Research And Application Of Graph Neural Network In Chinese Financial Event Extraction System

As an important task in the field of natural language processing,event extraction aims to extract structured information tuples that can describe an event from unstructured text,and realize the extraction and compression of key information.It is widely used in information extraction,risk control,emotional analysis,etc.However,the traditional event extraction research only proposes an extraction model for the English text field,and does not conduct in-depth research on the Chinese text field,especially the widely used Chinese financial text field.Moreover,due to the high and cumbersome manual labeling costs required in supervised learning,the available training data sets are small and the scope of application is narrow.Aiming at the above problems,the main contributions of this thesis mainly include the following three parts:1.The Chinese financial event extraction via graph attention network(CAEE)algorithm is proposed.The CAEE model considers the event extraction task as a character-level sequence labeling task,integrates the word information into the word embedding vector through word embedding,and then constructs an isomorphic graph based on the word-character structure through dependency syntax analysis,and then uses the graph attention network to aggregate to syntactic information.Finally,the final extraction result is obtained by joint extraction model.In order to verify the effectiveness of the model on Chinese texts,we collect and annotates Chinese financial news,and obtains a Chinese financial event extraction dataset.Compared with the existing event extraction model on this dataset and the ACE2005 dataset,CAEE has obtained a significant improvement,which proves the effectiveness of the model in the field of Chinese text.2.The adversarial Training for Distant Supervised Event Extraction(DSEE)is proposed,which trains an event detection model through annotated datasets and knowledge graphs,and annotates unstructured texts through triples of knowledge graphs.Then,the generator and the discriminator are trained through adversarial training,so that the generator has the ability to screen out the generated data sets with high reliability,and finally obtain the filtered remote supervision datasets.We prove the effectiveness of DSEE through manual verification and event extraction verification results.3.Combining the two algorithms of CAEE and DSEE,the Chinese financial event extraction prototype system is designed and implemented.The system mainly includes three modules: training data set generation,model training,model prediction,and together with three modules including user interaction,server and algorithm.The user interaction terminal mainly provides user visual operation interface and result display interface.The server side mainly responds to the request and calls the corresponding algorithm side interface.The algorithm side mainly performs event extraction and training data set generation tasks according to the CAEE and DSEE algorithms,and returns the corresponding results.