Exploring Author Context for Detecting Intended vs Perceived Sarcasm

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2854–2859 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2854 Exploring Author Context for Detecting Intended vs Perceived Sarcasm Silviu Vlad Oprea School of Informatics University of Edinburgh Edinburgh, United Kingdom silviu.oprea@ed.ac.uk Walid Magdy School of Informatics University of Edinburgh Edinburgh, United Kingdom wmagdy@inf.ed.ac.uk Abstract We investigate the impact of using author con- text on textual sarcasm detection. We define author context as the embedded representa- tion of their historical posts on Twitter and suggest neural models that extract these rep- resentations. We experiment with two tweet datasets, one labelled manually for sarcasm, and the other via tag-based distant supervision. We achieve state-of-the-art performance on the second dataset, but not on the one labelled manually, indicating a difference between in- tended sarcasm, captured by distant supervi- sion, and perceived sarcasm, captured by man- ual labelling. 1 Introduction Sarcasm is a form of irony that occurs when there is a discrepancy between the literal meaning of an utterance and its intended meaning. This discrep- ancy is used to express a form of dissociative at- titude towards a previous proposition, often in the form of contempt or derogation (Wilson, 2006). Sarcasm is omnipresent on the social web and can be highly disruptive of systems that harness this data (Maynard and Greenwood, 2014). It is therefore imperative to devise model for textual sarcasm detection. The effectiveness of such mod- els depends on the quality of labelled data used for training. Two methods are commonly used to label texts for sarcasm: manual labelling by human an- notators; and tag-based distant supervision. In the latter, texts are considered sarcastic if they contain specific tags, such as #sarcasm and #sarcastic. Most work on computational sarcasm detection extracts lexical and pragmatic cues available in the text being classified (Campbell and Katz, 2012; Riloff et al., 2013; Joshi et al., 2016; Tay et al., 2018). However, sarcasm is a contextual phe- nomenon and detecting it often requires prior in- formation about the author, audience and previous interactions between them, that originates beyond the text itself (Rockwell and Theriot, 2001a). In this work we investigate the impact of au- thor context on the current sarcastic behaviour of the author. We identify author context with the embedded representation of their historical tweets. We use the term user to refer to the author of a tweet and the phrase user embedding to refer to such a representation. Given a tweet t posted by user ut with user embedding et, we address two questions: (1) Is et predictive of the sarcastic na- ture of t? (2) Is the predictive power of et on the sarcastic nature of t the same if t is labelled via manual labelling vs distant supervision? To our knowledge, previous research that con- siders author context (Rajadesingan, Zafarani, and Liu, 2015; Bamman and Smith, 2015; Amir et al., 2016; Hazarika et al., 2018) only experiments on distant supervision datasets. We …

Content truncated to 500 words