MagicPai at SemEval-2021 Task 7: Method for Detecting and Rating Humor Based on Multi-Task Adversarial Training

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 1153–1159 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1153 MagicPai at SemEval-2021 Task 7: Method for Detecting and Rating Humor Based on Multi-Task Adversarial Training Jian Ma, Shuyi Xie, Haiqin Yang§, Lianxin Jiang, Mengyuan Zhou, Xiaoyi Ruan, Yang Mo Ping An Life Insurance, Ltd. Shenzhen, Guangdong province, China {MAJIAN446,XIESHUYI542,JIANGLIANXIN769,RUANXIAOYI687,MOYANG853}@pingan.com.cn § the corresponding author, email: hqyang@ieee.org Abstract This paper describes MagicPai’s system for Se- mEval 2021 Task 7, HaHackathon: Detecting and Rating Humor and Offense. This task aims to detect whether the text is humorous and how humorous it is. There are four sub- tasks in the competition. In this paper, we mainly present our solution, a multi-task learn- ing model based on adversarial examples, for task 1a and 1b. More specifically, we first vec- torize the cleaned dataset and add the pertur- bation to obtain more robust embedding rep- resentations. We then correct the loss via the confidence level. Finally, we perform interac- tive joint learning on multiple tasks to capture the relationship between whether the text is hu- morous and how humorous it is. The final re- sult shows the effectiveness of our system. 1 Introduction Humor is the tendency of experiences to provoke laughter and provide amusement. Regardless of gender, age or cultural background, it is a special way of language expression to provide an active atmosphere or resolve embarrassment in life while being an important medium for maintaining mental health (Lefcourt and Martin, 1986). Recently, with the rapid development of artificial intelligence, it becomes one of the most hot research topics in nat- ural language processing to recognize humor (Ni- jholt et al., 2003). The task of humor recognition consists of two subtasks: whether the text contains humorous and what level of the humor it is. Early humor recognition methods tackle this task mainly by designing heuristic humor-specific features on classification models (Khodak et al., 2018) and have proved that this automatic way can attain sat- isfactory performance. Nowadays, researchers try to resolve this task by statistical machine learning or deep learning technologies. The SemEval 2021 Task 7, HaHackathon: De- tecting and Rating Humor and Offense, consists of four subtasks: Subtask 1 simulates the previous hu- mor detection task, in which all scores are averaged to provide an average classification score. Subtask 1a is a binary classification task to detect whether the text is humorous. Subtask 1b is a regression task to predict how humorous it is for ordinary users in a value range from 0 to 5. Subtask 1c is also a binary classification task to predict whether the humor grade causes controversy if the text is classified as humorous. Subtask 2 aims to predict how offensive text for an ordinary user is in an integral value range from 0 and 5. Due to the highly subjective nature of humor detection, the data is labeled by people with differ- ent profile in gender, age group, political position, income level, social status, …

Content truncated to 500 words