Research Information System
Engineering Applications of Artificial Intelligence, vol.163, 2026 (SCI-Expanded, Scopus)
Relational triple extraction refers to identifying triples consisting of entities and relations form unstructured texts. The existing studies usually adopt an unidirectional extraction strategy, which fails to fully explore the semantic information related to entities and relations. And they rely heavily on initial extraction results when conducting multi-step extraction. To address this issue, we propose a novel Entity-Relation Enhanced Bidirectional Information Fusion approach (ER-EBIF). Specifically, we adopt a bidirectional extraction strategy of ”entity-to-relation” and ”relation-to-entity” to identify triples. One branch extracts potential relations, then extracts entities associated with those relations. The other branch initially extracts the potential subjects and objects as well as subsequently extracts relations between pairs of entities consisting of subjects and objects. Moreover, the contextual information is enhanced with a self-attention mechanism by integrating the information of potential relations and potential entities to better exploit the semantic information of entities and relations. Extensive experimental results on various datasets show that ER-EBIF exhibits better performance than other baselines and effectiveness in addressing the issue of dependency on initial results in multi-step extraction.