SCIRE at BioASQ 2025: LLM Driven Biomedical Named Entity Recognition for GutBrainIE 2025

In recent years, we have witnessed the rise of powerful Large Language Models (LLMs) and their flexibility in accomplishing a wide range of NLP tasks, often achieving state-of-the-art (SOTA) accuracy. However for Named Entity Recognition (NER), there is a specific need for token-level class assignments and bidirectional context, both preceding and following a token, to understand its role. As a result, bidirectional encoder-style transformer models (BERT-like models) have been the standard approach. However, fine-tuning these models on available datasets often faces the bottleneck of limited training data. In this paper, we propose an alternative approach that leverages the extensive knowledge base of decoder-style Transformer models. These modern LLMs are typically trained on vast amounts of text, which enables them to overcome the challenge of limited labeled data. Instead of training from scratch, we focus on aligning the responses of these LLMs to suit NER. To this end, we use two methods: (i) few-shot prompting, and (ii) fine-tuning on available examples. Our findings indicate that fine-tuning significantly outperforms prompting for biomedical NER, effectively aligning LLM outputs to the desired task outputs. Additionally, we propose an algorithm to parse the output of the LLM to extract relevant entities, their labels, and their start and end indices.