorically vs. literally.The proportions of metaphorical and literal cases are different with statisticalsignificance of p 0.01 by Pearson’s chi-square test. . . . . . . . . . . . . . . . 747.27.3xii

7.47.58.1Cosine similarity between the topic of a target sentence and the topic of its previous/next sentence, when target words were used metaphorically vs. literally.The means of the metaphorical and literal cases are different with statistical significance of p 0.01 by Welch’s t-test. . . . . . . . . . . . . . . . . . . . . . . 74Cosine similarity of the topic of a target sentence and the topic of the sentencesthat are nearest to the target sentence and have a different topic from the targetsentence. The means of metaphorical and literal cases are different with statistical significance only for the next sentence, with p 0.01 by Welch’s t-test. . . . 75System flow diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79xiii

List of Tables1.1Structure of this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2.12.2Overview of Section 2.3.1 Metaphor Detection . . . . . . . . . . . . . . . . . . 15A subspace of the property-norm semantic space. Attribute-based vectors werebuilt from this resource (Bulat et al., 2017) . . . . . . . . . . . . . . . . . . . . . 193.13.23.3Inter-reliability between two trained annotators for our annotation scheme. . .Questions to annotate (N: Nonliteral, C: Conventionalized, L: Literal). . . . . .Inter-reliability between two trained annotators for our annotation scheme (N:Nonliteral, C: Conventionalized). . . . . . . . . . . . . . . . . . . . . . . . .Inter-reliability among MTurkers. . . . . . . . . . . . . . . . . . . . . . . . .Data statistics (N: Nonliteral, C: Conventionalized). . . . . . . . . . . . . . .Inter-reliability between trained annotators and MTurkers (N: Nonliteral, C: Conventionalized). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24. 27Example of interview questions asked to one participant. . . . . . . . . . . . .Example of questions that are excluded for annotation. . . . . . . . . . . . . .Example of questions selected for annotation. . . . . . . . . . . . . . . . . . .Nominal logistic regression model of “intimacy” and affective “polarity” of questions on “target sentiment”. The odds ratio is defined as P(me) / P(positive) (topthree), P(negative) / P(positive) (middle three), and P(neutral) / P(positive) (bottom three). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Nominal logistic regression model of PTSD on “target sentiment” controlled on“intimacy” and “polarity”. The odds ratio is defined as P(me) / P(neutral) (topthree), P(negative) / P(neutral) (middle three), and P(positive) / P(neutral) (bottom three). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Nominal logistic regression model of depression on “target sentiment” controlledon “intimacy” and “polarity”. The odds ratio is defined as P(me) / P(neutral)(top three), P(negative) / P(neutral) (middle three), and P(positive) / P(neutral)(bottom three). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Logistic regression model of “source sentiment” on anxiety when “target” is self.The odds ratio is defined as P(anxiety 1) / 1-P(anxiety 1). Source sentiment[neutral] is a baseline dummy variable. . . . . . . . . . . . . . . . . . . . . . . 36. 37. 383.43.53.64.14.24.34.44.54.64.7xiv9. 29. 29. 29. 30. 42. 42. 43. 43

