Knowledge Graph-Augmented Semantic Fusion for Large Language Model-Based Long-Text Classification

Abstract

This paper proposes a text classification method that integrates large language models with knowledge graphs to address the problems of semantic redundancy, insufficient contextual dependency, and limited external knowledge utilization in long-text classification. The method first employs the contextual modeling ability of large language models to obtain global semantic representations of texts, and uses multi- layer self-attention mechanisms to capture cross-paragraph dependencies, thereby alleviating the semantic dispersion and redundancy issues found in traditional approaches. It then introduces a knowledge graph to structurally model entities and relations in the text, and applies embedding and alignment strategies to deeply fuse knowledge graph information with text representations, allowing the model to capture semantic information while gaining explicit knowledge constraints, which enhances the completeness and accuracy of semantic understanding. Furthermore, a series of sensitivity experiments is conducted from multiple perspectives, including hyperparameters, environmental conditions, and data characteristics, to systematically analyze the stability and robustness of the method and verify its applicability and consistency in complex contexts. Experimental results show that the proposed method outperforms comparison approaches on multiple metrics such as AUC, ACC, F1-Score, and Precision, demonstrating the advantages of combining large language models with knowledge graphs in long-text classification. In summary, this study not only achieves an organic integration of semantic modeling and knowledge modeling at the theoretical level but also provides a feasible technical pathway for effective classification of texts.