MERMAID: Metaphor Generation with Symbolism and Discriminative Decoding

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4250–4261 June 6–11, 2021. ©2021 Association for Computational Linguistics 4250 MERMAID: Metaphor Generation with Symbolism and Discriminative Decoding Tuhin Chakrabarty1 Xurui Zhang3∗, Smaranda Muresan1,4 and Nanyun Peng2 1Department of Computer Science, Columbia University 2 Computer Science Department, University of California, Los Angeles, 3Tsinghua University, 4Data Science Institute, Columbia University {tuhin.chakr, smara}@cs.columbia.edu thuzxr@gmail.com, violetpeng@cs.ucla.edu Abstract Generating metaphors is a challenging task as it requires a proper understanding of abstract concepts, making connections between unre- lated concepts, and deviating from the literal meaning. In this paper, we aim to generate a metaphoric sentence given a literal expres- sion by replacing relevant verbs. Based on a theoretically-grounded connection between metaphors and symbols, we propose a method to automatically construct a parallel corpus by transforming a large number of metaphor- ical sentences from the Gutenberg Poetry cor- pus (Jacobs, 2018) to their literal counterpart using recent advances in masked language modeling coupled with commonsense infer- ence. For the generation task, we incorpo- rate a metaphor discriminator to guide the de- coding of a sequence to sequence model fine- tuned on our parallel data to generate high quality metaphors. Human evaluation on an independent test set of literal statements shows that our best model generates metaphors better than three well-crafted baselines 66% of the time on average. Moreover, a task-based eval- uation shows that human-written poems en- hanced with metaphors proposed by our model are preferred 68% of the time compared to po- ems without metaphors. 1 Introduction Czech novelist Milan Kundera in his book “The unbearable lightness of being" said “Metaphors are not to be trifled with. A single metaphor can give birth to love." Metaphors allow us to communicate not just in- formation, but also feelings and complex attitudes (Veale et al., 2016). While most computational work has focused on metaphor detection (Gao et al., 2018; Stowe et al., 2019; Shutova et al., 2010; Tsvetkov et al., 2014; Veale et al., 2016; Stowe and Palmer, 2018), research on metaphor generation is ∗Work down when the author is interning at UCLA. Literal Input1 The wildfire spread through the forest at an amazing speed. GenMetaphor1 The wildfire danced through the forest at an amazing speed. Literal Input2 The window panes were rattling as the wind blew through them GenMetaphor2 The window panes were trembling as the wind blew through them Table 1: Examples of two generated metaphors Gen- Metaphor1 and GenMetaphor2 by our best model MER- MAID from their literal inputs. under-explored (Yu and Wan, 2019; Stowe et al., 2020). Generating metaphors could impact many downstream applications such as creative writing assistance, literary or poetic content creation. Relevant statistics demonstrate that the most frequent type of metaphor is expressed by verbs (Steen, 2010; Martin, 2006). We therefore focus on the task of generating a metaphor starting from a literal utterance (Stowe et al., 2020), where we transform a literal verb to a metaphorical verb. Ta- …

Content truncated to 500 words