A deep-learning framework to detect sarcasm targets

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6336–6342, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6336 A deep-learning framework to detect sarcasm targets 1Jasabanta Patro, 2Srijan Bansal, 3Animesh Mukherjee, Indian Institute of Technology Kharagpur, India – 721302 {1jasabantapatro, 2srijanbansal97}@iitkgp.ac.in , 3animeshm@cse.iitkgp.ac.in Abstract In this paper we propose a deep learning framework for sarcasm target detection in pre- defined sarcastic texts. Identification of sar- casm targets can help in many core natural lan- guage processing tasks such as aspect based sentiment analysis, opinion mining etc. To begin with, we perform an empirical study of the socio-linguistic features and identify those that are statistically significant in indi- cating sarcasm targets (p-values in the range (0.05, 0.001)). Finally, we present a deep- learning framework augmented with socio- linguistic features to detect sarcasm targets in sarcastic book-snippets and tweets. We achieve a huge improvement in the perfor- mance in terms of exact match and dice score as compared to the current state-of-the-art baseline. 1 Introduction Computational sarcasm is a very well studied re- search area in computational linguistics (Joshi et al., 2017). Sentiment analysis and opinion mining of sarcastic texts are known to be diffi- cult problems (Pang et al., 2008). For instance, in aspect based sentiment analysis, which deals with the identification of sentiment expressed to- ward different aspects or dimensions of the entities present in the text, it is very important to identify the sarcasm targets and sentiments toward them in the texts. Thus, if a user expresses a sarcastic ut- terance such as “My laptop has an awesome bat- tery life that lasts for 15 minutes”, the tool should recognize that the speaker is expressing a nega- tive sentiment toward the battery life of the lap- top, even though, it has a positive sentiment word ‘awesome’ in it. Similarly the opinion mining tool should identify the negative opinion of the user ex- pressed toward the entity “battery”. Sarcasm tar- get identification can also benefit natural language generation; for example, after detection of entity toward which a negative sentiment is expressed in a sarcastic text, a natural language generation sys- tem will have more context to generate a response. Similarly, a sentiment analysis tool will flag the sentiment in a sarcastic text toward the correct as- pect of a product or the entity which can help to build a more accurate product review. In this paper we present a novel method for sarcasm target iden- tification with the help of deep learning techniques in addition to a set of socio-linguistic features. There is a lot of literature that deal with the sar- casm detection in text (Joshi et al., 2017), but only Joshi et al. (2018) have addressed the problem of sarcasm target identification. The sarcasm target is defined as the entity or situation that is being mocked or ridiculed at in the sarcastic text. For- mally, the sarcasm target identification is …

Content truncated to 500 words