Research On Document-level Relation Extraction Based On Evidential Context

A large amount of unstructured text data is generated daily in the information society,and how to convert the important information in unstructured text into a structured form is a hot research topic in Natural Language Processing.Relation Extraction aims at extracting semantic relationships between given entities from text,which is an important part of natural language techniques such as Knowledge Graph and Information Extraction.Traditional Relation Extraction mainly revolves around the sentence-level.As the Information Extraction scenario changes and the demand grows,Document-level Relation Extraction(DocRE)is receiving more and more attention.Most of the DocRE methods ignore the effective use of the context that is strongly related to the relation facts.Therefore,this paper proposes three DocRE methods around the entry point of evidential context,and the main research contents are as follows:(1)In view of the lack of relation awareness of encoders in existing methods,this dissertation proposes a DocRED method based on relation awareness and contextual refinement.First,we design a pre-training task based on Contrastive Learning to enhance the ability of the encoder to perceive relation context,as well as alleviate the gap between the pre-training and fine-tuning phases in terms of training objectives,and then refine the lexical-level implicit contextual embedding with the help of attention mechanism,which is finally used to assist in the relation prediction of target entity pairs.Experiments on real datasets show that the extraction performance of this method outperforms multiple baseline models.(2)To address the problems of document-graph information redundancy and inference heterogeneity in existing methods,this dissertation proposes a DocRE method based on discourse knowledge and multi-class inference.First,this paper designs a heterogeneous graph construction and encoding module that integrates discourse knowledge and multiheaded dense graph attention mechanism to obtain sentence-level implicit context information through discourse features and effectively reduce the redundancy of documentgraph structure;second,the multi-class inference subgraph module supports multiple types of relation inference,which can effectively avoid the inference promiscuity.At the same time,a simple and efficient mask mechanism is applied to achieve asymmetric information propagation.Extensive experiments show that this method achieves performance beyond several baseline models on several datasets,and outperforms the baseline model for relation extraction on multiple inference types.(3)To address the problem of neglecting the validity judgment of evidence sentences in existing methods,this dissertation proposes an evaluation-based method for joint extraction of document-level relation and evidence sentences.Contexts related to relation facts can be used as evidence sentences for relation judgments,but previous studies tend to focus on how to filter out key sentences from the original text and neglect the judgment of their validity.This method introduces an evidence sentence evaluation mechanism for judging the validity of candidate evidence sentences,and the model completes the relation judgment based on full-text information and sentence-level explicit key context information provided by valid evidence sentences,respectively.The experimental results demonstrate that this method has significant advantages over existing methods in terms of relation and evidence sentence extraction effects.From the analysis of the experiments,it can be seen that the increased effect of the evidence sentence evaluator is significant even in the difficult scenarios of multiple mentions and long entity spacing distance.