Research On Event Argument Extraction Based On Entity Role Interaction And Multitasking Assistance

In order to extract valuable information from massive Internet data and apply the information to different fields,event extraction technology is widely used to extract structured content from unstructured information.Because of the complex semantic environment and diversity of event argument,the event argument extraction subtask has become one of the main factors limiting the development of event extraction.This paper is devoted to the study of event argument extraction task.In previous studies,candidate entity information is not fully explored,and the overall distribution relationship between event argument role and candidate entity as well as the interaction information between candidate entity and event argument role is not fully utilized.To solve these problems,a new model structure is designed in this paper,and the effectiveness of the model is proved by experiments.The research work of this paper mainly includes the following two points.(1)This paper proposes an event argument extraction model based on entity role interaction.In order to capture the overall distribution relationship between candidate entities and argument roles,a new model structure is designed in this paper.Event argument is extracted and modeled as Seq2 Seq structure,and two-way decoder is used to decode forward and backward.Different from previous word-level sequence labeling,entity level decoding is adopted in this paper.The distribution of event argument roles at the semantic level of surrounding entities is captured in the form of two-way decoding,and the distribution of candidate entities is also modeled.Through the two-way decoder,the recognized argument information is transmitted forward and backward,which helps other target arguments to extract.In order to make full use of the interactive information between the candidate entity and the argument role,the attention mechanism is used to integrate the argument role information with the candidate entity information in the process of decoding,so as to improve the effect of each step of decoding.Experimental results on ACE2005,an event extraction dataset,demonstrate the effectiveness of this model.(2)In this paper,a multi-task-assisted event argument extraction model is proposed.To solve the problem that candidate entity information is not fully explored,a new model architecture is proposed in this paper.Using the idea of multi-task learning,the named entity boundary detection is taken as an auxiliary task,and the event argument extraction is taken as the main task.Through the auxiliary task to learn the named entity boundary information,the main task is guided and the main task is taken as an auxiliary event argument extraction.In this paper,BERT model is used as the semantic coding layer.Event types and role information are added into the text,which is spliced with the original text and input into the BERT encoder,so that these information can be fully modeled with the original text,so as to obtain the context-dependent semantic vector,which contains rich role guidance information.And as a shared semantic coding layer in the multi-task learning structure.Based on the multi-task learning mechanism,the joint training of event argument extraction task and named entity boundary detection task is realized by using the correlation between tasks.Experimental results on ACE2005,an event extraction dataset,demonstrate the effectiveness of this model.