Research On Relation Extraction Algorithm Based On Information Embedding Of Article Structure

Relation Extraction(RE)is a popular research direction in natural language processing(NLP),with the goal of extracting relation between entities from text.This technology can be applied to various applications,including search engines,question-answering systems,intelligent customer service,and intelligent text analysis,among others.The input for document-level relation extraction is an article,which requires aggregation of information and inference of relationships.Existing work places entities and mentions with similar representations in the document-level graph,and the complex edges may lead to redundant information.Additionally,previous research only focuses on inference paths at the entity level,without considering the global interaction of entities across sentences.To address this,this thesis utilizes the article’s structural information,proposes a document-level relation extraction model called GRACR,which has information aggregation and cross-sentence inference networks.The research content and innovation are summarized as follows:(1)To address the problem of how to aggregate rich semantic information from articles,this thesis constructs a document-level graph to model articles and integrate complex semantic information.The document-level graph constructed in this thesis is a heterogeneous graph containing mention nodes and sentence nodes,which only includes sentence nodes and mention nodes to avoid redundant information.In addition,by connecting nodes through three different types of edges,the semantic information of entities is effectively aggregated,thereby improving the performance of the relationship extraction model.(2)To address the problem of inferring relationships between entity pairs across multiple sentences,this thesis introduces an entity-level graph constructed by connecting sentences and entities based on their positions.This thesis model utilizes the connections between entities and proposes an entity-level graph for inference.The entity-level graph is constructed by connecting sentences and entities based on their positions,connecting entity pairs across sentences that have a potential relation.By learning with the context relevant to entities through graph neural networks,each entity node can interact with its most relevant entity node,which is conducive to inferring potential relationships between entity pairs across long distances and multiple sentences.(3)To apply the model proposed in this thesis to real-world scenarios,an answerfocused relation extraction system was designed.The system’s business scenarios primarily include relation extraction and knowledge question answering.The system relies mainly on the relation triples extracted by the relation extraction model and presents them through visualization.By identifying user queries and querying knowledge base triples,the system meets the user’s demand for knowledge question answering.Experimental results show that our model achieves excellent performance on two datasets of document-level RE.It is especially effective in extracting potential relations of cross-sentence entity pairs.