Automated Identification of Offensive Language in Code-Mixed Social Media Content using a Hybrid VNN-BERT Model

The increased influence of social media platforms has transformed the way of communication and interaction with people globally. These platforms allow people to express their emotions, opinions, criticism, etc. The ability to connect with people easily is also said to hurt people's lives. Social media also covers the use of harsh words and hate speech which can have serious negative consequences, ranging from psychological harm to inciting violence and discrimination. Hence an automated approach is needed to detect offensive speech. Several deep learning and Natural Language Processing (NLP) techniques have been investigated in previous research to identify objectionable language in multilingual languages, and the results showed promising results. In order to find a single pre-trained model that is efficient for detecting hate speech and offensive content (HOS), the paper investigates multilingual transformer-based embedding models utilizing BERT. The classification of abusive speech is done using a hybrid model that combines the Vanilla Neural Network (VNN) with Bidirectional Encoder Representations from Transformers (BERT). The research develops effective machine-learning models to identify offensive language and hate speech in Tamil with an F1 score of 0.89.