Empirical and Computational Approaches to Metaphor and Figurative Meaning

One of the hallmarks of human intelligence is the ability to go beyond literal meanings of utterances to infer speakers’ intended meanings, a feat that remains elusive to the most advanced artificial systems. Figurative language such as metaphor, in particular, provides a striking case where complex meanings arise that cannot be derived from the literal semantics alone. For example, My lawyer is a shark is a literally false sentence, yet communicates relevant features of the lawyer in question (e.g. ruthless, but not swims). Metaphor and other types of figurative language raise a range of questions core to the study of language and cognition; as a result, how people derive and produce figurative meanings has been approached from various angles and with a diverse set of methods. Some psychologists focus on the cognitive mechanisms that underly interpretations of specific types of figurative use, such as how people align shared properties and analogous relations across domains in order to understand metaphor (Gentner & Wolff, 1997). Other researchers apply theories of communication to explain how people arrive at contextually appropriate interpretations of non-literal utterances (Steen, 2015; Kao, Wu, Bergen, & Goodman, 2014). Finally, natural language processing researchers seek to identify features and principles that can help artificial agents process and produce figurative language (Veale & Hao, 2007). In this symposium, we will discuss the methods that our speakers have employed to examine how people interpret and produce figurative meaning, as well as ways to combine complementary approaches. By bringing together experts with different theoretical perspectives and from a range of disciplines, we aim to discuss outstanding questions that may require a synthesis of tools to resolve. We will open the symposium with a brief overview of the landscape of metaphor and figurative meaning, provided by N. Goodman. We will then present four 20-minute talks, starting with F. Maravilla presenting joint work with D. Gentner on how conventionality, relationality, and aptness affect figurative language processing. G. Steen will then introduce the Deliberate Metaphor Theory and highlight the importance of communicative intent. Continuing this thread, J. Kao will present joint work with Goodman on applying Bayesian models of communication to interpret figurative uses. Finally, T. Veale will describe a computational system that employs cognitive principles to automatically generate figurative language. The symposium will end with a 20-30 minute discussion with the speakers and audience (facilitated by Goodman and Kao). Conventionality, Relationality and Aptness in Figurative Language Processing Francisco Maravilla & Dedre Gentner (Northwestern University) Figurative statements can be expressed either in comparison syntax, as similes (An X is like a Y), or in categorization syntax, as metaphors (An X is a Y). What determines preference for one form over the other? The Career of Metaphor account (Bowdle & Gentner, 2005; Gentner & Wolff, 1997) links this difference to language evolution. Novel figuratives are processed as comparisons, so they are preferred in simile form. In contrast, conventional figuratives can be expressed as metaphors (categorization form), because they have acquired a standard metaphoric abstraction. A second account (Aisenman, 1999) proposes that people prefer to express relational meanings as metaphors and attributional meanings as similes. A third account, Glucksberg and Keysar (1990) class inclusion theory of metaphor, claims that the preference for metaphor form depends on aptness, not conventionality (Glucksberg, 2003). We tested these accounts in three studies. In Experiment 1, conventionality and relationality significantly predicted preference for metaphor. In Experiment 2, we found that people provided relational interpretations for both forms. In Experiment 3 (underway) we test the claim that apt figuratives are preferred as metaphors, and ask how aptness relates to relationality. Our findings so far support the Career of Metaphor theory and Aisenman?s relationality hypothesis.