Amharic Language Hate Speech Detection Using Machine Learning

The extensive availability of social media platforms, as well as the adaptability of the Internet, has made it easier for users to participate in violent communication. The anonymity provided by online platforms makes them appealing to individuals engaging in hate speech to conceal their criminal activities. This poses a significant challenge in many countries, especially Ethiopia. As social media platforms continue to multiply and the volume of social media data grows exponentially, the identification of hate speech presents a formidable challenge. This challenge exacerbates conflicts between diverse ethnic groups and contributes to the dissemination of misinformation within communities. Despite the fact that there is a lot of research on hate speech identification, much of it is for high-resource languages, and there is still a lot of work to be done for low-resource languages. These pique our interest. As a response to these issues, this study uses machine learning methods to create an Amharic hate speech detection model. In this research, we prepared a new Amharic hate speech dataset from Facebook, labeled manually as hate and free based on standard guidelines and pre-processed, and labeled into two classes, and the data is augmented to balance the class category. We applied word2Vec embedding and TF-IDF feature selection techniques to train Random Forest, and Naïve Bayes machine learning models. To evaluate the models, we adopted an (80, 10, 10) train, validate, and test split. We utilized precision, recall, and F1-score as a means to compare the performance of these models. By combining the Naïve Bayes algorithm with Word2Vec and TF-IDF techniques, the best performance was achieved, resulting in an accuracy of 91.59%. The model achieves a promising result with a unique feature selection and appropriate pre-processing techniques.