Retrieval Augmented Document-Level Information Extraction from Scientific Articles Using Large Language Models

Identifying the main results of scientific papers is an important and challenging task. Unlike traditional named entity recognition (NER) in general domains, NER in the scientific field primarily aims to identify and classify specific scientific entity types from scientific literature. This requires handling professional terminology, complex sentence structures, and domain-specific contextual information. Recently, generation methods based on large language models (LLMs) have attempted to co-train models on multiple datasets while using prompt instructions to accurately extract target entities. However, there is currently insufficient exploration of document-level NER methods in the scientific field. To further fill this gap, we propose a document-level information extraction method for scientific literature based on LLMs. Specifically, the method first constructs a document-level labeled dataset to train a binary classifier based on BERT, which determines the relevance of each sentence to the target entities. In the information extraction phase, specific prompt templates are used to guide LLMs in entity recognition and extraction, while performance is enhanced through fine-tuning with the LoRA framework. Extensive experimental evaluations were conducted on two public datasets, covering zero-shot and supervised conditions. The results show that this method significantly improves the performance in document-level information extraction in the scientific field, surpassing traditional methods.