Document-level Relation Extraction Based On Dependency Syntax Analysis

In recent years,the rapidly growing Internet has generated an enormous amount of unstructured text,which contains a variety of information from the real world.The task of extracting relations between entities is based on the main semantic information from these disorganized and scattered texts into structured information that is easily understood by machines.Entity relation extraction includes two subtasks,named entity recognition and relation extraction,as one of the subtasks relation extraction is an important prerequisite for downstream tasks.A large amount of research has been done on single-sentence relation extraction tasks,and scholars have shifted their attention from single-sentence to document-level complex cross-sentence document-level relation extraction tasks.Compared with single-sentence relation extraction tasks,document-level relation extraction tasks have challenges such as cross-sentence extraction and inference of relations between multiple entities.At the same time,previous studies have also shown that the dependency relations obtained from dependency syntactic analysis can be useful for relation extraction tasks,but there are still some problems that need to be solved in the full utilization of dependency trees.In this paper,we explore and study the above three problems,and in general,the main work of this paper includes.1.For the problem of cross-sentence relation extraction and relation reasoning A relation extraction model based on graph fusion strategy is proposed,which constructs both explicit and implicit graphs by semantic information and dependency structure information,and secondly,fuses graph structures containing two different information by multiple graph fusion strategies to obtain the final representation vector of each entity mentioning nodes.In order to solve the relation inference problem,all the entities are wandered on the graph to capture cross-entity information to obtain cross-entity information for relation inference between entities.More then that,the updated entity representations are linked separately and passed through the feed forward neural network to obtain the predicted relational category vector,and finally of that,this relational category vector is activated by the softmax function to obtain the final relational category prediction.2.Study of full utilization of dependency trees.A model that allows for dependency category encoding is proposed,which uses a long and short term memory sequence network as the underlying architecture of the model.Inspired by the fact that each node of the extracted dependency tree is connected by only one parent node,the sentences to be extracted are fed into the LSTM to obtain the intermediate hidden vectors.According to the parent-child relation in the dependency tree and the dependency relation between the parent and child are spliced and output again by the bidirectional long and short term memory network model.At the output,each node already contains the information of dependent parents and the dependency categories between parents and children.Finally,the predicted relation vector is output by the relation classifier,and then the vector is nonlinearized for the final relation classification.3.In order to comprehensively and effectively evaluate the model proposed in this paper.The experimental results show that the proposed method is superior to previous studies.The evaluation method and some specific examples show that this paper can effectively deal with this problem of relation extraction and relation inference across sentences that cannot be solved by previous models.