Multiclass Classification of Hindi-English Code Mixed Misogyny Comments Using Recurrent Neural Networks

Users can communicate their ideas, feelings and opinions on social networking connections. However, they are increasingly being exploited for the dissemination of harmful and abusive content. Addressing this issue, The goal of this work is to fill the research void, a meticulously curated dataset in code-mixed Hindi-English has been introduced. This dataset consists of 12,698 YouTube comments and replies, which are systematically categorized at two levels. At the first level, comments are classified as either optimistic or pessimistic, while at the second level, they are further categorized based on their content. To facilitate the automatic detection of misogynistic sentiment within these comments, various computational approaches are explored. Among the methods evaluated, the Bidirectional Gated Recurrent Unit (BiGRU) model with FastText embeddings demonstrates the most effective performance across both classification tasks. The model achieves accuracy of 63.71% with subtask 1 and 64.78% with subtask 2 respectively. These results highlight the efficacy of deep learning methods in analyzing and identifying misogynistic sentiment within code-mixed social media discourse. This work contributes to the ongoing efforts in automatic hate speech detection and presents a valuable dataset for further advancements in multilingual and sentiment analysis research.