PeTailor: Improving Large Language Model by Tailored Chunk Scorer in Biomedical Triple Extraction

Biomedical triple extraction systems aim to automatically extract biomedical entities and relations between entities. While current unified information extraction models showcase state-of-the-art performance, they face challenges in understanding relationships between entities within intricate biomedical sentences. Furthermore, the absence of a high-quality biomedical triple extraction dataset impedes the progress in developing robust triple extraction systems. To tackle these challenges, we propose a novel retrieval-based framework for biomedical triple extraction, namely PeTailor, which explicitly retrieves the relevant document from our pre-built diverse chunk database using a novel tailored chunk scorer and integrates the retrieved information into the input of a Large Language Model (LLM) to generate the corresponding triple (head entity, relation, tail entity) for the input sentence. Additionally, we present GM-CIHT, an expert-annotated biomedical triple extraction dataset that covers a wider range of relation types. Experimental results show that our proposed PeTailor method achieves state-of-the-art performance on GM-CIHT and two standard biomedical triple extraction datasets