Enhancing online toxicity detection on gaming networks: a novel embeddings-based valence lexicon approach

Abstract

Online toxicity and violent speech on gaming networks pose significant threats to societal well-being, particularly among adolescents, and are linked to severe consequences such as suicide. This highlights an urgent need for effective toxicity detection methods tailored to these platforms. Traditional rule-based approaches are inherently limited, and the performance of predictive models in detecting online toxicity is critically dependent on the quality and representativeness of their training data. However, the distinct linguistic characteristics of discourse on gaming networks present unique challenges in curating representative training samples using existing valence lexicons, often resulting in suboptimal detection accuracy. In this study, we propose a novel framework for online toxicity detection on gaming platforms that addresses these linguistic challenges. Our approach introduces an extended embeddings-based valence lexicon that can be customized to any gaming platform. In this study, we specifically target Twitch. Through a comprehensive comparative evaluation against state-of-the-art lexicon-based techniques, the proposed method demonstrates superior performance. Notably, agreement analysis reveals a moderate alignment between human annotators and the proposed method, with a kappa score of 0.619. Experimental results further underscore the efficacy of our approach, showing an average performance improvement of 31.47% across LSTM, GRU, and CNN architectures compared to all baseline methods. The findings of this study have significant implications for improving platform-specific toxicity detection, enhancing the safety and inclusivity of gaming environments. The framework can be adapted to other online platforms, offering a scalable solution to address toxicity and protect vulnerable populations.

Keywords Deep learning, Implicit valence, Online toxicity detection, Toxicity prediction, Word embeddings