Detecting Hate Speech in Code‐Mixed Languages: A Robust Study Leveraging Data Augmentation, Transfer Learning, and Model Explainability Techniques in Resource‐Constrained Settings

ABSTRACT Hate speech is a harmful form of expression that promotes discrimination, hostility, and prejudice towards a specific group of individuals or communities. It has become increasingly essential to develop effective methods for detecting hate speech on online platforms to promote inclusivity and protect individuals from the negative effects of such speech. Despite its importance due to the rising use of multilingual social media platforms, hate speech/offensive language (HS/OL) detection in code‐mixed (CM) languages has not gotten the same level of attention from the research community as that of monolingual cases. CM languages pose a challenge due to mixing multiple languages within a single sentence or text, which leads to difficulties in text representation and language modeling. In addition, data imbalance is a common issue in HS/OL detection, as it is often a rare event, and the majority class tends to dominate the dataset. The present study suggests a new method for addressing these challenges by using data augmentation techniques to balance the dataset and leveraging transfer learning to improve model performance. The presented method was assessed in the Malayalam–English Code‐mix language, which poses additional challenges for HS/OL detection due to limited labeled data and lack of resources. The outcomes of the method indicate its effectiveness as the weighted F1 score exceeds 0.98, exceeding the effectiveness of the most advanced and up‐to‐date models. In addition to its effectiveness, the model's explainability is enhanced through a technique showcased in the paper that visualizes each neuron's output in the transfer models' last layer. This technique highlights the disparate firing patterns of the neurons when presented with offensive and non‐offensive inputs.