| The knowledge graph has received a lot of attention from scholars since its introduction.After years of research and development,it has become the foundation and core of artificial intelligence research and intelligent information technology.As a form of knowledge representation,knowledge graph is playing an increasingly important role in healthcare,finance and education,and are driving human society from data intelligence to knowledge intelligence.As a prerequisite for the application of knowledge graph,its construction is a challenging task.At present,constructing knowledge graph mostly requires manual intervention,which will consume a lot of time and human resources.So many researchers are devoted to the automatic construction of knowledge graph.Natural language is the most common form of knowledge representation,and how to extract the relation triplet from natural language is the key to achieving automatic knowledge graph construction.Relation triplet extraction is a core subtask of information extraction,which aims to jointly extract the subject and object with their relation from natural language text.Relation triplet extraction is an important and challenging natural language processing task.With the development of technologies related to natural language processing,the performance of relation triplet extraction has also been greatly improved.The current mainstream relation triplet extraction methods are based on sequence annotation.Compared to previous research,these methods have made great progress,but it also faced some problems,mainly the error propagation problem and various entity overlapping scenarios.In order to solve the above problems,this paper explores relation triplet extraction methods based on entity pair tagging,and our work mainly includes the following two aspects:(1)This paper proposes an entity pair recognition method based on 2D tagging,which can effectively solve the error propagation problem and various entity overlapping cases.And we named it as 2DEPT.This approach keeps a matrix for each relation and in each matrix the word pairs under the corresponding relation are classified through Biaffine attention.We decode the matrices corresponding to all relations to obtain relation triplets contained in the input text.To validate the performance of 2DEPT,we conducted experiments on two open datasets,NYT and Web NLG.The experimental results show that the proposed 2DEPT model realizes the best performance on all F1 values and performs better than baselines overall.In addition,experimental results on complex cases demonstrate that the 2DEPT model is indeed effective in dealing with error propagation problem and various entity overlapping scenarios,and has strong stability.(2)2DEPT is an all-relation oriented approach to entity pair tagging,which suffers from the problem of redundant relations.To address redundant relation problem,this paper modifies the 2DEPT model and proposes a relation triplet extraction method named RF2 DTagging,which is based on relation filtering and entity pair tagging.This approach can improve the efficiency of relation triplet extraction and avoid wasting computational resources.RF2 DTagging first uses a relation filter to filter out confident relations and then identifies the corresponding entity pairs for them.Comparative experimental results with 2DEPT show that RF2 DTagging gains significant efficiency improvement despite performance degradation,which has an important reference value for latency-sensitive business requirements.In addition,we also conducted experiments on three open Chinese datasets CCKS2019-Task3,CMe IE and Du E2.0,and the experimental results show that RF2 DTagging achieves the best performance for almost all evaluation criteria on three datasets,demonstrating the effectiveness of RF2 DTagging for the crosslinguistic relation triplet extraction task. |
