JCT at SemEval-2022 Task 4-A: Patronism Detection in Posts Written in English using Preprocessing Methods and various Machine Leaerning Methods

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 519 - 524 July 14-15, 2022 ©2022 Association for Computational Linguistics Abstract This paper describes our submissions to SemEval-2022 subtask 4-A - “Patronizing and Condescending Language Detection: Binary Classification". We developed different models for this subtask. We applied 11 supervised machine learning methods and 9 pre- processing methods. Our best submission was a model we built with BertForSequenceClassification. Our experiments indicate that pre-processing stage is a must for a successful model. The dataset for Subtask 1 is highly imbalanced. The F1-scores on the oversampled imbalanced training dataset were higher than the results on the original training dataset. 1 Introduction The explosion of social media in recent years also enables an increasing the number of patronizing and condescending language (PCL). Patronizing language is best described as expressions that are agreeable and show kindness to a person or group in a condescending manner, indicating that the person or group is inferior (McCune and Matthews, 1978). The discourse of condescension has three main characteristics: (1) It does not contain anything openly critical or negative, and often contains insincere praise; (2) it assumes a difference in status and worth between the writer and the person who wrote about him; and (3) this assumed difference is disputed by the listener (Huckin, 2002). PCL can harm individuals or groups of people and may cause harmful effects on society. Therefore, it is important to develop efficient computerized systems capable of detecting PCL (Lo and Wei, 2006). PCL detection is not a simple problem because it requires understanding the context of the situation, the relevant culture, and indirect clues. In social media texts, the problem is harder due to the different levels of ambiguities in natural language and the noisy nature of such texts. In contrast to the offensive language or hate speech detection field, where there has been relatively an extensive research (e.g., Basile et al., 2019; Zampieri et al., 2019; Zampieri et al., 2020), PCL is still a relatively new and open field of study in Natural language processing (NLP) and machine learning (ML) (Pérez-Almendros et al., 2020). Pérez-Almendros et al. (2020) introduced the Don’t Patronize Me! Dataset. This dataset contains paragraphs extracted from news stories, which have been annotated to indicate the presence of PCL at the text span level. This paper describes our research and participation in subtask 4-A for patronism detection in posts written in English. The full description of task 4 in general and 4-A, in particular, is given in Perez-Almendros et al. (2022). The structure of the rest of the paper is as follows. Section 2 introduces a background concerning patronism detection, text pre- processing, and TC with imbalanced classes. Section 3 describes subtask 4-A and its training dataset. In Section 4, we present the submitted models and their experimental results. Section 5 summarizes and suggests ideas for future research. 2 Related Work Various NLP methods have been applied in the detection of several types of harmful language such as …

Content truncated to 500 words