Sapphire at SemEval-2022 Task 4: A Patronizing and Condescending Language Detection Model Based on Capsule Networks

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 405 - 408 July 14-15, 2022 ©2022 Association for Computational Linguistics Sapphire at SemEval-2022 Task 4: A Patronizing and Condescending Language Detection Model Based on Capsule Networks Sihui Li Yunnan University, Yunnan, P.R. China sihui_li@mail.ynu.edu.cn Xiaobing Zhou * Yunnan University, Yunnan, P.R. China zhouxb@ynu.edu.cn Abstract This paper introduces the related work and the results of Team Sapphire’s system for SemEval- 2022 Task 4: Patronizing and Condescending Language Detection. We only participated in subtask 1. The task goal is to judge whether a news text contains PCL. This task can be considered as a task of binary classification of news texts. In this binary classification task, the BERT-base model is adopted as the pre-trained model used to represent textual information in vector form and encode it. Capsule networks is adopted to extract features from the encoded vectors. The official evaluation metric for sub- task 1 is the F1 score over the positive class. Finally, our system’s submitted prediction re- sults on test set achieved the score of 0.5187. 1 Introduction Patronizing and Condescending Language (PCL) can be considered when someone’s language has a superior attitude towards others, demeans others, or describes the situation of others in a compassion- ate way. Such expressions are often unconscious, and are used by people to try to induce action or raise awareness. Because of its subtlety and of- ten well-meaning when used, users often overlook the demeaning elements of this expression. Such elements may contribute to the stereotyped influ- ence of society on a group, making discrimination normalized and even leading to stronger exclusion (Pérez-Almendros et al., 2022). Detecting PCL in media text is a challenging task. Recognizing PCL based on Natural Language Processing (NLP) can alert speakers to examine the rationality of their speeches, so that speeches can be more inclusive and constructive, which in turn leads to more responsible communication. When processing corpus, the pre-trained model can convert text information into vector representa- tion, making it more suitable for NLP tasks. Early pre-trained models were designed to learn repre- sentational word embeddings, such as Word2Vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014). Although such methods can capture the se- mantics of words through word embeddings, they cannot capture the concepts in the context. With the introduction of new technologies, there are now pre-trained models that can learn to represent contextual word embeddings, such as the ELMo (Peters et al., 2018) model based on LSTM (Shi et al., 2015) and the BERT (Devlin et al., 2018) model based on Transformer Encoder (Vaswani et al., 2017). In recent years, using deep neural networks in NLP, such as Convolutional Neural Networks (CNNs) in text classification (Kim, 2014), has be- come mainstream. Capsule networks (Sabour et al., 2017), as a structure proposed on the basis of CNNs to improve spatial sensitivity in computer vision, is also used in text classification tasks (Yang et al., 2019; Ding et al., 2019). Kim et al. (2020) …

Content truncated to 500 words