Hate speech detection and gender bias mitigation on online social media

Abstract

The popularisation of online social media has allowed the quick proliferation of user- generated content. The large amount of content generated every second on social media plat- forms makes the proper moderation of its content arduous and time-consumed, resulting in an easy dissemination of hate speech. Even though significant advances have been made for au- tomatic hate speech detection, concerns have been raised about the robustness of the learning model and its impact due to its potentially biased behaviours, leading to questionable trends based on identity terms (e.g., women, black, or gay). In this thesis, we address unintended bias, specifically unintended gender bias, in the hate speech detection task. Firstly, we pro- posed a comprehensive study of hate speech, including a critical analysis of definitions of hate speech proposed across multiple platforms and in the scientific community. It also overviews the main approaches typically used in automatic hate speech detection. The results presented a critical analysis of theoretical and practical resources, discussing opportunities in this area and several challenges, including bias issues. Considering the unintended bias in the model for automatically detecting hate speech is essential to prevent potential unintended discrimination. Therefore, we proposed a new methodology using a multi-view ensemble for automatic hate speech detection and unintended gender bias mitigation. The proposed methodology consists of two modules: (1) a gender bias mitigation module based on the detection and replacement of bias-sensitive words and (2) a hate speech detection module using a multi-view stacked clas- sifier. The multi-view stacked classifier combines base classifiers trained with distinct feature representations. Experimental results over four benchmark datasets demonstrate the proposed approach’s effectiveness compared to state-of-the-art solutions, reducing the unintended bias without compromising the model performance. Furthermore, there are concerns whether un- intended bias may presents different behaviours depend on the feature extraction technique used. Therefore, we also proposed a framework to help analyse the biased behaviour of feature extraction techniques. In addition, a new comprehensive dataset to help the unintended gender bias evaluation is designed, called the Unbiased dataset. We have conducted an experimental study on various state-of-the-art feature extraction methods, focusing on their potential bias towards identity terms. Our findings indicate that the feature extraction technique can influ- ence the bias found in the final model, and its effectiveness can rely on the dataset analysed.