Mano_sub@DravidianLangTech 2026: Article-Aware Batching and Discriminative Fine-Tuning of MuRIL for Telugu Prompt-Style Classification

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Team Mano_sub's submission to the Telugu Prompt-Style Recovery task at DravidianLangTech 2026 focused on classifying Telugu text into nine stylistic categories: Formal, Informal, Optimistic, Pessimistic, Humorous, Serious, Inspiring, Authoritative, and Persuasive. They identified that standard random batching led to poor within-batch diversity due to 384 unique source articles appearing approximately 7.8 times each, causing majority-class collapse and a macro F1 of 0.022. Their proposed solution involved an article-aware batch sampler that enforces within-batch article diversity, combined with discriminative learning rates for full MuRIL fine-tuning. This approach yielded a mean macro F1 of 0.3834 (std=0.0189) on the development set. The fold 1 best model achieved macro F1=0.2765 on the official test set, a 5.6x improvement over their submitted 0.0491, which would have ranked 2nd among 13 teams.

Key takeaway

If you are an NLP Engineer working on text classification with datasets containing repeated source material, standard random batching can severely hinder model performance. You should investigate your dataset's structure for hidden biases and consider implementing custom batching strategies, such as article-aware sampling, combined with discriminative learning rates, to achieve substantial performance gains and avoid majority-class collapse.

Key insights

Article-aware batching and discriminative fine-tuning significantly improve Telugu prompt-style classification with MuRIL.

Principles

• Address dataset structural properties.
• Batching diversity impacts F1 scores.
• Discriminative learning rates enhance fine-tuning.

Method

Implement an article-aware batch sampler to ensure within-batch diversity, then apply discriminative learning rates during full MuRIL fine-tuning for prompt-style classification.

In practice

• Analyze dataset for structural biases.
• Customize batching for data diversity.
• Apply MuRIL for Dravidian language tasks.

Topics

• Telugu Language Processing
• Prompt-Style Classification
• MuRIL Fine-Tuning
• Article-Aware Batching
• Discriminative Learning Rates
• DravidianLangTech 2026

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• The Significance of Style Diversity in Annotation-Free Synthetic Data Generation — Annotation-free synthetic data generation for intent classification significantly improves utility by prioritizing style diversity during generation over post-hoc adaptation.
• Mini-Batch Class Composition Bias in Link Prediction — Batch normalization in GNNs for link prediction can create a mini-batch composition bias, hindering learning of true graph properties.
• Phonological Assimilation-Aware Neural Segmentation of Sanskrit Compounds — Neural segmentation of Sanskrit compounds is enhanced by explicitly modeling phonological assimilation.
• UrduMMLU: A Massive Multitask Benchmark for Urdu Language Understanding — UrduMMLU reveals current LLMs have uneven Urdu knowledge, particularly for regionally grounded content, despite strong performance from top models.
• Continual Adaptation for Pacific Indigenous Speech Recognition — Adapting speech foundation models to distant, low-resource languages causes severe representational drift and catastrophic forgetting.
• Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers — Finetuning multimodal embedding models on domain-specific data significantly boosts performance over general-purpose models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.