What Drives the Use of Metaphorical Language? Negative Insights from Abstractness, Affect, Discourse Coherence and Contextualized Word Representations

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 299 - 310 July 14-15, 2022 ©2022 Association for Computational Linguistics What Drives the Use of Metaphorical Language? Negative Insights from Abstractness, Affect, Discourse Coherence and Contextualized Word Representations Prisca Piccirilli and Sabine Schulte im Walde Institute for Natural Language Processing, University of Stuttgart, Germany {prisca.piccirilli,schulte}@ims.uni-stuttgart.de Abstract Given a specific discourse, which discourse properties trigger the use of metaphorical lan- guage, rather than using literal alternatives? For example, what drives people to say grasp the meaning rather than understand the mean- ing within a specific context? Many NLP ap- proaches to metaphorical language rely on cog- nitive and (psycho-)linguistic insights and have successfully defined models of discourse coher- ence, abstractness and affect. In this work, we build five simple models relying on established cognitive and linguistic properties – frequency, abstractness, affect, discourse coherence and contextualized word representations – to pre- dict the use of a metaphorical vs. synonymous literal expression in context. By comparing the models’ outputs to human judgments, our study indicates that our selected properties are not suf- ficient to systematically explain metaphorical vs. literal language choices. 1 Introduction Metaphors are "not just nice", but represent a "nec- essary" element of everyday thought and commu- nication (Ortony, 1975; Lakoff and Johnson, 1980; van den Broek, 1981; Schäffner, 2004, i.a.), and are ubiquitous in natural language text corpora (Gedi- gian et al., 2006; Shutova and Teufel, 2010; Steen et al., 2010, i.a.). From the perspective of natural language processing (NLP), automatic approaches to metaphor processing are therefore important for any task that requires natural language understand- ing, and NLP has been concerned with the detection (Köper and Schulte im Walde, 2016; Alnafesah et al., 2020; Ehren et al., 2020; Dankers et al., 2020, i.a.), the interpretation (Shutova, 2010; Bizzoni and Lappin, 2018; Mao et al., 2018, i.a.) and, the gen- eration (Stowe et al., 2021; Zhou et al., 2021) of metaphors.1 1See Tong et al. (2021) for a systematic, comprehensive review and discussion of the most recent metaphor processing systems and datasets. As to our knowledge, however, no study so far has raised the question of WHY a metaphorical ex- pression is used within a specific discourse, rather than an equally plausible literal alternative. For ex- ample, consider the discourse in table 1, where both the metaphorical expression grasp the meaning and its synonymous literal alternative understand the meaning seem equally acceptable (Piccirilli and Schulte im Walde, 2021, 2022). What are the fac- tors priming for one use or the other? Are there cues within the discourse which influence the se- lection of one usage over the other? To which extent can computational approaches based on dis- course properties model human behavior regarding the choice between synonymous metaphorical vs. literal language usage? According to psycholinguistics and computa- tional linguistics research, the processing of words is a function of their frequency of occurrence in the language (van Jaarsveld and Rattink, 1988; Wittmann et al., 2017, i.a.). Is the …

Content truncated to 500 words