Research On Joint Extraction Of Entity Relations By Fusing Entity Local Information

Entity and relation extraction is the core subtask of information extraction and the important foundation for the textual content comprehension task of natural language processing.Entity and relation extraction research presents the information content in text in a structured way,the text is elevated from the text representation level to the text understanding level,and solves the task of relationship classification among entities in the original text.In the early stage,the pipeline method was mainly used to deal with the entity and relation extraction task,which was simple and easy to implement,but it cut the correlation between subtasks and brought about problems such as error propagation.All the research methods basically stay in the efficient identification of entities,ignoring the effect of entity's local information on entity relationship extraction,and extract relation basically only considers the existence of a single relation.In this thesis,based on the Bert pre-trained language model and the fusion of localized information of entities,the joint entity recognition and relation extraction is studied.How to identify and represent entities in the joint framework and how to fuse localized effective information and to effectively identify multiple relationships existing in text is experimentally studied.The research content is mainly divided into the following aspects:(1)Propose a multi-head selection entity and relation joint extraction model which fuses information between entities.We use the BERT model to encode text information and Bi RNN-CRF architecture to implement named entity recognition module;The joint extraction method integrates the information between entities with the embedding of the entity label.The integrated information is then used to calculate the scoring matrix.At the same time,multiple relationships were extracted to achieve joint modeling through parameter sharing and loss value fusion.The experimental results show that,while the pre-trained language model can effectively represent text information,the fusion representation of information between entities can more effectively impove the model performance.(2)Propose a joint extraction model of entity recognition and relation extraction based on entity span and entity contextual information.We utilize sequential block search to identify all the named entities in the text sequence.The span representation of candidate entities combines randomly with local contextual information.The integrated information is then classified into entity relations with the biaffine mechanism.Negative sampling is used for model training in the two subtasks of the model,which speeds up the training process while avoiding over-fitting the model.The experimental results show that the contextual information of entities can improve the performance of relationship classification model more effectively than the information between entities,Strong negative sampling method can also participate in model learning effectively.(3)Domain datasets are constructed using crawler technique and manual annotation.In experiments,most of the datasets are English datasets,while the Chinese datasets are basically self-owned datasets annotated by hands.Other problems exist like different expression in different areas and a variety of technical terms.In order to verify the performance of the model on the datasets with multiple entities and multiple relationships,and the generalization of the model.Therefore,we use the crawler technology to get the data of the stock field,and annotate and correct it manually with the help of students majoring in Finance.Our new dataset proves the effectiveness of the model in Chapter 4 of this thesis and provides new research ideas.