4.84.95.15.25.35.45.56.16.26.37.1Nominal logistic regression model of PTSD on “source sentiment negation”when “target” is others. The odds ratio is defined as P(negative) / P(positive)(top three) and P(neutral) / P(positive) (bottom three). Target sentiment [positive] is a baseline dummy variable. . . . . . . . . . . . . . . . . . . . . . . . . . 44Nominal logistic regression model of Anxiety on “source sentiment negation”when “target” is others. The odds ratio is defined as P(negative) / P(positive) (topthree) and P(neutral) / P(positive) (bottom three). Target sentiment [positive] isa baseline dummy variable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44Corpus-wide unambiguous popular metaphor use statistics (among posts wherethe user used the metaphor at least once) (M: posts that contain each metaphor,L: posts that do not contain each metaphor). . . . . . . . . . . . . . . . . . .Metaphor candidates and their associated events . . . . . . . . . . . . . . . .Metaphor use statistics of data used for MTurk (* indicates metaphor candidatesfor which the literal usage is more common than the non-literal one, N: nonliteraluse L: literal use). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Metaphor candidates and their associated events . . . . . . . . . . . . . . . .Performance on metaphor disambiguation evaluation. (6) is significantly betterthan (5) [p 0.013] (fs.: used feature selection) . . . . . . . . . . . . . . . . . . 50. 51. 53. 53. 53Metaphor use statistics of data used for MTurk (* indicates metaphor candidatesfor which the literal usage is more common than the non-literal one, N: nonliteraluse L: literal use). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Performance on metaphor disambiguation evaluation. (Models) T: Tsvetkovet al. (2014), K: Beigman Klebanov et al. (2014), U: context unigram, GWC:global word category, GT: global topic dist., LC: lexical chain, LWC: local wordcategory, SR: semantic relatedness, AC: abstractness/concreteness. (Metrics) A:accuracy, P-M: precision on metaphors, R-M: recall on metaphors, P-L: precision on literal words, R-L: recall on literal words, F1: Average F1 score overM/L., *: statistically significant improvement over baselines . . . . . . . . . . . 65Performance on metaphor disambiguation task per target word with the best setting ALL-LC. Note that the performance results on target words candle and spiceare not reported because of their small number of instances. . . . . . . . . . . . 657.2Performance on metaphor identification task. (Models) C: Frame Contrast modelfrom Chapter 6, (Metrics) : Cohen’s kappa, F1: average F1 score on M/L, PL: precision on literals, R-L: recall on literals, P-M: precision on metaphors,R-M: recall on metaphors, A: accuracy, *: statistically significant (p 0.05)improvement over corresponding baseline by Student’s t-test. . . . . . . . . . . . 70Topics learned by Sentence LDA. . . . . . . . . . . . . . . . . . . . . . . . . . 718.18.28.3The bootstrapping process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80Dependencies from parsed result . . . . . . . . . . . . . . . . . . . . . . . . . . 81Examples of lexico-grammar patterns. r represents a reverse dependency. . . . . 81xv

8.48.58.6Performance on metaphor detection. (Metrics) : Cohen’s kappa, F1: averageF1 score on M/L, P-L: precision on literals, R-L: recall on literals, P-M: precisionon metaphors, R-M: recall on metaphors, A: accuracy, *: statistically significant(p 0.05) improvement over corresponding baseline by Student’s t-test. . . . . . 84Performance on metaphor detection for battle-related metaphors in the politicaldomain. (Metrics) : Cohen’s kappa, F1: average F1 score on M/L, P-L: precision on literals, R-L: recall on literals, P-M: precision on metaphors, R-M:recall on metaphors, A: accuracy, *: statistically significant (p 0.05) improvement over the above one by Student’s t-test, U: unigram model, C: frame contextmodel, T: frame transition model, F: frame facet model. . . . . . . . . . . . . . . 86Performance on metaphor detection for illness-related metaphors in the politicaldomain. (Metrics) : Cohen’s kappa, F1: average F1 score on M/L, P-L: precision on literals, R-L: recall on literals, P-M: precision on metaphors, R-M: recallon metaphors, A: accuracy, U: unigram model, C: frame context model, T: frametransition model, F: frame facet model. . . . . . . . . . . . . . . . . . . . . . . . 86A.1 Target nonliteral language use (a word or phrase) corresponding to our scope ofnonliteral language [A.1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96xvi

