Metaphor Generation with Conceptual Mappings

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6724–6736 August 1–6, 2021. ©2021 Association for Computational Linguistics 6724 Metaphor Generation with Conceptual Mappings Kevin Stowe1, Tuhin Chakrabarty2, Nanyun Peng3 Smaranda Muresan2, Iryna Gurevych1 1Ubiquitous Knowledge Processing Lab, Technical University of Darmstadt https://www.informatik.tu-darmstadt.de/ukp/ 2Columbia University, {tuhin.chakr,smara}@cs.columbia.edu 3University of California Los Angeles, violetpeng@cs.ucla.edu Abstract Generating metaphors is a difficult task as it requires understanding nuanced relationships between abstract concepts. In this paper, we aim to generate a metaphoric sentence given a literal expression by replacing rele- vant verbs. Guided by conceptual metaphor theory, we propose to control the generation process by encoding conceptual mappings be- tween cognitive domains to generate meaning- ful metaphoric expressions. To achieve this, we develop two methods: 1) using FrameNet- based embeddings to learn mappings between domains and applying them at the lexical level (CM-Lex), and 2) deriving source/target pairs to train a controlled seq-to-seq generation model (CM-BART). We assess our methods through automatic and human evaluation for basic metaphoricity and conceptual metaphor presence. We show that the unsupervised CM- Lex model is competitive with recent deep learning metaphor generation systems, and CM-BART outperforms all other models both in automatic and human evaluations.1 1 Introduction Recent neural models have led to important progress in natural language generation (NLG) tasks. While pre-trained models have facilitated advances in many areas of generation, the field of metaphor generation remains relatively unexplored. Moreover, the few existing deep learning models for metaphor generation (Yu and Wan, 2019; Stowe et al., 2020; Chakrabarty et al., 2020) lack any con- ceptualization of the meaning of the metaphors. This work proposes the first step towards metaphor generation informed by the conceptual metaphor theory (CMT) (Lakoff and Johnson, 1980; Lakoff, 1993; Reddy, 1979). CMT holds 1All code, models, and data are made avail- able at: https://github.com/UKPLab/ acl2021-metaphor-generation-conceptual Figure 1: Metaphor generation guided by concep- tual metaphors. Given a literal input, we can gener- ate metaphoric outputs based on different mappings be- tween conceptual domains. that we use conceptual mappings between domains (conceptual structures that group related concepts) to generate linguistic metaphors.2 Metaphoric map- pings consist of a source and a target conceptual do- main. The source domain is the conceptual domain from which we draw the metaphorical expressions, while the target domain is the conceptual domain that we try to understand. A classical mapping is ARGUMENT IS WAR, in which we conceptual- ize the target argumentation domain as the more concrete source domain of war: • They fought against the contract. • They defended their new proposal. We focus on verbs, as they are often the key component of metaphoric expressions (Steen et al., 2010; Martin, 2006). When used metaphorically, verbs typically evoke source domains (e.g. fought, defended in the above examples): they are con- crete, and are used to understand more abstract tar- gets (i.e., argumentation verbs such as argued, sup- ported) via conceptual mappings (Sullivan, 2013). We propose a novel …

Content truncated to 500 words