Joint Entity and Relation Extraction Based on Large Language Model and Graph Convolutional Networks

Joint extraction of entity and relation from unstructured text is a critical task in information extraction. Although current methods have achieved considerable performance, they still suffer from inherent limitations such as imbalanced label distribution in datasets, insufficient consideration of the complex structural features of entity relationships. To address these issues, this paper introduces a novel approach by proposing a joint entity and relation extraction framework based on Large Language Model (LLM) and Graph Convolutional Networks (GCN). Specifically, the model employs a strategy that enhances data using LLM, assisting the expansion of corpora with fewer labels. Subsequently, a GCN component is designed to fully explore the complex structural features of entity and relation within sentences. Finally, a biaffine classification component is developed to align entity and relation into matrix form, making it easier to capture triples. A wide range of experiments was carried out on two datasets, namely The New York Times (NYT) and Web Natural Language Generation (WebNLG). The results of the experiments show that this approach reaches state-of-the-art performance, maintaining stable extraction capabilities while capturing more entity pairs, further validating its effectiveness in addressing the overlapping triples problem.