In-context prompt framework for relation extraction in the coal chemical domain based on format information guidance

Large Language Models (LLMs) (e.g., GPT-3) have achieved significant results in Relational Extraction (RE) tasks through In-Context Learning (ICL). However, in specific domains, such as coal chemical industry, LLMs are less effective in RE due to the complex many-to-many relationship structure between multiple chemical reactants and products. To enhance the RE capabilities of LLMs in specific domains, there is a need to provide more adequate and specific cueing information to enhance the understanding of entity relationships in the domain.While domain-specific data usually have high quality and specific formats, existing methods fail to fully utilize such format information. However, this information is crucial for LLMs to understand the textual content and the relationships between entities.Therefore, this paper proposes the LLM-FIGRE framework. First, a sample retrieval model based on similarity calculation is trained by pre-extracting entity-relationship structures from unlabeled data to provide ICL candidate samples. Second, a graph structure is used to organize the format and content information of documents, and “format attractor coefficients” are introduced for nodes to guide the traversal of the graph. Finally, the samples containing format information are used as demonstrations in ICL.Experimental results on the dataset SemEval and the self-constructed coal chemical dataset Coal show that LLM-FIGRE performs excellently in domain-specific RE tasks.