The Indian constitution categorizes its population into groups such as Scheduled Castes, Scheduled Tribes, Other Backward Classes, and Forward Castes, reflecting historical inequalities that influence social dynamics and discrimination. Migrants who are relocating within the country for better opportunities are often viewed as outsiders, leading to concerns about job security and crime, fostering hate and discrimination against them. Social media has exacerbated these issues, becoming hotspots for caste and migration-related hate speech, especially in low-resource languages. This study introduces a novel dataset specifically curated to detect hate speech related to caste and migration in the low-resourced Tamil language. Using this dataset, we benchmarked the dataset with baseline experiments with the highest macro-F1 score of 0.73. We also created custom-modified models integrating custom loss functions, adapter-based fine-tuning, and parameter-efficient fine-tuning techniques. To support further research, we released the dataset, conducted a shared task, and ranked participant systems. By publicly releasing this dataset, we aimed to facilitate further study and improve the detection and mitigation of hateful content related to caste and migration on social media.
The K-means clustering technique is widely used in many fields, such as anomaly detection, customer segmentation, cyber-physical system, medical diagnoses, sentiment analysis, fraud detection, and other similar tasks. We used this k-means technique in handling imbalanced datasets by preserving minority class structure using the stratified resampling technique. For this experimental study, we used a benchmark dataset from Kaggle. It is a labeled dataset collected from online social media regarding fake news. This proposed model, The Stratified k-means Sampling (SKMS), is compared with Synthetic Minority Oversampling Technique (SMOTE) by empirically experimenting using different machine learning algorithms. Random Forest (RF) algorithm gives significant accuracy, and Support Vector Classification (SVC) produces a better F1-score than other algorithms. The SMOTE technique was compared with the same dataset using these same algorithms. While SKMS seeks to preserve the structure of the minority class, SMOTE aims to diversify the minority class by interpolating between existing samples. Depending on the dataset, one might be more relevant than the other.
12 August 2025 in Social Network Analysis and Mining
The growing prevalence of hate speech online has amplified acts of discrimination against marginalized populations, with the LGBTQIA+ community being particularly affected. In areas where under-resourced languages such as Malayalam are used, the issue grows more complex because of the absence of localized resources. This research offers an in-depth analysis of the production of counter-speech to address transphobia and homophobia in Malayalam. Our work incorporates both native Malayalam script and Malayalam written in Latin script, addressing the diverse linguistic practices of online users in Kerala. This paper introduces a two-stage pipeline to counter such online abuse. The first stage focuses on dataset creation through a human-in-the-loop process, beginning with 100 seed pairs of hate speech and their corresponding counter-speech manually curated. This set is expanded iteratively using language models culminating in 5,000 validated pairs. In the second stage, we propose a method to generate counter speech in Malayalam that leverages the Retrieval-Augmented Generation framework enhanced by REFINE (Retrieval Evaluation via Fluency, Inversion, and NEarness) for knowledge retrieval and constrained decoding. Evaluation metrics for both dataset quality and model performance demonstrate the effectiveness of this approach in producing diverse, fluent, and target-specific counter-speech. This research provides a foundational resource and scalable strategy for countering hate in low-resource regional languages. GitHub Link: https://github.com/Bharathi-AI-for-Social-Good/CN-Malayalam .