FiRC at SemEval-2023 Task 10: Fine-grained Classification of Online Sexism Content Using DeBERTa

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1824–1832 July 13-14, 2023 ©2023 Association for Computational Linguistics FiRC at SemEval-2023 Task 10: Fine-grained Classification of Online Sexism Content using DeBERTa Fadi Hassan† Abdessalam Bouchekif† Huawei Technologies Oy (Finland) Co. Ltd., Finland firstname.lastname@huawei.com Walid Aransa† Abstract SemEval 2023 shared task 10 “Explainable Detection of Online Sexism” focuses on de- tecting and identifying comments and tweets containing sexist expressions and also explai- ning why it is sexist. This paper describes the system that we used to participate in this sha- red task. Our model is an ensemble of different variants of fine-tuned DeBERTa models that employs a k-fold cross-validation. We have participated in the three tasks A, B and C. Our model ranked 2nd position in tasks A, 7th in task B and 4th in task C. Index Terms : Hate Speech, DeBERTa, En- semble Models, Cross-validation. 1 Introduction Social media platforms have been facing an in- crease in the number of hate speech targeting an individual or group, most often on the ground of religion, race, sexual orientation and gender, which makes these platforms hostile and toxic. Sexism is a form of discrimination that can be di- rected toward women. Sexism identification has been gaining attention recently among NLP com- munity, mainly through shared tasks. The EXIST 2021 challenge (Rodr´ıguez- Sanchez et al., 2022) is the first shared task aimed to attract and motivate research efforts in building sexism detection models for English and Spanish languages. The proposed systems in EXIST 2021 challenge were mainly based on transformers (monolingual or multilingual). Hence, the best model in sub-task 1 (for Spanish language) (Butt et al., 2021) is obtained by applying specific pre-processing techniques and by using BERT model with data augmentation. The new training data was created by translating the comments into German language and then back-translated into Spanish and English. In (del Arco et al., 2021), †. These authors contributed equally to this work the authors explore how transferred knowledge from tasks related to toxicity language may help in sexism tasks. Therefore, the proposed system follows a multitask learning approach where multiple tasks related to toxic comment identification are learned in parallel while using a shared representation. The model achieves convincing performance in both subtasks, 1st place in sub-task 1 English and 2nd place in sub-task 2 Spanish. The second edition of EXIST was organized in 2022 (Rodr´ıguez-Sanchez et al., 2022). Similar to the first edition, EXIST 2022 focused on “Sexism Identification and Sexism Categorization”, both in Spanish and English. EXIST 2021 data sets are used as training data and a new data set consisting of 1058 tweets is used as a test set. Unsurpri- singly, all participating teams used some sort of transformer architecture : Bert (Devlin et al., 2018), RoBERTa (Liu et al., 2019), DeBERTa (He et al., 2021) and BETO * (Ca˜nete et al., 2020). (Vaca-Serrano, 2022) obtained powerful models for both tasks using different transformer- based language models with RoBERTa-large, BERTweet-large (Nguyen et …

Content truncated to 500 words