Evaluating LLMs for Named Entity Recognition in Scientific Domain with Fine-Tuning and Few-Shot Learning

Entity extraction is a crucial step in constructing Knowledge Graphs (KGs) from natural language text. In the scientific domain, Named Entity Recognition (NER) is widely used to analyze research papers and facilitate the generation of knowledge graphs that capture research concepts. Given the vast scale of contemporary research output, this task necessitates automated pipelines to maintain efficiency while ensuring the quality of the extracted knowledge. Large Language Models (LLMs) present a promising solution to this challenge. As such, this paper explores the effectiveness of LLMs for NER in scientific texts, using the SciERC dataset as a benchmark. Specifically, it evaluates different LLM architectures, including encoder-only, decoder-only, and encoder-decoder models, to identify the most effective approach for NER in the computer science domain. By examining the strengths and limitations of each model type, this study aims to provide deeper insights into the applicability of LLMs for entity extraction, ultimately improving the construction of domain-specific KGs.