Linguistic Analysis Improves Neural Metaphor Detection

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 362–371 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 362 Linguistic Analysis Improves Neural Metaphor Detection Kevin Stowe, Sarah Moeller, Laura Michaelis, Martha Palmer University of Colorado, Boulder, CO 80309 [kest1439, samo9533, laura.michaelis, mpalmer]@colorado.edu Abstract In the field of metaphor detection, deep learn- ing systems are the ubiquitous and achieve strong performance on many tasks. However, due to the complicated procedures for manu- ally identifying metaphors, the datasets avail- able are relatively small and fraught with com- plications. We show that using syntactic fea- tures and lexical resources can automatically provide additional high-quality training data for metaphoric language, and this data can cover gaps and inconsistencies in metaphor an- notation, improving state-of-the-art word-level metaphor identification. This novel applica- tion of automatically improving training data improves classification across numerous tasks, and reconfirms the necessity of high-quality data for deep learning frameworks. 1 Introduction Humans use metaphors to conceptualize abstract and often difficult concepts by employing knowl- edge of more concrete domains. They are preva- lent in speech and text, and allow us to communi- cate more effectively and more imaginatively. The fact that they are commonplace and easily under- stood by humans makes appropriate interpretation of them essential for high quality natural language processing applications. The primary linguistic and cognitive theory of metaphor is conceptual metaphor theory (Lakoff and Johnson, 1980; Lakoff, 1993), which theo- rizes that metaphors are primarily a mental activ- ity, and the language is merely a side effect of these ”conceptual” metaphors. From this, it is posited that metaphors are agnostic with regard to syntactic structure: a conceptual mapping can be expressed through whatever syntax the speaker de- sires. This is apparent from evidence that many metaphoric predications have the same syntactic properties as their literal counterparts. Goldberg (1995) observes that metaphorical ditransitive sen- tences like ”It gave me a headache” do not dif- fer syntactically from literal ditransitive sentences like ”She gave me the account.” Accordingly, to find syntactic hallmarks of metaphorical meaning we do not look generally for particular syntactic constructions, but rather for mismatches of vari- ous kinds between specific verbs’ ordinary syntac- tic behavior and their behavior under metaphoric interpretation. Perhaps the primary source of verbal syntactic variability is the set of argument-structure con- structions identified by Goldberg (1995). One such construction is Caused Motion (CM), illus- trated by the sentence ”They pushed it down the hall”. CM can augment the array of semantic roles supplied by the verb, as in ”They laughed me out of the room”. Augmentation often entails a metaphoric construal: here the verb ”laugh”, otherwise a single-argument verb, is paired with both a theme argument (the direct object) and a PP location argument, and the resulting predica- tion expresses metaphorical rather than literal mo- tion (Hwang, 2014). Despite this connection between verbal syntax and metaphoric properties, most computational approaches to metaphor eschew syntax for more semantic features. While these have proven effec- tive, metaphor detection …

Content truncated to 500 words