LIDA: Lexical-based Imbalanced Data Augmentation for Content Moderation

Data augmentation (DA) has attracted considerable attention as an alternative for collecting more data without additional human annotation efforts, particularly in low-resource, sensitive, and class-imbalanced tasks. However, the majority of current approaches are designed for the general domain with often balanced data, while in specific tasks like content moderation, the data is often with a skewed distribution. The situation is further exacerbated by data sensitivity, making it unlikely or costly to obtain additional human annotations. To fill this research gap, our paper presents a lexical-based imbalanced data augmentation (LIDA) approach for content moderation. LIDA is an easy-to-implement and explainable DA method that utilizes sensitive lexicons and randomly inserts sensitive lexicons into negative samples for converting them into positive ones. In this way, LIDA can achieve a balanced dataset for avoiding skewed distribution problems. We validate our model on two datasets, namely Wiki-TOX and Wiki-ATT, to show the superior performance of our proposed algorithm compared to other rule-based data augmentation baselines, and p-values are presented to demonstrate its effectiveness and stability.