SUTNLP at SemEval-2023 Task 10: RLAT-Transformer for explainable online sexism detection

Abstract

There is no simple definition of sexism, but it can be described as prejudice, stereotyping, or discrimination, especially against women, based on their gender. In online interactions, sexism is relatively rare but still harmful. One out of ten American adults says that they have been harassed because of their gender and have been the target of sexism, so sexism is a growing issue. The Explainable Detection of Online Sexism shared task in SemEval-2023 aims at building sexism detection systems for the English language. In order to address the problem, we use large language models such as RoBERTa and DeBERTa. In addition, we present a novel method called Random Layer Adversarial Training (RLAT) for transformers which is based on adversarial training, and show its impact on boosting all subtasks’ scores. Moreover, we use other discriminative and generalization techniques for subtask A to boost performance. Using our methods to make predictions over subtask A, B, and C test sets, we obtained macro-F1 of 84.45, 67.78, and 52.52 respectively, outperforming proposed baselines on all subtasks. Our code is publicly available on Github.