Multi-Task Learning for Toxic Comment Classification and Rationale Extraction

This report explores the growing importance of content moderation on social media platforms and its role in fostering constructive discussions. It introduces toxic span prediction as a key technique for identifying and labeling toxic comments, which is essential for developing automated moderation systems. The strong correlation between toxic comment classification and toxic span prediction highlights the value of a joint learning approach. To address this, the report proposes a multi-task learning model that utilizes BERT transformers to enhance the contextual understanding of toxic comments. Additionally, it integrates a Bi-LSTM CRF layer to precisely identify toxic spans or rationales within the comments. This combined methodology aims to improve the efficiency of content moderation systems. To support multi-task learning in this area, a dataset is curated from the Jigsaw and Toxic Span Prediction datasets, serving as the basis for training and evaluation. The proposed model demonstrates significant performance improvements in both toxic comment classification and span identification compared to single-task models. In practical applications, social media platforms can leverage this multi-task learning model to gain deeper insights into toxic content. By incorporating BERT transformers and the Bi-LSTM CRF layer, these platforms can enhance their ability to detect harmful content with greater accuracy and efficiency, leading to a safer online environment.