shefnlp at SemEval-2023 Task 10: Compute-Efficient Category Adapters

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1069–1075 July 13-14, 2023 ©2023 Association for Computational Linguistics shefnlp at SemEval-2023 Task 10: Compute-Efficient Category Adapters Thomas Pickard and Tyler Loakman and Mugdha Pandya The University of Sheffield Sheffield, UK {tmrpickard1, tcloakman1, mugdha.pandya}@sheffield.ac.uk Abstract As social media platforms grow, so too does the volume of hate speech and negative sentiment expressed towards particular social groups. In this paper, we describe our approach to SemEval-2023 Task 10, involving the detec- tion and classification of online sexism (abuse directed towards women), with fine-grained categorisations intended to facilitate the devel- opment of a more nuanced understanding of the ideologies and processes through which on- line sexism is expressed. We experiment with several approaches involving language model finetuning, class-specific adapters, and pseudo- labelling. Our best-performing models involve the training of adapters specific to each subtask category (combined via fusion layers) using a weighted loss function, in addition to perform- ing naive pseudo-labelling on a large quantity of unlabelled data. We successfully outperform the baseline models on all 3 subtasks, placing 56th (of 84) on Task A, 43rd (of 69) on Task B, and 37th (of 63) on Task C. 1 Introduction Sexism, as defined within this task (Kirk et al., 2023), refers to any abuse or negative sentiment that is directed towards women based on their gen- der, or based on their gender combined with one or more other identity attributes. Whilst sexism may, in the general case, be gender agnostic (i.e. also applies to discrimination towards men or any other identified gender), it remains true that this form of discrimination disproportionately targets women. As the userbases of social media plat- forms grow, and disputes regarding the censor- ship and moderation of particular opinions and ideologies increase, extreme negative opinions to- wards particular groups are allowed to proliferate. Consequently, much recent research in NLP has been on the detection of various forms of online hate, including racism, homophobia, and sexism, to combat its growing prevalence (Alkomah and Ma, 2022). Within SemEval2023 Task 10 (Kirk et al., 2023), the 3 component subtasks relate to in- creasing levels of granularity in the classification of sexism in short English texts sourced from Reddit and Gab. Task A presents a binary detection task to determine whether social media posts express sexism, whilst Task B splits this further into a 4- way multi-class classification task regarding broad categories of sexist content (threats, derogation, an- imosity, and prejudiced discussions). Finally, Task C presents the most fine-grained level of classifica- tion, presenting an 11-class taxonomy (described as "vectors of sexism") which further splits the 4 classes from Task B into specific constituents. 2 Background Recent years have witnessed many works in the area of hate speech and offensive language detec- tion. As the field develops, research is beginning to diverge from simple offence detection into more fine-grained classification into areas such as racism, homophobia, and sexism, thereby getting closer to identifying not only whether language is hate- ful, …

Content truncated to 500 words