Hate Speech and Offensive Content Identification in Hindi and Marathi Language Tweets using Ensemble Techniques

Abstract

Hate Speech is described as any form of speech in which speakers attempt to ridicule, humiliate, or inculcate hatred in someone else’s minds based on characteristics such as religion, the colour of skin, race, or sexual preference. In recent years, social networking sites have been a major source of excessive amounts of hate speech. If unaddressed, these might cause anxiety and despair in the affected individuals or groups. As a result, the above-mentioned social networks utilize an assortment of algorithms to identify such hate speech. Detecting Hate Speech in English texts has been one of the hottest topics in recent years, with multiple types of research being published. However, in regional and indigenous languages, hate speech detection is a recent area with not much research being conducted. It is difficult to perform hate speech detection using data in regional languages due to a lack of large enough training data and a lack of resources about that domain. The HASOC [1] 2021 Hate Speech Detection Task solves one of the problems. It provides a dataset containing Tweet data in English, Hindi [2] and Marathi [3] languages. There were two subtasks as part of the main task. The subtask was to classify the hate speech and offensive texts in the Hindi and Marathi tweet dataset as Hate Speech (HATE), Offensive (OFFN) or Profane (PRF). This work compares the performance of different models on both subtasks and provides a conclusion on the best performing model. The Random Forest Classifier reports the most remarkable accuracy on the first subtask with a macro F1 score of 75.19% and 73.12% on the Marathi and Hindi tweet datasets. The XGBoost algorithm is the best performing algorithm on the second subtask with a 46.5% macro F1 score. Overall any of these models can get satisfactory results when dealing with hate speech detection in regional language. This work has been submitted to the FIRE2021 shared task, Hate Speech and Offensive Content Identification in English and Indo-Aryan Languages (HASOC-2021) by team DLRG. © 2021 Copyright for this paper by its authors.