Night Owls and Majestic Whales: Modeling Metaphor Comprehension as a Rational Speech Act over Vector Representations of Lexical Semantics

Anonymous

Abstract: While they are some of the few computational models that directly capture pragmatic processes underlying language reasoning, current Rational Speech Act (RSA) models of metaphor are (1) not easily scalable, and (2) do not align well with contemporary accounts of metaphor comprehension. The following research project leverages GloVe word vectors to capture pragmatic language reasoning in metaphoric utterances using an updated RSA framework. This updated framework better aligns model predictions with Relevance Theoretic and Construction Grammatical theories of metaphor semantics. The model yields high posterior probabilities for attributes of metaphors that humans deem relevant in metaphoric utterances over erroneous ones in 89% of all cases, validating the methodology to generate prior probabilities for a RSA framework. When presented with biased priors like listeners are in many naturalistic conversations, the model accurately matches human judgements of the most topical attribute of a topic/target indicated by a metaphoric utterance 90% of the time.

Paper Type: long