Computational Exploration of Metaphor Comprehension Processes

Computational Exploration of Metaphor Comprehension Processes Akira Utsumi (utsumi@se.uec.ac.jp) Department of Systems Engineering, The University of Electro-Communications 1-5-1 Chofugaoka, Chofushi, Tokyo 182-8585, Japan Abstract A central question in metaphor research is what processes are involved in metaphor comprehension, especially which of comparison and categorization processes governs metaphor comprehension. In this paper, I attempt to provide a compu- tational solution to this problem using comparison and cate- gorization algorithms based on word vectors in a multidimen- sional semantic space constructed by latent semantic analysis. These algorithms receive word vectors for the topic and the ve- hicle of a metaphor and compute a vector for the metaphorical meaning. The resulting vectors can be evaluated on the degree to which they mimic human interpretation of the same meta- phor. Using this simulation framework, I tested five compet- ing views of metaphor comprehension: the categorization view (Glucksberg, 2001), the comparison view (Gentner, 1983) and three hybrid views — the conventionality view (Bowdle & Gentner, 2005), the aptness view (Jones & Estes, 2005) and the interpretive diversity view (Utsumi & Kuwabara, 2005) — which claim that vehicle conventionality, aptness and interpre- tive diversity, respectively, determine a shift between both pro- cesses. The simulation result was that the interpretive diversity view outperformed the other four views on two different mea- sures. This result can be seen as computational evidence in favor of the interpretive diversity view. Keywords: Computational modeling; Latent semantic analy- sis (LSA); Metaphor comprehension Introduction How metaphors are psychologically comprehended is one of the most central topics for metaphor research on which a con- siderable number of studies have been made. Nevertheless, these studies are divided on this issue; some studies (e.g., Gentner, 1983; Gentner, Bowdle, Wolff, & Boronat, 2001) have proposed that metaphors are processed as comparisons or analogical mappings, while others (e.g., Glucksberg, 2001; Glucksberg & Keysar, 1990) have argued that metaphors are processed as categorizations. The comparison view by Gentner and her colleagues (Gen- tner, 1983; Gentner et al., 2001) argues that metaphors (and analogies) are processed as comparisons consisting of a pro- cess of structural alignment between representations of the topic and the vehicle followed by a process of projection of aligned features or relations into the topic. For example, in comprehending the metaphor “A rumor is a virus”, two con- cepts rumor and virus are aligned, salient alignments such as ones about contagion or infection prevention are found, and features and relations inferred from such alignments are pro- jected into the topic. Note that the initial alignment process is symmetric (i.e., the products of the process do not change even if the topic and the vehicle are reversed), while the sub- sequent projection process is asymmetric (i.e., directional). Hence, the intuition that the reversed metaphor “A virus is a rumor” seems meaningless can be attributed to the projection process, not to the alignment process. On the other hand, the categorization view by Glucks- berg and his colleagues (Glucksberg, 2001; Glucksberg & Keysar, 1990) claims that metaphors are seen as categoriza- tion (i.e., class-inclusion) statements expressing that the topic is a member of an abstract superordinate category exemplified by the vehicle. For example, the metaphor “My job is a jail” is comprehended so that the topic my job is categorized as an ad hoc category like “unpleasant and confining things” to which the vehicle jail typically belongs. Note that the topic also plays an important role in metaphor comprehension; it constrains dimensions by which it can be characterized. In the case of the above metaphor, my job facilitates attribution of features related to tasks and jobs, while blocking out irrel- evant features such as ones related to jail building. Very recent studies have tried to reconcile these two op- posite views of metaphor comprehension into a coherent me- taphor theory. However, they disagree on how both views are reconciled, in other words, what property of metaphor de- termines a shift between both processes. Bowdle and Gen- tner (2005) claim that it is vehicle conventionality that deter- mines such shift; their career of metaphor theory argues that, although metaphors are basically processed as comparisons, conventional metaphors are processed as categorizations by accessing stored categories, which are conventionalized by repeated figurative use. Jones and Estes (2005, in press) argue against the career of metaphor view and advocated that meta- phor aptness mediates both processes by empirically demon- strating that apt metaphors were more likely to be processed as categorizations than less apt metaphors. Glucksberg and Haught (in press) also reported that novel but apt metaphors were easy to comprehend in categorization form than in com- parison form, and concluded that the aptness or the quality of metaphors determines the choice of comprehension strategy. Against these views, Utsumi and Kuwabara (2005) and Utsumi (2006) claim that interpretive diversity determines whether metaphors are processed as comparisons or catego- rizations. Interpretive diversity is a measure of the seman- tic richness of literal or figurative utterances including meta- phors; it is high to the extent that more features constitute the utterance meaning and that their relative saliences are more evenly distributed. The interpretive diversity view then ar- gues that diverse metaphors are comprehended by the pro- cess of categorization, but less diverse metaphors with only a few features require the process of comparison because the