Event Sequence Relation Extraction From Scientific And Technological Literature Based On Transformer

Scientific events in scientific and technological literature play an important role in the analysis of the development of the scientific field.However,with the increasing number of scientific and technological literature,how to accurately extract the relationships between scientific events and analyze the development of scientific research is an urgent problem in the study of scientific and technological literature from the event perspective.In this thesis,we introduce hypergraph dual-channel partition network into the task of entity relationship extraction,and use hyperbolic space to model the asymmetric relationship between events for event sequence relation extraction.Finally,we design and implement the extraction of scientific research development in scientific and technological literature.The main work of this thesis is as follows:(1)Entity relationship extraction based on hypergraph dual-channel partition networkIn order to solve the problems of unbalanced feature interactions among subtasks and inadequate handling of inter-entity dependencies in entity relationship extraction tasks,this thesis proposes an entity relationship extraction model(HGCN-DCP)based on a hypergraph dual-channel partition network.First,the entities and relations in the triad are constructed as a hypergraph structure,and the hypergraph convolutional network is used for feature extraction.Then,the similarity between nodes is calculated using a self-attention mechanism,and the nodes and relations in the hypergraph are divided into different task partitions using a partition network,while a shared partition is reserved.Finally,a dual-channel classification model is used for entity recognition(NER)and relationship extraction(RE),respectively.In the RE model,the relationship table and GRU unit are combined to finally obtain the triadic representation of each relationship in the sentence and the relationship score.The experiments validate that the method improves the results in both NER and RE subtasks compared to the comparison methods by constructing a hypergraph structure and a dual-channel classification model in the entity relationship extraction task.(2)Event sequence relation extraction based on hyperbolic space TransformerIn the existing studies of event sequence relation extraction,the embedding vectors in Euclidean space cannot capture the rich asymmetric relations between events,while hyperbolic space has advantages for processing asymmetric relations.For this reason,this thesis proposes a Transformer improvement model(TF-HYS)based on hyperbolic space for extracting temporal information related to events in text.The encoder of the model first embeds the input onto a Pongaree sphere,performs attention operations and feedforward operations in the hyperbolic space,and generates a contextual representation of the input by hyperbolic cosine similarity and Poincaré distance.The decoder of the model introduces a hyperbolic tangent function in the hyperbolic feedforward layer for capturing the hierarchical structure of the contextual representation.Finally,the event time series relationships are extracted by temporal knowledge graph and hyperbolic multinomial logistic regression.Experiments verified that the method more accurately captured the asymmetric relationships between events and outperformed the comparison model on both MATRES and TCR datasets in the event sequence relation extraction task.(3)Extraction of scientific research development venation in scientific literatureWe have designed and implemented the extraction of sequence relation between scientific research events and the analysis of the development of scientific research in scientific and technological literature,the main functions include entity identification,relationship extraction,event collection,event sequence relation extraction,literature retrieval and development analysis.The scientific research development chain analyzes the scientific research sequence relation in scientific and technological literature to explore the development history of scientific and technological literature,and provides ideas and references for researchers to plan and select scientific research topics in the future.