Combining Sentence-based Relational Features with Biaffine Mechanism for Triple Extraction

Extraction of relational triples from unstructured text is a crucial task for knowledge graph construction. However, most methods of relational triple extraction ignore the influence of all relations in the input sentence on the extraction of corresponding relation between entity pairs. To address this problem, we present SRBM, a framework combining sentence-based relational features with biaffine mechanism for relational triple extraction. SRBM takes into account the feature of all relations in the sentence and leverages a sentence-relation-perception mechanism to allow entities to perceive the feature. SRBM also adopts biaffine mechanism to fully interact the subject and the corresponding object. We evaluate the SRBM model on NYT and WebNLG datasets. Compared with other baselines, experiments demonstrate that the SRBM model can achieve better results and thus extract overlapping triples and multiple relational triples more efficiently.