Multi Model Hybrid Approach for Youtube Hate Speech Detection

Abstract

"Hate Speech has been an issue since the dawn of human communication. Hate speech can be distinguished in multiple ways according to race, gender, religion, age and many ways that divide people. Ever since the internet saw light, The users have the function to express their ideas and views openly to the public without considering offensive language and hate speech. There are multiple attempts and approaches done to this domain to limit and prevent public hate speech. Social media plays the biggest role in this sector. In that YouTube is a video posting, viewing platform with free speech limited by YouTube policies which sometimes become ineffective. Therefore a system must be set in place with high accuracy using natural language processing (NLP) algorithms. Multiple approaches were taken to solve this arising issue by implementing single model approaches, hybrid model approaches and ensemble model approaches. These systems were done by using various algorithms such as Naive Bayes, Bidirectional Encoder Representation for Transformers (BERT), Long Short Term Memory (LSTM) and many more. Mostly single model implementations were conducted by many authors for detecting hate speech. Therefore presented in this research paper, the author proposes a development of a hate speech detection system for YouTube videos. The author provides a chrome extension and a single page web application integrated with the ensemble modeling architecture. This gives the user a user friendly interface and an easy to use architecture. A kaggle labeled dataset of hate speech was used to train the BILSTM and BERT models. Both these models produced stable and robust implementations. BILSTM performs by two unidirectional LSTM layers. Both these layers return a probability vector. The BILSTM model performs with an average of 98% accuracy. The BERT model performs with an average of 92% accuracy."