Chapter 1Setting the Stage“The metaphor is probably the most fertile power possessed by man.”– Jose Ortega y GassetIt is not difficult to imagine how inefficient and cumbersome it would be if we used only literal language to express our meaning. Almost every concept we refer to is richly multi-faceted,and when we communicate, we tend to focus on some facets more than others. Fortunately, human language provides many linguistic tools for abstraction and deictic reference, which allowsus to package meaning in ways that draw upon shared knowledge and suggest additional nuancesthat we do not fully articulate. These linguistic tools allow us to communicate many aspects efficiently without actually spelling them all out. This abstraction and deictic reference is supportedby the human ability of symbolic reasoning, understanding something (A) in terms of somethingelse (B).Metaphor is a form of abstraction and manifestation of symbolic reasoning in language. Asa type of figurative (nonliteral) language, metaphor highlights similarities between two unlikethings for rhetorical effect. For example, in the sentence Time is money (Lakoff and Johnson,1980), we compare time to money, which is a well-known limited resource, to represent that timeis valuable. In other words, the metaphor is seeing time (A) in terms of money (B). Seeing A interms of B suggests a wide variety of inferences, brought by our shared knowledge of B. Thisshared knowledge includes not only understanding the concept B in isolation, but also in its relationships to surrounding concepts. Put differently, shared knowledge includes the informationabout the connections B has and can evoke connections between parts of A and other related entities. For example, if we see pursuing a Ph.D degree in terms of a journey, we can easily makeconnections between pursuing a Ph.D degree and journey-related activities such as arriving at adestination, encountering obstacles, and falling in a road. Another example shows He is a snakesuggests not only properties of a snake, but also how people feel about or judge snakes in theculture. In this way, these inferences enable a speaker to find shorter and more effective ways ofsaying A by bringing in the information of B the listener already knows. Thus, metaphor allowspeople to deliver thoughts, feelings, and ideas effectively that might otherwise be difficult to sayonly by using literal language.At the same time, casting A in terms of B may convey the social implications of B. If Brepresents some elite knowledge, the metaphor can be a way of bestowing respect or signal1

ing group membership. Similarly, B representing some personal knowledge can be a way ofshowing intimacy or solidarity. For example, when Computer Science Ph.D students talk aboutthe romantic dating process in terms of a machine learning algorithm, they can feel a sense ofbelonging to a particular group. This kind of social effect occurs at the same time that propositional meaning is communicated. Thus, when considering metaphor, we need to address thetwo dimensions of meaning that are communicated within a discussion: first, the propositionalmeaning of what a speaker is saying, and second, the social meaning of who the speakers are,their personal identities, their relationships with one another, and the communities in which theinteraction is situated.Metaphor is a unique language tool in that a metaphor invites participants to contribute tothe meaning of a conversation in new ways, because it provides facets and perspectives of a newdomain. The new domain affects how other participants might respond and become involved inthe communication. For example, EX(1)–EX(4) from the same thread in the breast cancer discussion forum shows how conversational participants repeat and expand one another’s metaphorsby linking facets of wagon such as falling, part of a journey, and on weagon.EX(1) “falling off the wagon is no big thing in my opinion, the psychological goodfeelings of enjoyment weigh in big for feeling good.”EX(2) “**** falling off is part of this journey, it is stupid to deny yourself everything.”EX(3) “I am on the wagon so far today .ongoing battle.”EX(4) “**** - hope you stay on the wagon, or at least get back on after you fall!”Metaphor has fascinated scholars in a wide variety of fields including philosophy, cognitivescience, sociology, and computational linguistics. Broadly speaking, conceptions of metaphor inall of these fields embrace some version of the ideas communicated above. However, each oneapproaches metaphor from its own distinct methodological perspective and emphasizes someparts of the picture over others, depending on the shared values and goals of that scholarly community.Philosophers, for example, seek to understand how people work in terms of discrete logical formalizations that can be manipulated using prescriptive rules of inference. This placesphilosophers at the high abstraction end of the continuum. They posit formalizations that matchtheir subjective experience and reason about whether the implications are similarly consistentwith their experience. They seek examples as illustrations, but are not strongly empirical in theirapproach. For instance, within language pragmatics, the Gricean model of the cooperative principle (Grice, 1975) and the Relevance Theory model (Sperber et al., 1986) offer formalizationsconsistent with our subjective experience about language, based on a philosophical approach.These models attempt to describe how metaphor works as a linguistic phenomenon. The Griceanmodel explains that flouting Grice’s maxims invites a metaphorical interpretation. The RelevanceTheory model looks at metaphors as general examples of loose talk. While both models addresssome important questions in metaphorical language use, they do not offer direction in terms ofcomputation or empirical work since such approaches to metaphor fall outside the purview oftheir approach.2

In contrast, cognitive science, while still placing a value on general principles that can beused to make causal claims, places a greater value on empiricism and has a greater aversionto over-generalization. Though a propositional understanding of metaphor can reach into thesocial sphere, cogni

Metaphor is used as a language resource/tool to better represent one's point in communication. It can help achieving social goals such as illustrating attitudes in-directly. This thesis aims to understand metaphor from this social perspective in order to capture how metaphor is used in a discourse and identify a broad spectrum