Research On Key Technologies Of Text-Based Event Extraction

Internet contains a large amount of valuable information,so intelligent question answering system,information retrieval,public opinion monitoring and other intelligent information technologies have broad application prospects.As premise of the above tasks,text-based event extraction still faces many challenges.Event detection is the difficulty of event extraction as well as the premise of argument extraction,and argument extraction is the key of event structured representation.At present,the accuracy of event detection is not high.Moreover,the research on event detection is generally oriented to open datasets,whereas littele is based on practical scenarios.Research of event detection in this thesis serves the practical issue,and a three-stage annotation method for corpus in the actual scene is proposed,for experiments and effectiveness verification of event detection technology.And to improve the precision and recall of event detection,this thesis proposes three event detection models: event detection model based on knowledge engineering,trigger word detection based on deep learning,event detection with joint learning of semantic and syntactic representation.The three methods solve the difficulties in the detection task from different prospects and complement each other,which are fused in the verification system to improve the final detection effect.Upon the event detection results,this thesis proposes an argument extraction model based on dependency attention aware GCN model.Experiments on open datasets show that the model is effective.The details are as follows:(1)This thesis proposes a three-stage annotation method for corpus about South China Sea and Taiwan Strait,which reduces the difficulty of matching by filtering the not-event sentences and improves the tagging accuracy by trigger word expansion.The three-stage annotation method effectively improves the tagging quality and greatly reduces the manual workload.At the same time,based upon the first two stages of the three-stage annotation method,an event detection method based on knowledge engineering is constructed,which alleviates the rigidity and the trigger number limit when matching.(2)Although the event detection method based on knowledge engineering alleviates the problems the rigidity and the trigger number limit when matching,there are still triggers not covered by the trigger list in actual event sentences.To solve this problem thoroughly,using the semantic understanding ability of deep learning model to directly detect trigger words,this thesis constructs a trigger detection model based on BERTBi LSTM-Attention-CRF.This method avoids the mechanical matching and could be applied more universally,making it possible to detect triggers not covered by trigger list.(3)Aiming at the problem that trigger words in some event sentences are lack or not obvious,this thesis proposes an event detection model with joint learning of semantic and syntactic representation,named BDD.The model skips trigger words and directly takes the event sentence itself as the detection object.BDD proposes an LSTM based on dependency tree and an attention mechanism based on dependency vector,so as to improve the ability to represent the sentence meaning.(4)As for event argument extraction,in order to determine the argument boundary,this thesis proposes an argument extraction model based on dependency attention aware graph convolution network,DAGCN-AE,to strengthen the cohesion of the feature representations of words in the same argument,as well as strengthen the distinction of the feature representation of words in different arguments.