Social media platforms like X (twitter) have gained popularity for political debates and election campaigns in the last decade. This creates the need to moderate and understand the sentiments of the tweets in order to understand the state of digital campaigns. This paper focuses on political sentiment classification of Tamil X (Twitter) comments which proves to be challenging because of the presence of informal expressions, code-switching, and limited annotated datasets. This study focuses on categorizing them into seven classes: substantiated, sarcastic, opinionated, positive, negative, neutral, and none of the above. This paper proposes a solution to Political Multiclass Sentiment Analysis of Tamil X (Twitter) Comments - DravidianLangTech@NAACL 2025 shared task, the solution incorporates IndicBERTv2-MLM-Back-Translation model and TF-IDF vectors into a custom model. Further we explore the use of preprocessing techniques to enrich hashtags and emojis with their context. Our approach achieved Rank 1 with a macro F1 average of 0.38 in the shared task.

Detecting hate speech targeting caste and migration communities in code-mixed Tamil-English social media content is challenging due to limited resources and socio-cultural complexities. This paper proposes a multi-scale hybrid architecture combining classical and neural representations with hierarchical ensemble learning. We employ advanced preprocessing including transliteration and character repetition removal, then extract features using classical TF-IDF vectors at multiple scales (512, 1024, 2048) processed through linear layers, alongside contextual embeddings from five transformer models-Google BERT, XLM-RoBERTa (Base and Large), SeanBenhur BERT, and IndicBERT. These concatenated representations encode both statistical and contextual information, which are input to multiple ML classification heads (Random Forest, SVM, etc). A three-level hierarchical ensemble strategy combines predictions across classifiers, transformer-TF-IDF combinations, and dimensional scales for enhanced robustness. Our method scored an F1-score of 0.818, ranking 3rd in the LT-EDI-2025 shared task, showing the efficacy of blending classical and neural methods with multi-level ensemble learning for hate speech detection in low-resource languages.