Research And Implementation Of Event Argument Extraction

With the advent of the era of big data,people get a growing number of information.How to quickly extract the key points that people care from the complex information is particularly important.Event extraction technology can extract the events that people are interested in from the text and show them in a structured form.In business,politics,medical,social and other fields,event extraction can help people extract key information and make important decisions.Event arguments are the elements that used to describe events.This thesis focuses on the extraction of event argument,designs and implements the event argument extraction system based on this.At present,event argument extraction has the problem of small scale of labeled data.This thesis attempts to use the semi-supervised learning method based on Co-training,uses labeled data and a large number of unlabeled data,iteratively trains the two classifiers,and finally integrates the results of the two classifiers to improve the effect of event argument extraction.Meanwhile this thesis deeply studies the influence of different thresholds and different data sizes on the effect of event argument extraction.Especially for the argument with less data distribution,this thesis studies the language characteristics,manually formulates a small number of rules,and uses the Bootstrapping method to iteratively generate rules for pattern matching to extract argument.Further,through experiments,this thesis studies the effects of different machine learning methods such as Long Short-Term Memory,Support Vector Machine and Conditional Random Field,and different combinations of text features,such as part of speech,named entity,location and syntactic relationship on the effect of event argument extraction.On this basis,this thesis designs and builds an appropriate method for event argument extraction.Finally,based on the above research,this thesis implements an event argument extraction system,which can choose different extraction methods,upload data and train model online.