Enhancing Aggression Detection using GPT-2 based Data Balancing Technique

In the past few years, the use of Social Media platforms has seen a significant increase in the number of users and content. Owing to the increased usage of such platforms, instances of cyberbullying, trolling, hate-speech, negative provocation, community attacks, etc. has also increased tremendously. There is an urgent need to automatically identify such contents (posts/tweets) which can hamper the well-being of an individual or society in general. In this paper, different techniques are compared to classify posts /tweets of Hindi-English Code-Mixed origin into three categories: “Overly Aggressive”, “Covertly Aggressive”, and “Not Aggressive”. This paper has proposed a novel data balancing technique for text classification. A data balancing technique is proposed for analyzing textual data using Generative Pretrained Transformer (GPT-2) owing to its contextual understanding and more realistic data generation capability. The effect of balancing the data with GPT-2 is evaluated on six different learners. An ensemble of all the Six learners is also developed to achieve an overall improvement in the performance for aggression classification. Comparative analysis of different Machine learning and Deep learning models are performed with and without data balancing. The proposed ensemble has outperformed and achieved 0.65 F1 score.