Gulu at SemEval-2021 Task 7: Detecting and Rating Humor and Offense

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 1120–1124 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1120 Gulu at SemEval-2021 Task 7: Detecting and Rating Humor and Offense Maoqin Yang School of Information, Yunnan University, Yunnan, P.R. China maomaoq33@gmail.com Abstract Humor recognition is a challenging task in natural language processing. This document presents my approaches to detect and rate hu- mor and offense from the given English text. This task includes 2 tasks: task 1 which con- tains 3 subtasks (1a, 1b, and 1c), and task 2. Subtask 1a and 1c can be regarded as classifi- cation problems and take ALBERT as the ba- sic model. Subtask 1b and 2 can be viewed as regression issues and take RoBERTa as the basic model. And finally, team-Gulu scores in subtask 1a with a weighted average F1 score of 0.9190, in subtask 1b with an RMSE score of 0.7405, in subtask 1c with a weighted aver- age F1 score of 0.5561, and in subtask 2 with an RMSE score of 0.5807 on the private leader board. 1 Introduction For social animals like humans, humor is an ef- fective bonus. From the perspective of evolution- ary psychology, “humorous” often means superior creativity, in other words, a smarter mind. There- fore, it is important to recognize whether a sen- tence is humorous and how humorous the sentence is. It is a bit impractical to recognize such a huge data set by humans, so it becomes necessary for us to develop a system to automatically detect hu- mor. In this task, the organizer collects labels and ratings from a balanced age group of 18-70. An- notators also represent various genders, political positions, and income levels. Therefore, for some texts classified as humorous, we should once again prove whether they are controversial and predict the offensiveness of the text. For more specific content, please refer to the official website of the competition 1. 1https://competitions.codalab.org/ competitions/27446 Because the pre-trained and deep learning mod- els have shown excellent performance in many NLP problems such as classification and topic extraction(Zampieri et al., 2019), so I use deep learning methods to deal with those four tasks. According to the latest related research progress, the transformer-based language model has become my favorite model. In order to facilitate the un- derstanding of the corresponding model of each subtask, I made it into a table shown as Table 1. I choose A Lite BERT (ALBERT)(Lan et al., 2019) as my basic model in subtask 1a. In subtask 1b and subtask 2, I choose Bidirectional Encoder Representations for Transformers (BERT)(Devlin et al., 2018) model as my basic model. In subtask 1c, A Robustly Optimized BERT (RoBERTa)(Liu et al., 2019) has been chosen. To get a more ef- fective and higher accuracy model in subtask 1a, BiGRU combined with attention. To prove the ef- fectiveness of this model, there are also compar- ative experiments with other neural networks for task 1c. To obtain as much effective information as …

Content truncated to 500 words