Identifying Exaggerated Language

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7024–7034, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7024 Identifying Exaggerated Language Li Kong1, Chuanyi Li1, Jidong Ge1, Bin Luo1 and Vincent Ng2 1State Key Laboratory for Novel Software Technology, Nanjing University, China 2Human Language Technology Research Institute, University of Texas at Dallas, USA kl nju@126.com, {lcy, gjd, luobin}@nju.edu.cn, vince@hlt.utdallas.edu Abstract While exaggeration is one of the most preva- lent rhetorical devices, it is arguably one of the least studied in the figurative language processing community. We contribute to the computational study of exaggeration by (1) creating the first Chinese corpus focusing on sentence-level hyperbole detection, with the goal of facilitating a cross-lingual study on this phenomenon, (2) performing a statistical and manual analysis of our corpus, with the goal of gaining insights into the strategies hu- mans employ when creating hyperboles, and (3) addressing the automatic hyperbole detec- tion task with deep learning techniques. 1 Introduction Recent years have seen a surge of interest in the automatic processing of figurative language in the NLP community, as evidenced by the successful organization of the NAACL 2018 Workshop on Fig- urative Language Processing. Much of the work on figurative language processing conducted so far, however, has focused on metaphor and metonymy (Tsvetkov et al., 2014), and more recently, sar- casm (Hazarika et al., 2018), idioms (Liu and Hwa, 2018), and puns (He et al., 2019). In particular, hy- perbole, also known as exaggeration, is a relatively under-studied phenomenon in the community. This is somewhat surprising, especially given that the prevalence of hyperbole as a rhetorical device is only second to metaphor (Kreuz et al., 1996). Hu- mans exaggerate in different situations for various purposes, such as creating amusement, expressing emotion and drawing attention (Li, 2013). The vast amount of work on metaphor detection in the past few years was stimulated in large part by the availability of standard evaluation corpora. Progress on the computational study of exaggera- tion, on the other hand, is hindered by the lack of annotated resources. To our knowledge, HYPO, the first dataset that focuses on exaggeration, was only released in late 2018 (Troiano et al., 2018). HYPO consists of 709 hyperbolic sentences, each of which has a non-hyperbolic version created by manually paraphrasing its hyperbolic counterpart. Given the dataset, Troiano et al. introduced the task of automatic hyperbole detection, where the goal is to determine whether a sentence is a hyperbole. Given the status quo, our goal is to further the computational study of exaggeration. Specifically, our contributions in this work are three-fold. First, we create HYPO-cn, the first Chinese dataset on ex- aggeration. HYPO-cn consists of 4762 sentences, of which 2680 are hyperbolic and 2082 are non- hyperbolic. To stimulate research on the computa- tional study of exaggeration, we make HYPO-cn publicly available. We believe that this dataset can complement Troiano et al.’s English dataset and facilitate a cross-lingual study of exaggeration. Our second contribution involves conducting an empirical analysis of …

Content truncated to 500 words