People around the world use social media to express and discuss their views and ideas on various situations, as well as socio-cultural and political affairs. Freedom of speech allows individuals to use social media platforms to create and spread hatred among people. Given its societal impact, identifying such posts on social media is crucial. A notable amount of research exists in this field, particularly in the English language; however, this needs to be extended to Indian languages. The proposed system attempts to identify hate speech in Dravidian language social media posts, specifically Tamil, using a custom dataset annotated based on the severity of hate speech in line with Article 19 of the Universal Declaration of Human Rights (UDHR). Due to the limited availability of resources, the proposed system adopts a few-shot learning approach for detecting hate speech. It employs a 4-way k-shot approach to classify Tamil texts into four class labels: ‘Lawful’, ‘MaybeRestricted’, ‘MustbeRestricted’, and ‘NotHate’. Different sentence embeddings, such as Bidirectional Encoder Representations from Transformers (BERT) -base-multilingual, distiluse-base-multilingual, and language-agnostic embeddings, are used to capture contextual information during the hate speech detection process. The performance of the 4-way k-shot learning was examined for various values of k, and the analysis showed that the 4-way 80-shot learning with language-agnostic embeddings provided the best performance. The proposed system was compared with traditional and transformer-based models to evaluate its performance. Analysis of the effect of data augmentation on the performance of the 4-way 80-shot learning model using language-agnostic sentence embeddings showed a performance improvement, achieving an accuracy of 71. 4% and a macro F1-score of 0. 572. The proposed model also outperformed the existing methodologies for hate speech detection. All resources are made available through the GitHub repository (https: //github. com/JerinMahibha/HateSpeech-Severity DOI: 10. 5281/zenodo. 17116472).
Mahibha et al. (Tue,) studied this question.