Research On Event Extraction Methods With Topic Feature And Implicit Sentence Structure

Event Extraction aims to extract the information of events in the text and transform it into structured data,which plays an important role in information retrieval and event map construction.The existing methods can be divided into pipeline methods and joint methods.The pipeline methods will cause error accumulation,and most of the sentence-level event extraction joint methods lack the overall information of the text so that they can't deal with the ambiguity of triggers,while the document-level joint methods are complex in modeling.And syntactic features are important for the task,but only a few methods introduce syntactic information based on pre-training tools into event extraction;In fact,it is common for multiple events or overlapping of arguments,but few models take this into account.A joint method based on topic features and implicit sentence structure is proposed to solve the above problems.Firstly,a document-level topic feature is introduced into the sentence-level model.Secondly,an influence matrix of implicit syntactic information from BERT is extracted and modeled with event extraction jointly.And multiple triggers and multiple roles of the entities are extracted.Through these methods,the event extraction task can be improved.The main contributions of this thesis are:1.Propose a joint model of event extraction based on topic features: by combining BERT representation vector and LDA results,the topic features are obtained to improve the ambiguity of triggers,and a joint model of subtasks is established to solve the accumulation problem.The design of the model improves the recall rate of the model by facing multiple events and overlapping arguments.2.Propose a joint model of event extraction based on implicit sentence structure: by extracting the implicit syntactic information from BERT results,which is trained with EE subtasks jointly to solve the accumulation of error.The design of the model improves the recall rate of the model by facing multiple events and overlapping arguments.3.Design and experiment on public data sets.The common evaluation tasks and indicators are used to evaluate the model,and the results are compared and analyzed with the existing models under the same evaluation indicators.The results show that these methods perform well in event extraction tasks,verifying the effectiveness of these methods.Research on the event extraction joint model based on topic features and implicit sentence structure can effectively improve the performance of the event extraction model,which is of great significance to the event extraction and related tasks.