UniBoe’s at SemEval-2023 Task 10: Model-Agnostic Strategies for the Improvement of Hate-Tuned and Generative Models in the Classification of Sexist Posts

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1138–1147 July 13-14, 2023 ©2023 Association for Computational Linguistics UniBoe’s at SemEval-2023 Task 10: Model-Agnostic Strategies for the Improvement of Hate-Tuned and Generative Models in the Classification of Sexist Posts Arianna Muti1 and Francesco Fernicola1,2 and Alberto Barrón-Cedeño1 1DIT – Università di Bologna, Forlì, Italy 2EURAC Research, Bolzano, Italy [arianna.muti2 , francesco.fernicola2 , a.barron] @unibo.it Abstract We present our submission to SemEval-2023 Task 10: Explainable Detection of Online Sex- ism (EDOS). We address all three tasks: Task A consists of identifying whether a post is sex- ist. If so, Task B attempts to assign it one of four classes: threats, derogation, animosity, and prejudiced discussions. Task C aims for an even more fine-grained classification, divided among 11 classes. We experiment with fine- tuning of hate-tuned Transformer-based models and priming for generative models. In addition, we explore model-agnostic strategies, such as data augmentation techniques combined with active learning, as well as obfuscation of iden- tity terms. Our official submissions obtain an F1 score of 0.83 for Task A, 0.58 for Task B and 0.32 for Task C. 1 Introduction The shared task on Explainable Detection of On- line Sexism (EDOS) defines sexism as any abuse or negative sentiment directed towards women based on their gender, or based on their gen- der combined with one or more other identity at- tributes (e.g. Black women, Muslim women, Trans women) (Kirk et al., 2023). The EDOS shared task focuses on English posts from Reddit and Gab and proposes three hierarchi- cal sub-tasks. Task A Binary Sexism Detection: systems have to predict whether a post is sexist or not. Task B Category of Sexism: if a post is sexist, systems have to predict one of four mutually ex- clusive categories: (1) threats, (2) derogation, (3) animosity, or (4) prejudiced discussions. Task C Fine-grained Vector of Sexism: if a post is sexist, systems have to predict one among 11 mutually exclusive subcategories, e.g., threats of harm, descriptive attacks (see Table 1, bottom). In this paper, we present our approach to address all three subtasks. For Task A and B, we employ hate-tuned models built upon BERT (Devlin et al., 2019) and RoBERTa (Zhuang et al., 2021). We experiment with two model-agnostic techniques: obfuscation of identity terms and data augmenta- tion with and without active learning. These strate- gies apply only to the model inputs and not to the internal model structure. The first strategy involves masking gender- identifying information such as names and pro- nouns and aims at reducing unintended bias (Nozza et al., 2019; Dixon et al., 2018). The second strat- egy uses feedback from the model to iteratively select new training examples that positively influ- ence the performance on the validation set. For Task C we employ a single hate-tuned model, RoBERTa-hate, and explore the potential of genera- tive models in such a fine-grained text classification setting via in-context learning. 2 Background Sexist language can take many forms, ranging from overtly misogynistic …

Content truncated to 500 words