Qmul-sds at exist: Leveraging pre-trained semantics and lexical features for multilingual sexism detection in social networks

Abstract

Online sexism is an increasing concern for those who experi- ence gender-based abuse in social media platforms as it has affected the healthy development of the Internet with negative impacts in society. The EXIST shared task proposes the first task on sEXism Identifica- tion in Social neTworks (EXIST) at IberLEF 2021 [30]. It provides a benchmark sexism dataset with Twitter and Gab posts in both English and Spanish, along with a task articulated in two subtasks consisting in sexism detection at different levels of granularity: Subtask 1 Sexism Iden- tification is a classical binary classification task to determine whether a given text is sexist or not, while Subtask 2 Sexism Categorisation is a finer-grained classification task focused on distinguishing different types of sexism. In this paper, we describe the participation of the QMUL-SDS team in EXIST. We propose an architecture made of the last 4 hidden states of XLM-RoBERTa and a TextCNN with 3 kernels. Our model also exploits lexical features relying on the use of new and existing lexicons of abusive words, with a special focus on sexist slurs and abusive words targeting women. Our team ranked 11th in Subtask 1 and 4th in Sub- task 2 among all the teams on the leaderboard, clearly outperforming the baselines offered by EXIST.