Research On Document-Level Relation Extraction Model Based On Global-to-Local Network And Knowledge Injection

Relation extraction(RE)aims to identify the semantic relations between named entities in text.It is an important part of the Information Extraction and has import research and application value.While previous work focuses on extracting semantic relations within a sentence,recent studies have witnessed it raised to the document level,a.k.a.document-level RE task,which requires the extraction system to be able to perform complex reasoning with entities and mentions throughout an entire document,such as logical reasoning,coreference reasoning and common-sense reasoning.To cope with the limitations of existing works,this thesis proposes two document-level RE models based on deep learning:(1)Compared with sentence-level RE,document-level RE needs to model richer information in document,and it needs to rely on multiple mentions of entities in different sentences to model the complex interactions between multiple entities.But existing works are not good in the use of entity mention information.This thesis proposes a novel model GLRE to document-level RE,by encoding the document information in terms of entity global and local representations as well as context relation representations.Entity global representations model the semantic information of all entities in the document,entity local representations aggregate the contextual information of multiple mentions of specific entities for specific entity pairs,and context relation representations encode the topic information using other relations.Experimental results demonstrate that our model achieves superior performance on two public datasets(Doc RED and CDR)for document-level RE.Compared with the best result of the comparison methods,the F1-scores on Doc RED and CDR are improved by 3.4 and 5.4,respectively,and it is particularly effective in extracting relations between entities of long distance and having multiple mentions.(2)Document-level RE needs coreference reasoning and common-sense reasoning based on entities and their mentions.The processing of these reasoning requires the support of external knowledge.Therefore,this thesis proposes an entity-driven knowledge injection model(KIRE)to enhance current RE models.KIRE injects the knowledge from knowledge graph and pre-trained coreference resolution model(involving coreference triple facts,entity attribute triple facts and relation triple facts)into current RE models through multi-task learning mechanism to improve their performance.For coreference triple facts,coreference knowledge is introduced based on knowledge distillation mechanism.For entity triple facts,the entity attribute triples and relation triples are sequentially encoded to obtain the entity representation based on the knowledge graph,and then the learned entity representation is integrated into the RE model based on fusion reconstruction.The experimental results on two benchmark dataset Doc RED and DWIE demonstrate that the generalization of KIRE to both graphbased and sequence-based models,and the stable improvement to the state-of-the-art document-level RE models,with an increase of F1-scores up to 2.62.