Explainable AI for Reliable Detection of Cyberbullying

The advent of Internet has brought a pivotal revolution in communication technology. This revolution is observed through tremendous use of devices like cell phones and social networking websites that provide a platform for people to interact virtually. The use of social networking websites has provided economical, educational and societal benefits. The ease of access to information and the liberty to openly share and publish information on social networking sites also has a downside. Past decade has witnessed an upsurge of objectionable and cyberbullying content shared on social media. The massiveness and the sensitive bullying content shared on social media platforms make manual method of detecting such content practically challenging. Automated approaches like Artificial Intelligence (AI) based techniques are widely being adopted to detect and moderate online cyberbullying content. Natural Language Processing (NLP) and Deep Learning (DL) are at forefront in automating the detection of cyberbullying content. Inspite of the wide adoption of DL models for detection of bullying content, the decision and predictions made by these models are difficult to comprehend. Explainable AI (XAI) is a promising field that provide interpretations to the decision made by DL models. Local Interpretable Model-Agnostic Explanations (LIME) XAI technique provide better explanations by highlighting the most pertinent features that contributed to model's decision. This paper proposes a unified BiLSTM-LIME model for multiclass classification of cyberbullying content on Twitter platform.