Transfer learning approach for identifying negative sentiment in tweets directed to football players

Abstract

In recent times, the rising cases of racial abuse toward football players are gravely concerning. This trend is becoming a recurrent decimal on social networks with little or no proactive measures to serve as a deterrent to others in most cases. There was a public outcry due to massive racial abuse remarks targeting some England footballers during the Euro2020 final between England and Italy on social media. This motivated us to investigate and recommend a better solution to the social pandemic through the application of transfer learning. This will go a long way to maintain unity in diversity in the game of football. Tweets with related hashtags to the football match were scraped. The researchers framed this problem as a multi-class classification task. The Valence Aware Dictionary and sEntiment Reasoner (VADER) compound score was employed to label the dataset as positive (POS), negative (NEG), and neutral (NEU). Seven pre-trained models were fine-tuned and corresponding models were built for identifying negative sentiments toward the football players. Three contributions were made – a review of literature on racial abuse targeting footballers, a new dataset to further research in this direction and a robust transformer-based model identified for policing Twitter during football games. Among the models built, the distilled version of Bidirectional Encoder Representations from Transformer (DistilBERT) proved superior with an F1-score of 0.99.