Pre-Training Language Models for Identifying Patronizing and Condescending Language: An Analysis

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3902–3911 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3902 Pre-Training Language Models for Identifying Patronizing and Condescending Language: An Analysis Carla P´erez-Almendros, Luis Espinosa-Anke, Steven Schockaert School of Computer Science & Informatics, Cardiff University, UK {perezalmendrosc, espinosa-ankel, schockaerts1}@cardiff.ac.uk Abstract Patronizing and Condescending Language (PCL) is a subtle but harmful type of discourse, yet the task of recognizing PCL remains under-studied by the NLP community. Recognizing PCL is challenging because of its subtle nature, because available datasets are limited in size, and because this task often relies on some form of commonsense knowledge. In this paper, we study to what extent PCL detection models can be improved by pre-training them on other, more established NLP tasks. We find that performance gains are indeed possible in this way, in particular when pre-training on tasks focusing on sentiment, harmful language and commonsense morality. In contrast, for tasks focusing on political speech and social justice, no or only very small improvements were witnessed. These findings improve our understanding of the nature of PCL. Keywords: Patronizing and Condescending Language, Pre-Training Strategies, Natural Language Processing 1. Introduction The study of unfair, ideological, offensive or mislead- ing discourse has become an important and well nour- ished topic of interest within the NLP research com- munity. Most works on this topic address messages with a flagrant and clear intention of harming others, such as hate speech or offensive language detection (Zampieri et al., 2019; Zampieri et al., 2020; Basile et al., 2019). Others seek to address deceiving discourse, as in fake news (Conroy et al., 2015) or propaganda (Da San Martino et al., 2020). However, there are other kinds of discourse that, although equally harm- ful, present themselves in a more subtle way, making it more difficult, both for humans and NLP systems, to detect them. This is the case of Patronizing and Condescending Language (PCL), a type of discourse that presents a person or a community as superior to other(s). The use of PCL is often unconscious and well- intended, especially when referring to vulnerable com- munities (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. PCL has been extensively studied in linguistics and social sci- ences (Margi´c, 2017; Giles et al., 1993; Huckin, 2002; Chouliaraki, 2006). Within NLP, however, only a few authors have addressed this topic. Two examples are Wang and Potts (2019), who compiled a corpus derived from Reddit comments, and our earlier work (Perez- Almendros et al., 2020), where we compiled a cor- pus from news stories about vulnerable communities. In addition, although not focusing on PCL, Sap et al. (2019) studied how certain uses of language indicate power relations, Mendelsohn et al. (2020) discussed the dehumanization of minorities through language and Zhou and Jurgens (2020) investigated the interplay be- tween authoritative …

Content truncated to 500 words