UPB at SemEval-2021 Task 7: Adversarial Multi-Task Learning for Detecting and Rating Humor and Offense

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 1160–1168 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1160 UPB at SemEval-2021 Task 7: Adversarial Multi-Task Learning for Detecting and Rating Humor and Offense R˘azvan-Alexandru Sm˘adu, Dumitru-Clementin Cercel, Mihai Dascalu University Politehnica of Bucharest, Faculty of Automatic Control and Computers razvan.smadu@stud.acs.upb.ro {dumitru.cercel, mihai.dascalu}@upb.ro Abstract Detecting humor is a challenging task since words might share multiple valences and, de- pending on the context, the same words can be even used in offensive expressions. Neural network architectures based on Transformer obtain state-of-the-art results on several Natu- ral Language Processing tasks, especially text classification. Adversarial learning, combined with other techniques such as multi-task learn- ing, aids neural models learn the intrinsic prop- erties of data. In this work, we describe our adversarial multi-task network, AMTL- Humor, used to detect and rate humor and of- fensive texts from Task 7 at SemEval-2021. Each branch from the model is focused on solving a related task, and consists of a BiL- STM layer followed by Capsule layers, on top of BERTweet used for generating contextual- ized embeddings. Our best model consists of an ensemble of all tested configurations, and achieves a 95.66% F1-score and 94.70% accu- racy for Task 1a, while obtaining RMSE scores of 0.6200 and 0.5318 for Tasks 1b and 2, re- spectively. 1 Introduction Sentiment analysis studies expressed opinions and feelings, affective states, and subjective informa- tion introduced while conveying ideas to others. In recent years, social media has encountered a major change in ways of interaction and in the freedom of expressing moods to others. For example, individ- uals frequently use emojis, or even abbreviations, to emphasize their feelings. Therefore, new ma- chine learning systems are developed to predict sentiment valences; however, this task is not triv- ial, as the detection of sentiments is challenging even for humans. This is mainly generated by the diversity of contexts and differences among people, such as socio-cultural status, age, or gender. For example, certain jokes might be confusing due to past experiences or the current mood while reading the joke. Similarly, language ambiguity can intro- duce difficulties in understanding the text and in altering the perceived emotions. Moreover, individuals express a wide range of sentiments, such as happiness, humor, or anger. Words that are used to express such feelings, might share multiple valences and meanings, depending on the context. Therefore, a sequence of words that is, for example, humorous in a certain context, can be considered offensive in another situation. Such examples can rely on bad jokes regarding a person’s position in society, ethnicity, or political ideology. Another characteristic that impacts a text’s level of humor is how clear the idea is conveyed to the reader. In this regard, SemEval-2021 Task 7 - Ha- Hackathon: Detecting and Rating Humor and Of- fense (Meaney et al., 2021) introduced four sub- tasks: Task 1a (classification) includes the class of the text, which can be either humorous or not; Task 1b (regression) presents …

Content truncated to 500 words