SemEval-2022 Task 4: Patronizing and Condescending Language Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 298 - 307 July 14-15, 2022 ©2022 Association for Computational Linguistics SemEval-2022 Task 4: Patronizing and Condescending Language Detection Carla Perez-Almendros Luis Espinosa-Anke School of Computer Science & Informatics Cardiff University {perezalmendrosc,espinosa-ankel,schockaerts1}@cardiff.ac.uk Steven Schockaert Abstract This paper presents an overview of Task 4 at SemEval-2022, which was focused on detect- ing Patronizing and Condescending Language (PCL) towards vulnerable communities. Two sub-tasks were considered: a binary classifica- tion task, where participants needed to classify a given paragraph as containing PCL or not, and a multi-label classification task, where par- ticipants needed to identify which types of PCL are present (if any). The task attracted 77 teams. We provide an overview of how the task was organized, discuss the techniques that were em- ployed by the different participants, and sum- marize the main resulting insights about PCL detection and categorization. 1 Introduction The study of unfair, misleading or offensive lan- guage has attracted the interest of many scholars from the NLP research community. Most rele- vant tasks in this context focus on explicit, aggres- sive and flagrant phenomena, such as fake news detection or fact-checking (Conroy et al., 2015; Nakov et al., 2018; Atanasova et al., 2019; Barrón- Cedeno et al., 2020); detecting propaganda tech- niques (Da San Martino et al., 2020); modeling offensive language (Zampieri et al., 2019, 2020) identifying hate speech (Basile et al., 2019); and ru- mour propagation (Derczynski et al., 2017). How- ever, there also exist subtler but equally harmful types of language, which have received less atten- tion by the NLP community, and which, due to their subtle nature, we can expect to be more diffi- cult to detect. This is the case, among others, for Patronizing and Condescending Language (PCL), which was the focus of Task 4 at SemEval-2022. An entity engages in patronizing or condescend- ing communication when its use of language re- veals a superior attitude towards others. These attitudes, when normalized, routinize discrimina- tion and make it less visible (Ng, 2007). Further- more, the use of PCL is often unconscious and well-intended, especially when referring to vulner- able communities (Wilson and Gutierrez, 1985; Merskin, 2011). This good will can make PCL especially harmful, as the audience receives this discriminatory language with low defense and is often unaware of its effects. Research in sociolinguistics presents PCL as a subtle, often unconscious but harmful and discrim- inative kind of language (Mendelsohn et al., 2020). It creates and feeds stereotypes (Fiske, 1993), which result in greater exclusion, rumour spread- ing and misinformation (Nolan and Mikami, 2013). PCL also tends to strengthen power-knowledge re- lationships (Foucault, 1980), calling for charitable action instead of cooperation and presenting those who can help as saviours of those in a less privi- leged position (Bell, 2013; Straubhaar, 2015). Fur- thermore, PCL tends to conceal who is responsible for very deep-rooted societal problems, sometimes by implicitly or explicitly blaming the underprivi- leged communities or individuals for their situation, and often involves …

Content truncated to 500 words