HumorHunter at SemEval-2021 Task 7: Humor and Offense Recognition with Disentangled Attention

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 275–280 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 275 HumorHunter at SemEval-2021 Task 7: Humor and Offense Recognition with Disentangled Attention Yubo Xie, Junze Li, and Pearl Pu School of Computer and Communication Sciences ´Ecole Polytechnique F´ed´erale de Lausanne, Switzerland {yubo.xie, junze.li, pearl.pu}@epfl.ch Abstract In this paper, we describe our system submit- ted to SemEval 2021 Task 7: HaHackathon: Detecting and Rating Humor and Offense. The task aims at predicting whether the given text is humorous, the average humor rating given by the annotators, and whether the humor rat- ing is controversial. In addition, the task also involves predicting how offensive the text is. Our approach adopts the DeBERTa architec- ture with disentangled attention mechanism, where the attention scores between words are calculated based on their content vectors and relative position vectors. We also took ad- vantage of the pre-trained language models and fine-tuned the DeBERTa model on all the four subtasks. We experimented with sev- eral BERT-like structures and found that the large DeBERTa model generally performs bet- ter. During the evaluation phase, our system achieved an F-score of 0.9480 on subtask 1a, an RMSE of 0.5510 on subtask 1b, an F-score of 0.4764 on subtask 1c, and an RMSE of 0.4230 on subtask 2a (rank 3 on the leader- board). 1 Introduction Humor, appreciated by people with almost any age or cultural background, is perhaps one of the most fascinating human behaviors. Besides providing en- tertainment, humor can also be beneficial to mental health by serving as a moderator of life stress (Lef- court and Martin, 2012), and plays an important role in regulating human-human interaction. As Reeves and Nass (1996) have pointed out, people respond to computers in the same way as they do to real people, which indicates that modeling hu- mor computationally could bring positive effects in human-computer interaction (Nijholt et al., 2003). Despite being universal to human beings, the extent to which people find something humorous varies according to one’s age, gender, or socio-economic status, making humor a highly subjective experi- ence. This poses many challenges to the field of computational humor. Abundant research has been done to enable computers to automatically decide whether humor is entailed in a given piece of text. Early work (Mihalcea and Strapparava, 2005; Mi- halcea et al., 2010) uses manually engineered fea- tures to recognize humor in text, while more recent work (Chen and Soo, 2018; Weller and Seppi, 2019) adopts deep learning approaches and pre-trained language models. SemEval 2021 Task 7: HaHackathon: Detect- ing and Rating Humor and Offense (Meaney et al., 2021) aims at detecting and rating humor as well as offense in short English text. There are four subtasks involved. Subtask 1a is a binary classifica- tion task, predicting if the text would be considered humorous for an average user. Subtask 1b is a regression task and predicts the humor rating of the text if it is considered humorours. Subtask …

Content truncated to 500 words