Bidirectional Network-Based Relational Triple Extraction with Prior Relation Mechanism
Youzi Xiao,Shuai Zheng,Zhiqiang Tian,Jiancheng Shi,Xiaodong Du,Jun Hong
DOI: https://doi.org/10.1007/s10115-024-02241-0
IF: 2.7
2024-01-01
Knowledge and Information Systems
Abstract:Relational triple extraction is a fundamental task for extracting semantic information from unstructured text, which is also the cornerstone for constructing large-scale knowledge graphs. In recent years, the relational triple extraction task has achieved significant gains. However, some challenges still need to be addressed in practical applications, such as relational prediction redundancy, error accumulation, and overlapping triples. To address these challenges, we propose a novel bidirectional network with a prior relation mechanism named BPRM in this paper. Our model consists of a prior relational mechanism, a bidirectional network, and a relation-based alignment strategy. Specifically, our model first mines potential relations by the prior relation mechanism, which reduces redundant relations and discovers implied relations. Then, the bidirectional network fuses prior relation features to tag the subject and the object from two opposite directions, which can address the error accumulation and the overlapping triple problem. Moreover, extracting subjects and objects from both directions can reduce the impact of accumulation errors. Finally, the relation-based alignment strategy removes redundant entity pairs. Our proposed model gains the highest F1-scores of 92.7