Star-transformer based semantic enhanced union relation extraction

• Published in: The Journal of supercomputing Vol. 81; no. 10
• Main Authors: Hu, Feng, Pei, Wei, Wu, Yirong, Hu, Qin, Wang, Ben, Sun, Shuifa
• Format: Journal Article
• Language: English
• Published: New York Springer Nature B.V 11.07.2025
• Subjects: Complexity; Consumption; Knowledge representation; Natural language processing; Semantics; Sentences
• DOI: 10.1007/s11227-025-07591-2
• ISSN: 1573-0484; 0920-8542

Relation extraction is a crucial task in natural language processing and knowledge graph construction. However, existing methods often suffer from three major limitations: (1) they overlook the intrinsic semantic associations between subjects and objects; (2) they have limited generalization in long-text scenarios; and (3) their resource consumption grows rapidly with sentence length due to matrix-based tagging strategies. To address the issues, we propose ST-ICRE (star-transformer intrinsic correlation relation extraction), a novel relation extraction model that leverages the intrinsic correlations between subjects and objects. The model effectively captures the associations between the subject and the object by adopting the relay node and satellite node interaction mechanism within the star-transformer structure, thereby improving the sentence’s semantic representation. Moreover, to mitigate the resource consumption of the conventional “filling-the-form” approach, we introduce an entity tagging method with linear-level complexity. The experimental results on both Chinese and English datasets (DuIE2.0, CMeIE, SciERC, FinRED) demonstrate that ST-ICRE outperforms strong baseline models, achieving F1-score improvements of 4.4% on DuIE2.0 and 11.4% on SciERC. Notably, ST-ICRE shows significant performance in handling complex scenarios with overlapping entity relationships.