Design And Implementation Of Neural Network Event Extraction Eystem Based On Attention Mechanism

With the continuous development of society,the amount of text information on the network is growing explosively.Although this rich information can increase the possibility of users to acquire the required knowledge,it reduces the speed of users to acquire effective information.To solve these problems,many scholars began to study information extraction,and events,as the core form of information expression,have attracted much attention in recent years.Event extraction is to transform unstructured text containing event information into a structured form,which is divided into two tasks: event detection and event element extraction.It is widely used in information search,intelligent question answering,machine summary,and other fields,and has profound research significance.The traditional event extraction methods can not make full use of context information and ignore key features.Therefore,based on the gated recurrent unit,and adding attention mechanism,convolution layer and Encoder-Decoder framework,a series of studies and applications of event extraction are carried out in this thesis.The specific work and contributions are as follows.(1)Construct an event detection model(ATT-Conv-BiGRU)based on the convolutional gated recurrent unit.In feature selection,this thesis doesn't rely on part-ofspeech tagging and entity recognition results and combines text word vector and entity type information to form context representation learning vector,to directly utilize the relationship between event element information and candidate trigger words.Secondly,this thesis extracts sentence-level features by using a bi-directional gated recurrent unit network and extracts word-level features by convolution operation,to solve the problem of misclassification of event categories caused by the existence of synonyms in traditional methods.Finally,the two features are cascaded as the input of the golden standard attention mechanism output layer and using softmax to predict the event detection results.Experiments show that the model improves the accuracy and efficiency of event detection.(2)Construct an element extraction model(ATT-ED-BiGRU)based on the EncoderDecoder framework.Firstly,the text vector is composed of three parts: event trigger word type,candidate element position,and word vector.Second,this thesis uses bidirectional and unidirectional gated recurrent unit networks to capture semantic information at the encoding layer and decoding layer,respectively.Finally,the attention layer can effectively show the impact of event trigger words and other event arguments on the candidate event arguments.This solves the problem that when an event sentence contains multiple events,the event argument role is easy to identify confusion,thereby improving the accuracy of element extraction.(3)Construct an event extraction system in the Chinese financial field.This thesis uses a semi-automatic annotation tool and manual secondary tagging to build a financial news corpus.In combination with the event detection model in chapter 3 and the event factor extraction model in chapter 4,a C/S structure-based news event extraction system in the financial field is designed and implemented.Secondly,this thesis also designed an iterative learning framework to solve the problems of difficult domain transplantation and insufficient corpus size in the actual application of the system and improve the portability and robustness of the system.Finally,this thesis uses the event map and event timeline function to display event extraction results,making the recording of financial events more systematic.