Dartmouth at SemEval-2022 Task 6: Detection of Sarcasm

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 912 - 918 July 14-15, 2022 ©2022 Association for Computational Linguistics Dartmouth at SemEval-2022 Task 6: Detection of Sarcasm Rishik Lad, Weicheng Ma, Soroush Vosoughi Department of Computer Science, Dartmouth College {rishik.lad.23,weicheng.ma.gr,soroush.vosoughi}@dartmouth.edu Abstract This paper introduces the result of Team Dart- mouth’s experiments on each of the five sub- tasks for the detection of sarcasm in En- glish and Arabic tweets. This detection was framed as a classification problem, and our contributions are threefold: we devel- oped an English binary classifier system with RoBERTaBASE, an Arabic binary classifier with XLM-RoBERTaBASE, and an English multil- abel classifier with BERTBASE. Preprocessing steps are taken with labeled input data prior to tokenization, such as extracting and appending verbs/adjectives or representative/significant keywords to the end of an input tweet to help the models better understand and generalize sarcasm detection. We also discuss the results of simple data augmentation techniques to im- prove the quality of the given training dataset as well as an alternative approach to the ques- tion of multilabel sequence classification. Ul- timately, our systems place us in the top 14 participants for each of the five subtasks. 1 Introduction Sarcasm is a form of irony that occurs when there is a discrepancy between the literal and intended meanings of a text or utterance. This discrepancy typically manifests itself in the form of dislike, contempt, or derogation. Furthermore, sarcasm can be divided into multiple types: general sarcasm, irony, satire, understatement, overstatement, and rhetorical question. This task concerns itself with the detection of sarcasm in online tweets (Abu Farha et al., 2022). This is an important issue to solve because the na- ture of sarcasm can interfere with the effectiveness of natural language processing models, particularly when conducting sentiment analysis, opinion min- ing, or other emotion-based tasks. Machine learn- ing models deployed for such business use cases can be negatively impacted when processing sar- castic texts and provide inaccurate results, thereby harming an organization’s bottom line. Therefore, it is critical that machine learning models be devel- oped that can understand how to detect, ingest, and truly understand sarcasm. This paper discusses how to identify sarcastic tweets in binary and multi-label classification con- texts for both English and Arabic, as directed by SemEval-2022 Task 6 (Abu Farha et al., 2022). There are five subtasks: general sarcasm detection in standalone English and Arabic tweets (Task A), identification of sarcastic category in an English tweet (Task B), and identification of the sarcastic tweet in an English/Arabic pair of tweets (Task C). Our experiments show that pre-trained trans- former models demonstrate a strong ability of solv- ing most of the subtasks of this challenge. Specif- ically, we fine-tune a RoBERTaBASE model to de- tect the presence of sarcasm in tweets for the En- glish components of Tasks A and C. For the Ara- bic version of Tasks A and C, we apply an XLM- RoBERTaBASE model. For Task B, which is framed as a multilabel sequence …

Content truncated to 500 words