GUTS at SemEval-2022 Task 4: Adversarial Training and Balancing Methods for Patronizing and Condescending Language Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 432 - 437 July 14-15, 2022 ©2022 Association for Computational Linguistics GUTS at SemEval-2022 Task 4: Adversarial Training and Balancing Methods for Patronizing and Condescending Language Detection Junyu Lu, Hao Zhang, Tongyue Zhang, Hongbo Wang, Haohao Zhu, Bo Xu, Hongfei Lin* School of Computer Science and Technology, Dalian University of Technology, China (dutljy,zh373911345,zty9818,hldwhb2017,zhuhh)@mail.dlut.edu.cn (xubo,hflin)@dlut.edu.cn Abstract Patronizing and Condescending Language (PCL) towards vulnerable communities in gen- eral media has been shown to have poten- tially harmful effects. Due to its subtlety and the good intentions behind its use, the audi- ence is not aware of the language’s toxicity. In this paper, we present our method for the SemEval-2022 Task4 titled "Patronizing and Condescending Language Detection". In Sub- task A, a binary classification task, we intro- duce adversarial training based on Fast Gra- dient Method (FGM) and employ pre-trained model in a unified architecture. For Subtask B, framed as a multi-label classification prob- lem, we utilize various improved multi-label cross-entropy loss functions and analyze the performance of our method. In the final evalua- tion, our system achieved official rankings of 17/79 and 16/49 on Subtask A and Subtask B, respectively. In addition, we explore the rela- tionship between PCL and emotional polarity and intensity it contains. Our code is available on Github 1. 1 Introduction Patronizing and Condescending Language (PCL) expresses a superior attitude towards vulnerable communities (e.g. women, refugees, poor families), and describes them or their situation in a charitable way that evokes feelings of compassion (Pérez- Almendros et al., 2022). Although it is generally used involuntarily and with good intentions, the use of PCL can potentially be very harmful, as it feeds stereotypes, routinizes discrimination and drives to greater exclusion. Due to the subtlety of PCL, PCL detection is difficult for both humans and NLP systems and has aroused broad attention. To address the challenge of patronizing and con- descending language detection in general media, Pérez-Almendros et al. (2022) introduce the Task 1https://github.com/Nutpok/GUTS-at-SemEval-2022- Task-4.git 4 at SemEval-2022, and build a dataset with an- notated paragraphs extracted from news articles in English. Given a paragraph, systems must predict whether it contains condescending language or not (Subtask A), and whether it contains any of the 7 subtypes identified in the PCL taxonomy (Subtask B). For Subtask A, a binary classification task, we introduce adversarial training based on Fast Gradi- ent Method (FGM) (Miyato et al., 2016), enhanc- ing the robustness of the model. And in Subtask B, a multi-label classification problem, there is a long-tailed distribution of each label. To address the class imbalance problem, we utilize various improved multi-label cross-entropy loss functions: Focal loss (Lin et al., 2017), Class-balanced focal loss (Cui et al., 2019) and Distribution-balanced loss (Wu et al., 2020). We analyze the performance of our methods and demonstrate the contribution of each component of the architecture. In addition to completing basic evaluation tasks, we also explore the relationship between PCL and emotional polarity and intensity it contains …

Content truncated to 500 words