Comparative Analysis of Deep Learning Models for Offensive Comment Detection in Code-Mixed Dravidian Languages, with a Focus on Tamil-English

Abstract: Offensive and hate speech detection on social media has become an important natural language processing problem because of the scale at which harmful user-generated content is produced and the practical impossibility of moderating it manually. The problem is considerably harder for code-mixed Tamil-English text, where Tamil words are frequently transliterated into Roman script, spelling is inconsistent, and sentence structure borrows freely from both languages, all of which limit the effectiveness of models built primarily for monolingual, formally written text. This paper presents a comparative review of fifteen system-description papers submitted to the DravidianLangTech- EACL2021 shared task on offensive language identification in Tamil, Kannada and Malayalam code-mixed text, evaluated on the Dravidian Offensive Language Dataset released for that task, with the base paper SJ_AJ@DravidianLangTech-EACL2021 used as the primary point of reference. The fifteen studies are compared on dataset, preprocessing, feature representation, model architecture, training strategy and reported evaluation metric, rather than evaluated in isolation. The review finds that classical machine learning with TF-IDF or character n-gram features, non-transformer deep learning architectures such as BiLSTM and CNN-BiLSTM hybrids, and pretrained multilingual transformers such as mBERT and XLM-RoBERTa, often strengthened with ensembling, task-adaptive pretraining, transliteration or class-imbalance handling, together form the dominant methodological landscape for this task. Reported weighted F1-scores vary considerably across the three languages and techniques, with a severe class imbalance in the underlying dataset (over 70% of Tamil-English comments labelled Not-offensive) repeatedly identified as a limiting factor. The review further finds that interpretability of model predictions is addressed only marginally, mostly through confusion-matrix error analysis, and that no reviewed study applies a dedicated explainability technique. This observation motivates the proposed direction of incorporating explainable deep learning methods into Tamil-English code-mixed hate speech detection, so that model decisions can be examined for correctness as well as for the reasoning behind them.

Keywords: Hate Speech Detection; Code-Mixed Text; Tamil-English; Offensive Language Identification; Transformer Models; Explainable AI; Comparative Study