DH-FBK at SemEval-2022 Task 4: Leveraging Annotators’ Disagreement and Multiple Data Views for Patronizing Language Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 324 - 334 July 14-15, 2022 ©2022 Association for Computational Linguistics DH-FBK at SemEval-2022 Task 4: Leveraging Annotators’ Disagreement and Multiple Data Views for Patronizing Language Detection Alan Ramponi Fondazione Bruno Kessler Trento, Italy alramponi@fbk.eu Elisa Leonardelli Fondazione Bruno Kessler Trento, Italy eleonardelli@fbk.eu Abstract The subtle and typically unconscious use of pa- tronizing and condescending language (PCL) in large-audience media outlets undesirably feeds stereotypes and strengthens power-knowledge relationships, perpetuating discrimination to- wards vulnerable communities. Due to its sub- jective and subtle nature, PCL detection is an open and challenging problem, both for com- putational methods and human annotators. In this paper we describe the systems submitted by the DH-FBK team to SemEval-2022 Task 4, aiming at detecting PCL towards vulnerable communities in English media texts. Motivated by the subjectivity of human interpretation, we propose to leverage annotators’ uncertainty and disagreement to better capture the shades of PCL in a multi-task, multi-view learning frame- work. Our approach achieves competitive re- sults, largely outperforming baselines and rank- ing on the top-left side of the leaderboard on both PCL identification and classification. No- ticeably, our approach does not rely on any external data or model ensemble, making it a viable and attractive solution for real-world use. 1 Introduction Detecting patronizing and condescending language (PCL) is an open, challenging, and underexplored research area in natural language processing (Pérez- Almendros et al., 2020; Wang and Potts, 2019). A patronizing and condescending attitude is ex- pressed as a good-natured and beneficial attitude from a person of authority towards others, who are typically depicted in a subtly compassionate way (Pérez-Almendros et al., 2020). PCL is a mildly perceived phenomenon. It is often unconscious, driven by good intentions, and expressed through flowery wordings (Wong et al., 2014; Huckin, 2002). This makes PCL identifica- tion and classification difficult both for NLP sys- tems and human annotators (cf. Figure 1), as it cannot be linked to specific words. Nonetheless, Figure 1: Example showing that patronizing and con- descending language is a subtle linguistic phenomenon that human annotators (a1, a2) often perceive differently, and thus annotate in different ways. it undesirably conveys harmful messages in many ways, as it promotes stereotypes, and a superiority and discriminatory mindset (Fiske, 1993). This is particularly damaging when used by large-audience media outlets, since it drives greater exclusion of already vulnerable communities (Pérez-Almendros et al., 2020). Automatically detecting PCL has the potential to enable a range of applications and re- search directions, such as suggestion tools for news editors to mitigate condescension in writing before publication, and studies on the interplay between condescension and sociodemographic factors. To encourage research on patronizing and con- descending language detection, the SemEval-2022 Task 4 (Pérez-Almendros et al., 2022) has recently been proposed. The shared task aims at investigat- ing methods for the identification of PCL (subtask 1; Figure 1, top) and the categorization of the lin- guistic techniques which are used to express it (sub- task 2; Figure …

Content truncated to 500 words