RECOGNIZING HATE SPEECH IN MULTIPLE MODELS USING ML

Abstract

Malicious and dangerous content is quickly spreading throughout social media platforms, which is a major concern
for modern society. The detection of A number of activities rely on hate speech on sites like Twitter. These include
extracting problematic events, creating AI chatterbots, suggesting material, and analyzing sentiment. With the
proliferation of hate speech and damaging information, researchers have devoted a lot of time and energy to the
difficult problem of recognizing hostile material. Sorting tweets into hateful, offensive, or neutral categories is the
goal. The many expressions of the same idea and the complicated structure of natural language elements make this
job very difficult. Anger may take several forms and target different groups.
Most of the prior work has either used representation-learning methods followed by linear classifiers or depended on
human feature extraction. However, deep learning techniques have lately shown considerable gains in accuracy for
complicated issues in text, vision, and voice applications. This research proposes a method for automatically
categorizing hostile phrases and offensive language using transfer learning models. Make use of Kaggle's
categorized tweet datasets for this study and run experiments. The results show that the multilingual BERT model,
as well as its pre-trained variant, provide better results. In particular, as compared to other algorithms, the pre-
trained BERT model significantly improves the categorization accuracy of abusive tweets by as much as 92%.