Research And Implementation Of Document-Level Event Argument Detection Methods

Event argument extraction is an important branch of natural language processing,and it has important application value in many fields.With the development of machine learning related technology,people are increasingly demanding the integrity and accuracy of event argument detection.In this paper,the relevant methods of document-level event meta-detection are analyzed and studied.The main work and contributions include:(1)Established the benchmark model.First,Bi LSTM is used as the benchmark model to obtain single event document data using the open dataset RAMS.Since the object of this study is event arguments,its detection task is to detect event arguments from text information,identify and classify them under the trigger word type,code the input data using Bi LSTM,obtain the context characteristics of different sentences,and fuse the text content representation with the document sequence feature to get the document-level semantic representation.Finally,the probability score for each word is obtained through the Softmax layer,and the arguments under that event type are identified and classified.(2)Filtered valid feature encoders.In the case of Bi LSTM as the base encoder,three different encoders,CNN,LSTM and Transformer,are used to do comparative experiments.Text encoding of different lengths is converted into pre-set fixed-length vectors.By analyzing the feature extraction features of different encoders,the experimental results of argument detection under different encoders are compared.The Transformer model has the highest F1 value.After that,a meta-detection model based on sequence labeling is established to explore the effect of pre-training model on tasks.The results of BERT+Softmax and BERT+CRF models are compared to verify the validity of CRF models in sequence labeling tasks.(3)A text generation approach is used for argument detection,while improvements are made for generating templates.Using an end-to-end document-level event argument extraction model,this paper defines the problem as a primitive generation task under a given template.On this basis,for each event type,the template is decomposed into sub-templates equal to the number of arguments,and different template improvement strategies are proposed.By fusing semantic information or argument information into the template content,the effects of different prior information on text-based argument extraction tasks are studied.The results show that sub-templates incorporating argument information achieve the best results.Explains that statements with rich event information and strong logic have a positive impact on tasks.(4)Developed and implemented a theorem extraction system.Event argument extraction system integrates text generation-based argument extraction methods to support event argument extraction for user-uploaded document data.This paper makes a comparative analysis on the effectiveness of the methods used and demonstrates the experimental results.The event element detection method used in this paper has low cost,high robustness and good application prospects.