Research And Implementation Of Key Technologies For The Construction Of Knowledge Graph Of Airline Unsafe Events

With the continuous development of China civil aviation,the increasing number of flights inevitably increases the incidence of aviation unsafe incidents.Guaranteeing aviation safety is the primary task of all civil aviation units,and it is becoming more and more important to carry out corresponding research on aviation unsafe incidents to reduce the incidence of aviation unsafe incidents.As a knowledge representation method,the knowledge graph,whose huge semantic network can represent the explicit relationship between massive data and infer the hidden relationship based on the existing relationship,and then fully mine the effective information in the text data,has become a hot research direction in recent years.This thesis focuses on the key technologies of knowledge graph construction of airline unsafe incidents in the field of civil aviation safety,focusing on the basic tasks of knowledge graph construction,including named entity recognition in the field of airline unsafe incidents and airline unsafe incidents.Domain Named Entity Relationship Extraction Two information extraction tasks.Aiming at the goal of building a knowledge map of airline safety incidents,this thesis organizes and analyzes airline safety reports,and defines the named entities and the relationship between entities in the field of airline safety incidents according to the characteristics of the reports.Based on this entity and entity relationship definition standard,through the form of manual labeling,the airline safety report data is used to construct the named entity recognition dataset and entity relationship extraction dataset of airline unsafe events,which are then used to train the information extraction model and carry out knowledge mapping.build.In terms of model construction,a method of named entity recognition and entity relationship extraction based on the fusion of domain dictionaries for vector embedding is proposed.On the basis of character vector embedding,domain words are introduced for vector fusion and embedding,so as to identify entity boundaries more accurately.Based on the constructed domain named entity recognition dataset and entity relationship extraction dataset,it is trained and tested on various deep learning models represented by Bi LSTM-CRF and largescale pre-trained language models represented by BERT.The experimental results show that the proposed method of merging domain dictionaries for vector embedding can improve the performance of two types of tasks,named entity recognition and entity relationship extraction,under different models,which verifies the effectiveness of the method.Based on the realization of two tasks,Neo4 j is used to realize the construction of airline unsafe event knowledge graph,and a visualization system of airline unsafe event knowledge graph is designed and implemented,including the query of event knowledge stored in the current knowledge graph,and the Multi-model named entity recognition and entity relationship extraction are carried out for new incidents,which realizes the semi-automatic construction of the knowledge map of airline unsafe incidents.