Rˆ3: Reverse, Retrieve, and Rank for Sarcasm Generation with Commonsense Knowledge

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7976–7986 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7976 R3: Reverse, Retrieve, and Rank for Sarcasm Generation with Commonsense Knowledge Tuhin Chakrabarty1,2∗, Debanjan Ghosh3, Smaranda Muresan2,4 and Nanyun Peng1 1Information Sciences Institute, University of Southern California 2Department of Computer Science, Columbia University 3Educational Testing Service, 4Data Science Institute, Columbia University {tuhin.chakrabarty, smara}@columbia.edu dghosh@ets.org, npeng@isi.edu Abstract We propose an unsupervised approach for sar- casm generation based on a non-sarcastic input sentence. Our method employs a retrieve-and- edit framework to instantiate two major char- acteristics of sarcasm: reversal of valence and semantic incongruity with the context, which could include shared commonsense or world knowledge between the speaker and the lis- tener. While prior works on sarcasm gener- ation predominantly focus on context incon- gruity, we show that combining valence rever- sal and semantic incongruity based on com- monsense knowledge generates sarcastic mes- sages of higher quality based on several crite- ria. Human evaluation shows that our system generates sarcasm better than human judges 34% of the time, and better than a reinforced hybrid baseline 90% of the time. 1 Introduction Studies have shown that the use of sarcasm or verbal irony, can increase creativity on both the speakers and the addressees (Huang et al., 2015), and can serve different communicative purposes such as evoking humor and diminishing or enhanc- ing critique (Burgers et al., 2012). Thus, develop- ing computational models that generate sarcastic messages could impact many downstream appli- cations, such as better conversational agents and creative or humorous content creation. While most computational work has focused on sarcasm detec- tion (Davidov et al., 2010; Gonz´alez-Ib´a˜nez et al., 2011; Riloff et al., 2013; Ghosh et al., 2015; Joshi et al., 2015b; Muresan et al., 2016; Ghosh and Veale, 2017; Ghosh et al., 2017, 2018), research on sarcasm generation is in its infancy (Joshi et al., 2015a; Mishra et al., 2019). Sarcasm generation ∗ The research was conducted when the author was at USC/ISI. Literal Input 1 I hate getting sick from fast food. GenSarc1 I love getting sick from fast food. GenSarc2 [I love getting sick from fast food.] [ Stomach ache is just an additional side effect.] Human 1 Shout out to the Mc donalds for giving me bad food and making me sick right before work in two hours. Literal Input 2 I inherited unfavorable genes from my mother. GenSarc3 I inherited great genes from my mother. GenSarc4 [I inherited great genes from my mother.] [Ugly goes down to the bone.] Human 2 Great I inherited all of my mother’s GOOD genes Table 1: Table showing a literal or non sarcastic input sentence and respective sarcastic outputs. GenSarc1 and GenSarc3 simply reverses the valence, while Gen- Sarc2 and GenSarc4 add commonsense context to cre- ate incongruity or enhance the humorous effect. is a challenging problem since the generated utter- ance should have at least five characteristics (a.k.a. “sarcasm factors”) (Burgers et al., 2012): 1) …

Content truncated to 500 words