Hate Speech Detection in Sinhala and Romanized Sinhala Mixed with Emojis using Hybrid CNN-RNN Model

Abstract

ABSTRACT The paper addresses the critical issue of the hate speech detection in online Sinhala and Romanized sinhala content associated with emojis. The widespread of online hateful content is particularly significant in low-resource languages such as Sinhala, where natural language processing tools and resources are limited. The study examined the development and evaluation of a novel deep learning approach to effectively identify the words of hatred in the Sinhala language and the Romanized Sinhala text, and complicated the inclusion of emojis, which are prevalent in the online communications of the Sinhala language. In order to address this challenge, we propose a combined model of the Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM) as our primary architecture. This model is designed to use the strengths of both CNNs to capture local n-gram patterns and BiLSTMs to understand the sequential context within the Sinhala text. CNN extracts local signs of hate speech, while BiLSTM captures long-term dependencies and contextual information in Sinhala comments. An Embedding layer, shared by both branches, learns dense vector representations for Sinhala words and emojis optimized for the task of detecting offensive speeches. In order to compare, we also implemented and evaluated independent CNN, BiLSTM, LSTM, Support Vector Machine (SVM), Random Forest, Logistic Regression, and Naive Bayes models. The results show that our CNN-BiLSTM combined model is superior to the comparable models. The CNN-BiLSTM model achieved a high accuracy of 0.9251 and a balanced F1 score of 0.92 for the “Offensive” class, indicating the robustness of detecting hatred speech while minimizing both positive and negative false information. The analysis of the confusion matrix further revealed a balanced distribution of errors, highlighting the effectiveness of the model in the treatment of offensive and non-offensive Sinhala content. Although simple models such as SVM and Random Forest showed competitive accuracy, their precision and F1 points were generally lower, especially in the “Offensive” class. Neural network architectures, in particular CNN-BiLSTM and CNN, have consistently surpassed traditional machine learning methods. Keywords: Hate Speech Detection, Sinhala Language, Low-Resource Language, CNNBiLSTM, Deep Learning, Natural Language Processing, Romanized Sinhala, Emojis