Entity-Aware Graph Neural Modeling for Structured Information Extraction in the Financial Domain

Abstract

This paper addresses the challenging problem of structured information extraction in the financial domain. It proposes a joint extraction model that integrates graph neural networks with an entity-aware mechanism. The goal is to improve the recognition of entities and relations in complex financial texts. The method constructs a Syntax-Semantics Hybrid Graph by incorporating dependency syntax and semantic co-occurrence relations. It models words, entities, and contextual information in a graph structure, effectively capturing long-range dependencies and hidden connections between non-contiguous entities. At the same time, the model introduces an Entity-Aware Representation Enhancement mechanism. Based on a pre-trained language model, it strengthens the explicit representation of entities. This guides the model to focus on core semantic areas during sequence encoding, improving the accuracy of entity boundary detection and relation extraction. The proposed method demonstrates strong performance across multiple evaluation tasks. It shows high stability and robustness, especially in handling high entity density, input noise perturbation, and adaptation to different pre-trained models. Comparative and ablation experiments confirm the effectiveness and complementarity of syntactic-semantic structure fusion and entity-aware enhancement in financial text extraction. The structured modeling and semantic representation strategy proposed in this paper provides a technical foundation for deeper understanding and high-quality knowledge construction in financial corpora. It also offers a new research perspective for extraction tasks in complex language scenarios.