Modelling Context with User Embeddings for Sarcasm Detection in Social Media

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 167–177, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Modelling Context with User Embeddings for Sarcasm Detection in Social Media Silvio Amir Byron C. Wallace† Hao Lyu† Paula Carvalho M´ario J. Silva INESC-ID Lisboa, Instituto Superior T´ecnico, Universidade de Lisboa †iSchool, University of Texas at Austin samir@inesc-id.pt byron.wallace@utexas.edu xalh8083@gmail.com pcc@inesc-id.pt mjs@inesc-id.pt Abstract We introduce a deep neural network for automated sarcasm detection. Recent work has emphasized the need for mod- els to capitalize on contextual features, be- yond lexical and syntactic cues present in utterances. For example, different speak- ers will tend to employ sarcasm regard- ing different subjects and, thus, sarcasm detection models ought to encode such speaker information. Current methods have achieved this by way of laborious feature engineering. By contrast, we pro- pose to automatically learn and then ex- ploit user embeddings, to be used in con- cert with lexical signals to recognize sar- casm. Our approach does not require elab- orate feature engineering (and concomi- tant data scraping); fitting user embed- dings requires only the text from their previous posts. The experimental results show that the our model outperforms a state-of-the-art approach leveraging an ex- tensive set of carefully crafted features. 1 Introduction Existing social media analysis systems are ham- pered by their inability to accurately detect and in- terpret figurative language. This is particularly rel- evant in domains like the social sciences and poli- tics, in which the use of figurative communication devices such as verbal irony (roughly, sarcasm) is common. Sarcasm is often used by individuals to express opinions on complex matters and regard- ing specific targets (Carvalho et al., 2009). Early computational models for verbal irony and sarcasm detection tended to rely on shallow methods exploiting conditional token count reg- ularities. But lexical clues alone are insufficient Figure 1: An illustrative tweet. to discern ironic intent. Appreciating the con- text of utterances is critical for this; even for hu- mans (Wallace et al., 2014). Indeed, the exact same sentence can be interpreted as literal or sar- castic, depending on the speaker. Consider the sarcastic tweet in Figure 1 (ignoring for the mo- ment the attached #sarcasm hashtag). Without knowing the author’s political leanings, it would be difficult to conclude with certainty whether the remark was intended sarcastically or in earnest. Recent work in sarcasm detection on social me- dia has tried to incorporate contextual information by exploiting the preceding messages of a user, to e.g., detect contrasts in sentiments expressed to- wards named entities (Khattri et al., 2015), infer behavioural traits (Rajadesingan et al., 2015) and capture the relationship between authors and the audience (Bamman and Smith, 2015). However, all of these approaches require the design and im- plementation of complex features that explicitly encode the content and (relevant) context of mes- sages to be classified. This feature engineering is labor intensive, and depends on external tools and resources. Therefore, deploying such systems in practice is expensive, time-consuming …

Content truncated to 500 words