Offensive Language Classification of Code-Mixed Tamil with Keras

Abstract

This paper presents the method adopted for completing Task 1 of Dravidian-CodeMix-HASOC (Hate Speech and Offensive Content Identification in English and Indo-European Languages) Shared Task proposed by the Forum of Information Retrieval Evaluation in 2021, for offensive language detection. For detecting offensive language, a custom model architecture using convolutional neural networks was created using Keras for supervised learning, and trained on a dataset of YouTube comments, written in code-mixed Tamil in both Roman and Tamil scripts. The 5 layer neural network was built only using Keras, and required simple tokenized data, padded to an appropriate length. Recurrent neural networks and transfer learning were not used, and an F-score of 0.835 was achieved with the created CNN model.