BERT-based Models vs. Large Language Models for Low-Resource Named Entity Recognition: A Comparative Study on Marathi

2026-07-25 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

A comparative study on Marathi Named Entity Recognition (NER) evaluated the performance of fine-tuned MahaBERT-v2 models against an existing MahaNER baseline and prominent general-purpose Large Language Models (LLMs) including Gemini, LLaMA-3.3-70B, and Gemma. Focusing on low-resource language processing, the research fine-tuned MahaBERT-v2 on various MahaNER dataset variants. Experimental results, measured by precision, recall, and F1-score, consistently showed that the fine-tuned MahaBERT-based models achieved superior performance. These specialized models recorded F1-scores ranging from 0.88 to 0.91, surpassing the existing MahaNER model's 0.8843 and significantly outperforming LLM-based approaches, which yielded F1-scores between 0.57 and 0.69. This highlights the continued importance of specialized architectures and domain-relevant data for effective low-resource language NER.

Key takeaway

For NLP Engineers developing Named Entity Recognition solutions for low-resource languages like Marathi, you should prioritize fine-tuning specialized BERT-based models over deploying general-purpose Large Language Models. Your efforts in acquiring and annotating domain-relevant data for these task-specific architectures will yield significantly higher F1-scores, as general LLMs currently underperform in such language-specific, low-resource contexts.

Key insights

Task-specific, language-focused models outperform general-purpose LLMs for low-resource language NER.

Principles

• Specialized architectures are crucial for low-resource language processing.
• Domain-relevant data enhances language-specific model effectiveness.

Method

Fine-tune BERT-based models on specific language datasets and compare against general LLMs using F1-score, precision, and recall.

In practice

• Deploy fine-tuned BERT models for low-resource NER tasks.
• Invest in language-specific data annotation for better performance.

Topics

• Named Entity Recognition
• Low-Resource Languages
• BERT-based Models
• Large Language Models
• Marathi NLP
• Model Fine-tuning

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Translation as a Computationally Efficient Bridge: Feasibility of English BERT for Low-Resource Languages — Translating low-resource language data to English for fine-tuning existing English BERT models offers a computationally efficient and often effective NLP solution.
• Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats — Lightweight LLMs offer competitive biomedical NER performance, with specific output formats proving more effective than multi-format tuning.
• From BERT to T5: A Study of Named Entity Recognition — BERT and T5 models were fine-tuned and compared for Named Entity Recognition using different architectural approaches.
• The Future of Indian Language Technologies: Natural Language Processing for Low-Resource Languages — NLP for low-resource Indian languages is crucial for digital inclusion, overcoming data scarcity through advanced models and collaborative efforts.
• My Deep Dive into Large Language Models: An Architectural Journey — The shift to LLMs redefines NLP through scalable Transformer architectures, enabling versatile language tasks and requiring advanced customization.
• NASIMLab at SemEval-2026 Task 9: A Comparative Analysis of Fine-Tuned Small Language Models vs. Generative Large Language Models for Multilingual Polarization Type Detection — Fine-tuned small encoder-only models consistently outperform generative large language models for multilingual polarization detection, especially in low-resource languages.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.