RCTE: Relation Candidate-guided few-shot relational Triple Extraction
Bowen Liao, Zhikun Lu, Yingqi Guo · 2024
Few-shot Relational Triple Extraction (FRTE) task aims to extract new relational triples from unstructured text with limited labeled samples. Recently, FRTE approaches prefer the relation-then-entity paradigm to avoid the entity discrepancy problem and the tag explosion problem suffered by entity-then-relation paradigm and unified paradigm, respectively. Although the relation-then-entity approaches becomes mainstream, they ignore an important problem that is the error transmission from the upstream relation extraction process to the downstream entity identification process. To solve the problem, we propose a novel Relation Candidate-guided few-shot relational Triple Extraction approach, namely RCTE. It adopts relation-then-entity paradigm and based on gate mechanism and beam search framework. We design a gate mechanism. Then we provide a beam search framework to identify entities for each relation candidate and select the relational triple with the largest joint prediction probability as the final predicted result, alleviating the error transmission by augmenting relation candidates and comprehensively measuring prediction probabilities of two extraction subprocesses. The experiments on the FewRel dataset verify the validity, rationality and generalization of our method, and it achieves 45.28% and 41.87% micro F1 score in the 5-way-5-shot and 10-way-10-shot tasks, respectively, the results show that it achieves grate performance compared to other baseline models.