CYBERBULLYING DETECTION ON SOCIAL MEDIA USING DEEP LEARNING MODELS

Abstract

Application of deep learning models for cyberbullying detection in social media is an upcoming area for both researchers and practitioners for finding, exploring and analysing the extensibility of human-based expressions. Automated cyberbullying detection is typically a classification problem in natural language processing where the intent is to classify each abusive or offensive comment or post or message or image as either bullying or non-bullying. It needs high-level semantic analysis as well. Most of the earlier attempts on cyberbullying detection rely on manual feature extraction methods. Such methods are not only time-consuming and cumbersome, but often fail to correctly capture the meaning of the sentence. This fosters the need to build an intelligent analytic paradigm for detecting cyberbullying in social media data to lower down its hazard with minimal human intervention. Motivated by it, this research utilizes deep learning models for cyberbullying detection in social media as they trivialize the need of explicit feature extraction and are highly skilful, fast and more efficient in retrieval of essential features and patterns by themselves. In our research, we have applied deep learning for cyberbullying detection on textual and non-textual social media content. With high volume and variety of user-generated content on complex social media platforms, the challenges to detect cyberbullying in real-time have amplified. The influx of content makes it challenging to timely regulate online expression. Moreover, the anonymity and context-independence of expressions in online posts can be ambiguous or misleading. Nowadays, cyberbullying, through varied content modalities is also very common. At the same time, cultural diversities, unconventional use of typographical resources and easy

availability of native-language keyboards augment to the variety and volume of user- generated content compounding the linguistic challenges in detecting online bullying

posts. In an effort to deal with this antagonistic online delinquency referred to as cyberbullying, this research computationally analysed the content, modality and language-use in social media using deep learning models. This research has shown that the use of embeddings with deep learning architectures show better representation learning capabilities and simplify the feature selection process with enhanced classification accuracy as compared to baseline machine learning methods. The goal of the research is to automatically detect cyberbullying on textual, multimodal and mash-up social media content using deep learning models. In our research, we build models for these using deep architectures including capsule network, convolution neural network, multi-layer perceptron, self-attention mechanism, bi-directional gated recurrent unit, long short-term memory & bi-directional long short-term memory using embeddings such as GloVe, fastText and ELMo on social media like Askfm.in, Formspring.me, MySpace, Twitter, YouTube, Instagram and Facebook. The results show superlative performance as compared to SOTA as well.