Chride at SemEval-2023 Task 10: Fine-tuned Deberta-V3 on Detection of Online Sexism with Hierarchical Loss

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1670–1675 July 13-14, 2023 ©2023 Association for Computational Linguistics Chride at SemEval-2023 Task 10: Fine-tuned Deberta-V3 on Detection of Online Sexism with Hierarchical Loss Letian Peng, and Bosung Kim Computer Science and Engineering University of California, San Diego La Jolla, CA 92093 {lepeng,bosungkim}@ucsd.edu Abstract Sexism is one of the most concerning problems in the internet society. By detecting sexist expressions, we can reduce the offense toward females and provide useful information to understand how sexism occurs. Our work focuses on a newly-published dataset, EDOS, which annotates English sexist expressions from Reddit and categorizes their specific types. Our method is to train a DeBERTa-V3 classifier with all three kinds of labels provided by the dataset, including sexist, category, and granular vectors. Our classifier predicts the probability distribution on vector labels and further applies it to represent category and sexist distributions. Our classifier uses its label and finer-grained labels for each classification to calculate the hierarchical loss for optimiza- tion. Our experiments and analyses show that using a combination of loss with finer-grained labels generally achieves better performance on sexism detection and categorization. Codes for our implementation can be found at https://github.com/KomeijiForce/SemEval2023 _Task10. 1 Introduction The advent of the internet has drastically changed the way we communicate and interact with one another. It has enabled people from different parts of the world to connect and share their thoughts and ideas on various platforms. However, with the increased usage of the internet, there has been a concerning rise in the prevalence of sexism in online communities and platforms. This issue has been especially rampant towards women and other marginalized groups, with the negative sentiment and abuse contributing to the deterioration of the friendly atmosphere that online groups strive to foster. While various censorship systems have been im- plemented to filter out sexist content, they often fail to provide a deeper understanding of the underlying reasons for the sexism. This is where SemEval2023 Shared Task 10 (Kirk et al., 2023) comes in, aim- ing to categorize online comments into fine-grained vectors of sexism. Doing so, it takes the first step towards exploring the root causes of sexism and developing more effective strategies to combat it. The labels used in SemEval2023 Shared Task 10 are annotated hierarchically as sexist, category, and vector, providing a nuanced understanding of the different forms of sexism present in online com- ments. The statistics of the labels are depicted in Figure 1, highlighting the prevalence and diversity of the issue. In our approach, we recognize the hierarchical nature of the problem and propose a method to im- prove expression representations by training a clas- sifier with all three hierarchies of labels. We obtain a probability distribution on vector labels from the classifier’s output and use this to compute category and sexist probabilities. To optimize the model, we calculate the loss of probability distributions in different hierarchies and aggregate them to obtain the final loss. For encoding, we leverage …

Content truncated to 500 words