Research On Document-level Financial Event Extraction Based On Syntactic Semantic Enhancement

Event extraction is an important and complex task in the field of information extraction.The main goal of closed domain event extraction is to extract specific information from unstructured text data according to predefined event types and templates to obtain structured data.Because of the complexity of event information,it is very difficult to build a complete large-scale document-level event extraction corpus.Therefore,the document-level event extraction technology without trigger words has become a research hotspot.However,the existing models ignore the behavioral semantics as the core of the event.In addition,the distance information between entities is not fully utilized in event extraction.In order to solve the problem of ignoring behavioral semantics in existing models,a document-level financial event extraction method based on dependency syntax semantic enhancement(PTPCG-DSSE)is proposed in this thesis.This method uses dependency parsing technology,part of speech tagging tools and graph attention network(GAT)to capture the behavioral semantics of entities and the semantic meaning of dependency sentences of entities,and can better model the event semantic information of entities.PTPCG-DSSE model adds the dependency-based entity behavioral semantic enhancement module and the GAT-based entity dependency semantic enhancement module on the basis of PTPCG model.The motivation of the model is as follows: first,the behavior semantics represented by trigger words cannot be used in event extraction without trigger words,so it is necessary to capture the behavior semantics that the entity depends on through the core verbs of the entity;secondly,the GAT can capture the semantic features of neighboring nodes in the dependency graph and provide dependency syntactic information for event extraction.To sum up,enhancing the French meaning of dependency sentences can provide effective clues for document-level event extraction of untriggered words.In addition,distance information is an important syntactic and semantic information,but it is rarely used in event extraction.In order to make full use of the entity distance information in the text,this thesis captures the document-level syntactic semantics through the shortest sentence distance of the entity to provide help for event extraction.At the same time,the finer-grained distance information is obtained by calculating the shortest dependent distance of the entity.In this thesis,three different distance embedding methods are designed,and three models such as PTPCG-ED-division,PTPCG-ED-plus and PTPCG-ED-concat are constructed,and their performance is tested by experiments.In this thesis,experiments are carried out on a large-scale document-level financial event extraction data set.The experimental results show that the F1 value of PTPCG-DSSE model is 1.1% higher than that of PTPCG model,reaching 80.5%.It is proved that the semantic meaning of dependency sentences can play a certain role in event extraction.The average F1 value of PTPCG-ED-concat in single-event and multi-event documents is 86.0% and 69.5%respectively,which is 0.4% and 1.5% higher than that of PTPCG-DSSE model,which proves that entity distance plays an important role in the extraction of multi-event documents.