MOVER: Mask, Over-generate and Rank for Hyperbole Generation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 6018 - 6030 July 10-15, 2022 ©2022 Association for Computational Linguistics MOVER: Mask, Over-generate and Rank for Hyperbole Generation Yunxiang Zhang, Xiaojun Wan Wangxuan Institute of Computer Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {yx.zhang,wanxiaojun}@pku.edu.cn Abstract Despite being a common figure of speech, hy- perbole is under-researched in Figurative Lan- guage Processing. In this paper, we tackle the challenging task of hyperbole generation to transfer a literal sentence into its hyperbolic paraphrase. To address the lack of available hyperbolic sentences, we construct HYPO-XL, the first large-scale English hyperbole corpus containing 17,862 hyperbolic sentences in a non-trivial way. Based on our corpus, we pro- pose an unsupervised method for hyperbole generation that does not require parallel literal- hyperbole pairs. During training, we fine-tune BART (Lewis et al., 2020) to infill masked hy- perbolic spans of sentences from HYPO-XL. During inference, we mask part of an input literal sentence and over-generate multiple pos- sible hyperbolic versions. Then a BERT-based ranker selects the best candidate by hyperbol- icity and paraphrase quality. Automatic and human evaluation results show that our model is effective at generating hyperbolic paraphrase sentences and outperforms several baseline sys- tems. 1 Introduction Hyperbole is a figure of speech that deliberately exaggerates a claim or statement to show emphasis or express emotions. If a referent has a feature X, a hyperbole exceeds the credible limits of fact in the given context and presents it as having more of that X than warranted by reality (Claridge, 2010). Take the following example, “I won’t wait for you: it took you centuries to get dressed.” It over-blows the time for someone to get dressed with a single word “centuries” and thus creates a heightened effect. From a syntactic point of view, Claridge (2010) classifies hyperbole into word-level, phrase- level and clause-level types, and conclude that the former two types are more common in English. Although hyperbole is considered as the second most frequent figurative device (Kreuz and Roberts, 1993), it has received less empirical attention in the NLP community. Recently Tian et al. (2021) addressed the generation of clause-level hyperbole. In this paper, we instead focus on word-level and phrase-level hyperbole, which can be unified as span-level hyperbole. To tackle the hyperbole generation problem we need to address three main challenges: • The lack of training data that either consists of large-scale hyperbolic sentences or literal- hyperbole pairs, which are necessary to train an unsupervised or supervised model. • The tendency of generative language models to produce literal text rather than hyperboles. • Trade-off between content preservation and hyperbolic effect of the generated sentences. In order to address the above challenges, we pro- pose MOVER (Mask, OVEr-generate and Rank), an unsupervised approach to generating hyperbolic paraphrase from literal input. Our approach does not require parallel data for training, thus allevi- ating the issue of scarce data. Still, we need …

Content truncated to 500 words