Detection of Bodo Toxic Words using Recurrent Neural Network

Abstract

The detection of toxic words particularly in low resource support language like Bodo, characterize specific challenges in natural language processing (NLP). This study explores the application of Recurrent Neural Networks (RNN) which is a deep learning approach, for detecting toxic words in Bodo language. By the sequential nature of RNN, the model captures contextual dependencies that are crucial for understanding the nuances of a language. The application of deep learning algorithms in NLP has revolutionized the field, enabling significant advancements in understanding and generating human languages. In this paper, an imbalanced dataset is curated with 11465 instances collected from networking sites, dictionaries, comprising various forms of toxic expressions in Bodo language, annotated for training purposes. The model is trained to classify toxic or non-toxic words only. Moreover, performance of the model was evaluated using the metrics precision, recall, accuracy and F1-score. The results shows classifying toxicity with test loss of 0.0156 and test accuracy of 0.9947. The performance demonstrated precision of 0.8679, recall of 0.7731, accuracy of 0.5070 and F1-score of 0.8178. The findings highlight the model's effectiveness in identifying toxic words, contributing to the wide content of raising positive communication. The work emphasize the importance of NLP solutions for low resource language that illuminates the way for future studies in multi-lingual and multi-class toxicity identification.