Felix&Julia at SemEval-2022 Task 4: Patronizing and Condescending Language Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 357 - 362 July 14-15, 2022 ©2022 Association for Computational Linguistics Abstract This paper describes the authors’ submission to the SemEval-2022 task 4: Patronizing and Condescending Language (PCL) Detection. The aim of the task is the detection and classification of PCL in an annotated dataset. The authors of this paper worked on two different models with finetuned hyperparameters focusing on number of epochs, training batch size, evaluation batch size, gradient accumulation steps and learning rate. The authors submitted one RoBERTa model and one DistilBERT model. Both systems performed better than the random and RoBERTA baseline given by the task organizers. The RoBERTA model finetuned by the authors performed better in both subtasks than the DistilBERT model. 1 Introduction With the rise of social media, online hate speech has skyrocketed and has posed a problem for social media platforms: How can the giant number of messages on social platforms be surveilled, so that hateful comments can be reported and deleted? The answer to this is hate speech detection, a discipline within Natural Language Processing that has become increasingly popular and successful in recent years. However, apart from hate speech, there is also other harmful language that should be studied. This is what the research by Perez-Almendros, Espinosa-Anke and Schokaert dives into. They collected an annotated dataset that focuses on a type of harmful language that is not so easily 1 The code can be found here: https://github.com/julia- ecrevisse/SemEval2022 detected: patronizing and condescending language (PCL). The researchers describe PCL as follows: “An entity engages in PCL when its language use shows a superior attitude towards others or depicts them in a compassionate way.” (Perez-Almendros et. al., 2020) One of the difficulties in detecting PCL as opposed to openly hateful speech, is that persons who use PCL often do not intend to do harm, but instead want to support the groups that they name. PCL is usually aimed at vulnerable communities which makes the detection of PCL even more important. The SemEval 2022 task 4 competition is based on this research by Perez-Almendros et. al: the detection of PCL. In two subtasks, one a binary classification and one a multi-label classification, the participants of the SemEval competition were tasked to submit up to two models. The authors of this paper participated in both subtasks and propose a finetuning based approach on a pre-trained RoBERTa language model. Two models were submitted to the task organisers, one RoBERTa model and one DistilBERT model. The RoBERTa model performed better in both subtasks.1 In this paper, the authors will first describe the task set out by the SemEval 2022 competition in sections 2. In section 3, the authors explain the relevance of related work to this topic. Section 4 gives an overview over the BERT model. In the subsequent section, the two pre-trained models are presented. In the section experimental setup, the two different finetuned models are described in detail. Section 7 presents the results of …

Content truncated to 500 words