Design And Implementation Of Knowledge Extraction System For Overlapping Relations In Complex Semantic Context

Knowledge acquisition is crucial for large-scale knowledge graph construction.The key point is to understand the unstructured natural language text and extract structured knowledge of relational triple from it.However,few existing works excel in solving the overlapping triple problem within complex semantic context,leading to unsatisfactory performance in practice.To address the above issue,in this work,we introduce a fresh perspective to revisit the relational triple extraction task and propose a novel BERT-based framework,Cas Rel,derived from a principled problem formulation.Different from previous works,we are the first to model relations as functions that map subjects to objects in a sentence,which naturally handles the overlapping problem,providing an effective paradigm for knowledge acquisition with the knowledge graph research community.Experiments on six benchmarks with various evaluation metrics validate the effectiveness and generality of the proposed method.According to the empirical results,the Cas Rel framework delivers consistent performance gain across different scenarios.In particular,it achieves a surprisingly groundbreaking improvement in the overlapping problem,outperforming the strongest baseline by 17.5% and 30.2% absolute gain in F1-score on two public datasets NYT and Web NLG,respectively.This certainly ensures the high performance of knowledge extraction in complex contexts.The implemented knowledge extraction system in this work is open-sourced and is publicly available at https://github.com/weizhepei/Cas Rel.