A Deeper Look into Sarcastic Tweets Using Deep Convolutional Neural Networks

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1601–1612, Osaka, Japan, December 11-17 2016. A Deeper Look into Sarcastic Tweets Using Deep Convolutional Neural Networks Soujanya Poria, Erik Cambria, Devamanyu Hazarika, Prateek Vij Nanyang Technological University 50 Nanyang Ave, Singapore 639798 {sporia, cambria}@ntu.edu.sg {devamanyu, prateek}@sentic.net Abstract Sarcasm detection is a key task for many natural language processing tasks. In sentiment anal- ysis, for example, sarcasm can flip the polarity of an “apparently positive” sentence and, hence, negatively affect polarity detection performance. To date, most approaches to sarcasm detec- tion have treated the task primarily as a text categorization problem. Sarcasm, however, can be expressed in very subtle ways and requires a deeper understanding of natural language that stan- dard text categorization techniques cannot grasp. In this work, we develop models based on a pre-trained convolutional neural network for extracting sentiment, emotion and personality fea- tures for sarcasm detection. Such features, along with the network’s baseline features, allow the proposed models to outperform the state of the art on benchmark datasets. We also address the often ignored generalizability issue of classifying data that have not been seen by the models at learning phase. 1 Introduction Sarcasm is defined as “a sharp, bitter, or cutting expression or remark; a bitter gibe or taunt”. As the fields of affective computing and sentiment analysis have gained increasing popularity (Cambria, 2016), it is a major concern to detect sarcastic, ironic, and metaphoric expressions. Sarcasm, especially, is key for sentiment analysis as it can completely flip the polarity of opinions. Understanding the ground truth, or the facts about a given event, allows for the detection of contradiction between the objective polarity of the event (usually negative) and its sarcastic characteristic by the author (usually positive), as in “I love the pain of breakup”. Obtaining such knowledge is, however, very difficult. In our experiments, we exposed the classifier to such knowledge extracted indirectly from Twitter. Namely, we used Twitter data crawled in a time period, which likely contain both the sarcastic and non- sarcastic accounts of an event or similar events. We believe that unambiguous non-sarcastic sentences provided the classifier with the ground-truth polarity of those events, which the classifier could then contrast with the opposite estimations in sarcastic sentences. Twitter is a more suitable resource for this purpose than blog posts, because the polarity of short tweets is easier to detect (as all the information necessary to detect polarity is likely to be contained in the same sentence) and because the Twitter API makes it easy to collect a large corpus of tweets containing both sarcastic and non-sarcastic examples of the same event. Sometimes, however, just knowing the ground truth or simple facts on the topic is not enough, as the text may refer to other events in order to express sarcasm. For example, the sentence “If Hillary wins, she will surely be pleased to recall Monica each time she enters the Oval Office :P :D”, which refers to the …

Content truncated to 500 words