📄 Lannisters@DravidianLangTech 2026: A Comparative and Ablation Study of Multilingual Transformers for Gender-Targeted Abuse Detection in Tamil Social Media Platforms

Summary

The paper addresses automatic detection of abusive comments aimed at women on Tamil social media. Noisy spelling, dialect variation, code-mixing, and sarcastic or indirect phrasing make keyword or rule-based filtering impractical there. Built for the DravidianLangTech 2026 shared task, the system treats the problem as binary classification (Abusive vs. Non-Abusive).

The authors clean the text with a dedicated pipeline. It decodes HTML, normalizes Unicode, strips URLs and mentions, keeps hashtags and emojis, and maps obfuscated slurs to a placeholder. They then enlarge the minority Abusive class with random deletion, word swapping and back-translation via English. DistilBERT, mBERT and XLM-RoBERTa are fine-tuned under identical settings.

XLM-RoBERTa performs best, reaching 91.17% validation accuracy and 0.8865 macro F1. An ablation suggests that preprocessing, augmentation and tuning each improve results. Remaining errors concentrate on sarcasm, implicit abuse and dialectal or Tamil-English mixed comments. All scores come from a local split created after augmentation, and no official test result is reported, so the numbers may be optimistic. The work still offers a reproducible baseline for gender-targeted abuse detection in a low-resource language.

Key points

• Data: 3,652 training and 913 test Tamil comments. Augmenting the Abusive class raised the total to 7,190, split 80:20 into 5,752 training and 1,438 validation instances.

• Validation accuracy by model: DistilBERT 87.4%, mBERT 89.2%, XLM-RoBERTa 91.17%. XLM-RoBERTa also scored weighted F1 0.9119 and macro F1 0.8865.

• XLM-RoBERTa setup:

  • 6 epochs, learning rate 2e-5, batch size 16, weight decay 0.01, label smoothing 0.1.
  • Maximum sequence length 256, dropout 0.1.
  • Class-weighted cross-entropy and early stopping on macro F1.
  • Retrained on all data before final predictions, using a Tesla T4 GPU.

• Ablation (accuracy / macro F1):

  • Base model: 0.78 / 0.78
  • With preprocessing: 0.88 / 0.85
  • Adding augmentation: 0.92 / 0.89
  • Adding tuning: 0.93 / 0.91

  The final 0.91 does not match the 0.8865 reported elsewhere. The per-class F1 values (Abusive 0.94, Non-Abusive 0.83) also look odd given that Non-Abusive is the majority class.

• Code is available at github.com/Nandhinibalu05/SocialMediaTamilAbuse. The authors acknowledge limitations: a small dataset, limited variety in the augmented data, and difficulty with sarcasm and bias.