Exploring LLM-based Approaches for Open-ended Named Entity Recognition in Low-resource Domains

December 12, 2025

This work explores how Large Language Models (LLMs) can perform Named Entity Recognition (NER) in low-resource and domain-specific scenarios, focusing on the legal domain and Portuguese language. We present a modular framework that separates and evaluates the effect of entity information, in-context demonstrations, reasoning instructions, and error reflection, across 10 datasets. We conduct systematic experiments with all 10 datasets, demonstrating that smaller, cost-efficient models like GPT-4o mini can achieve competitive results when paired with optimized prompting and retrieval. Our final configuration, using representative central example spans and 10% of the training database as in-context demonstrations, retrieved via cosine similarity, achieves an average micro-F1 of 73.5%, showcasing that our refined prompting approaches and retrieval can offset the need for fine-tuning in resource-constrained NER settings.

Authors in the community:

• 

  M

  Miguel Angelo Medeiros Rebocho

  ist180869

Supervisors of this institution:

• 

  B

  Bruno Emanuel Da Graça Martins

  ist24686

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico