Civil aviation service domain knowledge graph construction method and system and storage medium

Abstract

The invention discloses a civil aviation service domain knowledge graph construction method and system and a storage medium. The method comprises the steps that S1, a BERT-BiLSTM-CRF algorithm model carries out entity extraction processing, and an entity vector sequence, a feature vector sequence and a labeling sequence which are correlated with one another are obtained; s2, the convolutional neural network model extracts sentence vectors and contained entity vectors, and n filters are adopted for recognition and extraction to obtain an entity-relation-entity triple database; s3, integrating the annotation information through a conditional random field entity node integration model, correspondingly storing the annotation information in the entity and taking the annotation information as an attribute value of the entity; and S4, utilizing the triple database and the attribute value of the integrated entity to carry out link fusion construction to obtain the civil aviation knowledge graph. According to the method, the comprehensive and accurate civil aviation knowledge graph of the relation between entities can be obtained based on the civil aviation knowledge text database, knowledge questions, answers and queries of passengers can be met, the knowledge graph can be used as training and educational resources, and the overall service level is improved.

Description

Technical Field

[0001] The present invention relates to the field of civil aviation data processing and application, and in particular to a method, system and storage medium for constructing a knowledge graph in the field of civil aviation services. Background Art

[0002] As an important part of global economic and social development, the civil aviation industry has a profound and extensive impact on modern society. With the continuous growth of global air traffic, the complexity and diversity of civil aviation services are also increasing. In order to ensure efficient and reliable flight operations, timely and accurate acquisition and processing of a large number of navigation notices, regulations and standards has become an important task for airlines and civil aviation management departments. However, the current civil aviation service field faces many challenges, especially in terms of information management and knowledge acquisition.

[0003] The civil aviation industry pursues the development requirements of "sincere service" and smart civil aviation construction, and there is an urgent need to further improve the quality of civil aviation services and meet passenger satisfaction. With the rise of intelligent question-answering robots and AI big data, early, such as intelligent service robots set up in the terminal hall, or civil aviation or airline websites for AI intelligent question-answering and information query, or display screens set up in front of aircraft seats for passengers to query and learn, and other aspects, knowledge promotion and information query in the civil aviation field are important aspects of improving services, and how to build a complete set of knowledge graphs or further obtain a complete and comprehensive knowledge graph is an important foundation for intelligent question-answering robots and AI big data.

[0004] In the field of civil aviation, there is a huge amount of data content. It is difficult for civil aviation staff to acquire knowledge and retrieve information, and ordinary passengers know very little. Therefore, there is an information gap between ordinary passengers and civil aviation staff. For example, passengers cannot quickly obtain the required civil aviation knowledge through simple inquiries to improve the scientificity and accuracy of travel decisions; in addition, airlines cannot better manage and utilize various types of navigation notices and standard information to improve operational efficiency and service quality, which seriously restricts the improvement of civil aviation services and is not conducive to improving the intelligence and information level of civil aviation services.

[0005] Constructing a knowledge graph for civil aviation knowledge query and question-answering and relying on the knowledge graph to form important links in massive data content makes it easier to obtain the data content corresponding to the entity nodes of interest in a more comprehensive and efficient manner. In addition to satisfying passengers' knowledge Q&A and query needs (for ordinary passengers, due to their lack of professional background, they often feel confused and overwhelmed when faced with complex aviation service information and professional terminology), it also mainly provides important data support for civil aviation staff's knowledge learning, retrieval and query, etc., and is also a technical problem that is currently urgently needed to be solved by major airlines and related institutions. Summary of the invention

[0006] The purpose of the present invention is to solve the technical problems pointed out by the background technology, and to provide a method, system and storage medium for constructing a knowledge graph in the field of civil aviation services. Entity vectors and annotations are extracted through the BERT-BiLSTM-CRF algorithm model, and then the entity correspondence in the sentence vector is identified through a convolutional neural network model including an attention mechanism, and entity-relationship-entity triples are constructed. The annotation information is integrated and stored as the attribute value of the entity through the conditional random field entity node integration model, and then the civil aviation knowledge graph is obtained by linking and fusion.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A method for constructing a knowledge graph in the field of civil aviation services, the method comprising:

[0009] S1. Construct a BERT-BiLSTM-CRF algorithm model and a civil aviation knowledge text database. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation fit from the civil aviation knowledge text database and obtains interrelated entity vector sequences, feature vector sequences, and annotation sequences.

[0010] S2. Construct a convolutional neural network model based on the attention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract the local feature matrix C = (c1, c2, ..., c n ), construct a triple database of entity-relationship-entity;

[0011] S3. Construct a conditional random field entity node integration model based on semi-supervised learning or / and weakly supervised learning. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents the feature function that extracts features from the input labeled sequence, λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, where j represents the position of the input annotation sequence; the annotation information in the integrated annotation sequence is stored in the entity and serves as the attribute value of the entity;

[0012] S4. Use the entity-relationship-entity of the triple database in step S2 and the attribute values ​​of the integrated entity to link and fuse to obtain a civil aviation knowledge graph.

[0013] In order to better implement the present invention, the present invention also includes the following method:

[0014] A. Use the cosine similarity method to calculate the similarity of entities in different sentence vectors:

[0015] Where similarity(e, c) represents the judgment result of the similarity between e and c, e represents the selected entity, and c represents the entity used for similarity comparison, that is, the candidate entity. represents the vector of entity e, A vector representing entity C;

[0016] B. Entity C with a similarity greater than the similarity threshold or / and the highest similarity is merged with entity e in terms of attribute values ​​and relationships.

[0017] Preferably, the sentence vectors and the entity vectors contained therein in the civil aviation knowledge text database are extracted respectively by using sliding windows, and the entity vectors contained in the sentence vectors are extracted by using n sliding windows corresponding to the filters.

[0018] Preferably, the local feature matrix also includes the corresponding weight matrix W∈R w*d , where R represents the real number system, w represents the sliding window size, and d represents the dimension after embedding concatenation; the sentence vector is recorded as a sequence (q1, q2, ...q i , …, q m ), q i The data representing position i in the sentence is convolved according to the following formula to obtain the sequence:

[0019] c i =Wq i-w+1：i +b,Wq i-w+1：i Indicates q i-w+1 to q i The weight of b represents the bias vector, c i Represents the local features of position i in the sentence,

[0020] And taking the entities in the sentence vector as the core, a local feature matrix of the entity correspondence in the sentence vector is extracted.

[0021] Preferably, the model update expression using the semi-supervised learning method in the conditional random field entity node integration model is as follows:

[0022] P(y|x;θ)=Model(x;θ), where θ represents the model parameters;

[0023] θ new =argmaxlogP(Y|X;θ1),θ new represents the model parameters after iterative update, θ1 represents the model parameters before iterative update, X represents the original labeled data set, and Y represents the integrated labeled data set.

[0024] Preferably, the entity vector sequence of the BERT-BiLSTM-CRF algorithm model is arranged in sequence according to the entity vectors, the feature vectors in the feature vector sequence are features corresponding to the entity vectors extracted from the text data and arranged in sequence according to the feature vectors, and the annotation sequence includes semantic dependency information between entities.

[0025] Preferably, in step S1, the BERT-BiLSTM-CRF algorithm model includes a BERT masked language model, a BiLSTM network and a conditional random field CRF module. The BERT masked language model uses a masked language method to generate a deep bidirectional language representation for civil aviation knowledge text data in a civil aviation knowledge text database. The BiLSTM network uses forward and reverse methods to capture bidirectional long-distance semantic dependencies of information. The conditional random field CRF module extracts entity nodes to obtain mutually related entity vector sequences, and uses a set of feature functions to obtain mutually related feature vector sequences and annotation sequences.

[0026] Preferably, the present invention also includes the following method:

[0027] S5. Continuously update and collect civil aviation knowledge data and build a civil aviation knowledge database. The user inputs keywords or entity words corresponding to the entity, and obtains the civil aviation knowledge graph of each entity in the civil aviation knowledge database according to steps S1 to S4. The civil aviation knowledge graph is used to search and analyze the keywords or entity words corresponding to the entity in the civil aviation knowledge database, and the search results corresponding to the entity are obtained.

[0028] A knowledge graph construction system for the field of civil aviation services comprises a BERT-BiLSTM-CRF algorithm model, a convolutional neural network model, a conditional random field entity node integration model and a civil aviation knowledge graph model structure. The BERT-BiLSTM-CRF algorithm model has a civil aviation knowledge text database inside. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation fit on the civil aviation knowledge text database and obtains mutually related entity vector sequences, feature vector sequences and annotation sequences. The convolutional neural network model is constructed based on an attention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract a local feature matrix C=(c1, c2, ..., c n ), construct an entity-relationship-entity triple database; the conditional random field entity node integration model is constructed based on semi-supervised learning and / or weakly supervised learning methods, and the conditional random field entity node integration model integrates the annotation information in the annotation sequence. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents the feature function that extracts features from the input labeled sequence, λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, and j represents the position of the input annotation sequence; the conditional random field entity node integration model stores the annotation information in the integrated annotation sequence in the entity and uses it as the attribute value of the entity; the civil aviation knowledge graph model structure uses the entity-relationship-entity of the triple database and the attribute value of the integrated entity to link and fuse and obtain the civil aviation knowledge graph.

[0029] A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for constructing a knowledge graph in the field of civil aviation services of the present invention.

[0030] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0031] (1) The present invention extracts entity vectors and annotations through the BERT-BiLSTM-CRF algorithm model, and then uses a convolutional neural network model including an attention mechanism to identify the entity correspondence in the sentence vector and construct entity-relationship-entity triples. The conditional random field entity node integration model is used to integrate the annotation information and store it as the attribute value of the entity, and then the linking and fusion are used to obtain the civil aviation knowledge graph.

[0032] (2) The present invention can obtain a comprehensive and accurate civil aviation knowledge graph of the relationship between entities based on the civil aviation knowledge text database, which is convenient for the subsequent use of the civil aviation knowledge graph to provide valuable reference information for civil aviation managers and decision makers through data analysis, assisting them in making scientific decisions. The structured data of the civil aviation knowledge graph can also be used for in-depth data mining to discover new knowledge and rules hidden behind the data, thereby promoting the innovative development of the industry.

[0033] (3) The civil aviation knowledge graph of the present invention covers multi-level and multi-dimensional information on professional terms of civil aviation services and the general conditions of transportation of each airline, which can realize efficient integration and sharing of information and provide technical support for the high-quality development of civil aviation services. The civil aviation knowledge graph of the present invention can not only meet the needs of passengers for knowledge questions and inquiries, but also provide important data support for knowledge learning, retrieval and query for civil aviation staff. It can also be used as a training and education resource to provide systematic learning materials for practitioners and improve the overall service level. At the same time, it not only enhances the efficiency of information utilization, but also promotes the development of various industries and lays a solid foundation for continuous innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A method flow chart of the method for constructing a knowledge graph in the field of civil aviation services of the present invention;

[0035] Figure 2 This is a schematic diagram of a portion of entities enumerated in step S1 of the embodiment;

[0036] Figure 3 A schematic diagram of a sequence of annotations corresponding to some entities enumerated in step S1 of the embodiment;

[0037] Figure 4 This is a schematic diagram showing the relationship of a portion of the content of the civil aviation knowledge graph captured in the embodiment;

[0038] Figure 5 This is a principle structure diagram of the system for constructing a knowledge graph in the field of civil aviation services according to the present invention;

[0039] Figure 6 This is a usage scenario diagram of the civil aviation knowledge graph for civil aviation data retrieval in the embodiment. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below in conjunction with embodiments:

[0041] Example

[0042] like Figure 1 As shown, a method for constructing a knowledge graph in the field of civil aviation services includes:

[0043] S1. Construct a BERT-BiLSTM-CRF algorithm model and a civil aviation knowledge text database. The civil aviation knowledge text data of the civil aviation knowledge text database is derived from text data such as civil aviation laws, regulations, notices, and management regulations (of course, it also includes text data obtained from websites or web page data). It covers texts of multi-level and multi-dimensional information such as professional terms for civil aviation services and general transportation conditions of various airlines, and extracts entity nodes using the present invention. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation compatibility (entities are correspondingly associated with the civil aviation field) from the civil aviation knowledge text database and obtains mutually related entity vector sequences, feature vector sequences, and annotation sequences. The entity vector sequences, feature vector sequences, and annotation sequences are all arranged in sequence with entity nodes as key associated nodes (that is, the entity vector sequences, feature vector sequences, and annotation sequences all include entity nodes and are arranged in sequence according to the entity nodes to form a sequence set). In this embodiment, a total of nearly 20,000 entity nodes (referred to as entities, some entities are exemplified as Figure 2As shown in the figure, the relationship between them provides basic data and relationship label information support for the subsequent construction of the civil aviation knowledge graph in the civil aviation field. The entity vector sequence of the BERT-BiLSTM-CRF algorithm model is arranged in sequence according to the entity vector. The feature vector in the feature vector sequence is the feature corresponding to the entity vector extracted from the text data and arranged in sequence according to the feature vector; the annotation sequence includes the semantic dependency information between entities, and is also sorted according to the entity nodes to form a sequence set.

[0044] In some embodiments, the BERT-BiLSTM-CRF algorithm model includes a BERT masked language model, a BiLSTM network and a conditional random field CRF module. The BERT masked language model uses a masked language method to generate a deep bidirectional language representation for the civil aviation knowledge text data in the civil aviation knowledge text database (mining out word vectors with deep bidirectional up and down associations as the basis for further screening of subsequent entity vectors). The BiLSTM network uses forward and reverse methods to capture bidirectional long-distance semantic dependencies of information (the BiLSTM network has a stronger ability to extract previous and next context feature information and can capture bidirectional long-distance semantic dependencies of information). The conditional random field CRF module extracts entity nodes (selects entity nodes from word vectors) to obtain a sequence of mutually related entity vectors, and uses a set of feature functions (including several feature functions) to obtain a sequence of mutually related feature vectors and a labeling sequence.

[0045] S2. Construct a convolutional neural network model based on the attention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract the local feature matrix C = (c1, c2, ..., c n), construct an entity-relationship-entity triple database. Preferably, the sentence vectors and the entity vectors contained in the civil aviation knowledge text database are extracted using sliding windows respectively, and the entity vectors contained in the sentence vectors are extracted by using n sliding windows corresponding to the filters, so that the entity vectors in the sentence vectors are extracted in sequence using sliding windows; preferably, the subject, predicate and object entities are also emphasized in the sentence vectors. An attention mechanism module is embedded in the convolutional neural network model of the present invention to further extract the relationship between the entity vectors in the sentence vectors. The entity vector sequence strategy in the sentence vector is built on the attention mechanism, and a pure, noise-free combined vector expression can be obtained. The present invention constructs a civil aviation knowledge graph model structure to assist in the embedding editing and optimization processing of the civil aviation knowledge graph. First, each sentence is converted into a feature vector (with entity vector as the core) through a convolutional neural network model. The feature vector corresponding to the sentence with the highest support is identified and gradually accumulated. Preferably, a specific set is constructed, which contains all kinds of combinations of sentence feature vectors involving the same entity pair. Then, the ability score of each set of feature vectors mapped to the relationship label under the multi-category classification task is calculated to identify the feature vector combination with the highest score. The feature vector set with the highest score integrates the semantic information of the positive examples and the potential noise sentence features, which provides a solid foundation for the present invention to train the classifier for the target entity pair, thereby constructing a comprehensive and appropriate model framework for the relationship extraction task. When extracting entity relationships, there are differences in the length of each sentence, which results in the ability of civil aviation service field text to reveal the relationship between entity nodes at any position in the sentence (see Figure 3 As shown, some entity correspondences and annotation sequences are listed); therefore, the present invention can adopt a strategy of extracting various local features from sentences to infer the corresponding relationship types of civil aviation service entities, and introduce convolution operations to process these extracted local features.

[0046] In some embodiments, the local feature matrix C=(c1, c2, ..., c n ) also includes the corresponding weight matrix W∈R w*d (During the convolution operation, a random initialization weight can be defined first), where R represents the real number system, w represents the sliding window size, and d represents the dimension after embedding connection; the sentence vector is recorded as a sequence (q1, q2, ...q i , …, q m ), the sequence is a sequence with entity vector as the core, q i The data representing position i in the sentence is convolved according to the following formula to obtain the sequence:

[0047] c i =Wq i-w+1：i +b,Wq i-w+1：iIndicates q i-w+1 to q i The weight of , b represents the bias vector (the formula of the present invention contains the bias vector b, which can help the model better fit the training data, of course, the formula may not contain the bias vector b), c i Represents the local features of position i in the sentence (this will get all the local features in the sentence vector, use multiple different filters to complete the local feature extraction, the convolution layer finally outputs a matrix set (c1, c2, ..., c n ), can further find all positive instances that reflect label relationships), and extract the local feature matrix of the entity correspondence relationship in the sentence vector with the entity in the sentence vector as the core. Due to the use of the attention mechanism in the present invention, when the sentence length is not enough to meet the convolution window requirements or the sentence is too long, which may cause information loss, the attention mechanism module can identify and focus on the key parts of the sentence, thereby better capturing the label relationship when constructing the feature vector without strictly relying on the sliding window size.

[0048] In some embodiments, after obtaining the entity vectors in each sentence vector, the present invention further includes the following method:

[0049] A. Use the cosine similarity method to calculate the similarity of entities in different sentence vectors (after information processing, the data is stored in the data layer and the pattern layer. Although the entities are extracted from the unstructured and semi-structured data in the civil aviation service field, there is still a lot of redundant and erroneous information. Data aggregation processing is required to eliminate redundancy and errors to improve the accuracy of entity-relationship-entity. The process of entity linking is to select a group of pending entities from the civil aviation service field entities, and calculate the similarity through the attribute value to achieve the purpose of linking to the correct entity object):

[0050] Where similarity(e, c) represents the judgment result of the similarity between e and c, e represents the selected entity, and c represents the entity used for similarity comparison, that is, the candidate entity. represents the vector of entity e, A vector representing entity c;

[0051] B. Entity c with the highest similarity and / or entity e is subjected to attribute value and relationship fusion processing (if the relationship between two entities is represented differently in multiple data sources, the fusion can be performed by calculating the consistency of the relationship).

[0052] The further method is as follows: first, determine a selected entity e, and screen out an entity set for similarity comparison (referred to as the entity set to be compared) based on the annotation information in the annotation sequence in the BERT-BiLSTM-CRF algorithm model. In the entity set to be compared, the cosine similarity method is used to calculate the similarity of each similarity comparison entity C in turn, and each similarity comparison entity C whose similarity judgment is greater than the similarity threshold is selected, and its attribute value and relationship are merged with the entity e (for multiple entity words with the same meaning); or the entity c with the highest similarity judgment is selected to fuse the attribute value and relationship with the entity e (for two entity words with the same meaning).

[0053] S3. Construct a conditional random field entity node integration model based on semi-supervised learning or / and weakly supervised learning. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents a feature function for extracting features from the input annotation sequence (in step S1, a feature function set including multiple feature functions is used for feature extraction, f k is the kth characteristic function in the characteristic function set), λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, and j represents the position of the input annotation sequence; the annotation information in the integrated annotation sequence is stored in the entity and serves as the attribute value of the entity.

[0054] In some embodiments, the model update expression using the semi-supervised learning method in the conditional random field entity node integration model is as follows:

[0055] P(y|x; θ) = Nodel(x; θ), where θ represents the model parameters; the above method is used to identify the attributes corresponding to the entity node, and labeling refers to assigning a known target value or label to the sample in the entity node. In the supervised learning process, the model learns the relationship between the input data (features) and its corresponding output label, so that it can predict whether it fits the entity node when encountering new data. On this basis, part-of-speech tagging is performed to help identify the descriptive attributes of the entity noun. A semi-supervised learning model is used to train the model using some labeled data and a large amount of unlabeled data; the common method in this step is self-training. The basic idea is to first train a basic model with labeled data, first use the model to predict unlabeled data, and then use a part of the unlabeled data with accurate prediction results for a second training to update the semi-supervised learning model so that it is more suitable for civil aviation data text, and select data with high prediction confidence from the unlabeled training results and continuously update and iterate.

[0056] Self-training semi-supervised learning method, model update can be simplified as follows.

[0057] θ new =argmaxlogP(Y|X;θ1),θ new represents the model parameters after iterative update, θ1 represents the model parameters before iterative update, X represents the original annotated data set, and Y represents the integrated annotated data set. Semi-supervised learning is suitable for training data with high labeling accuracy. In the process of automatic labeling, it can be used for data text annotation of knowledge graphs in the civil aviation field, and then it can be connected with the conditional random field CRF module. In this link, the conditional random field CRF module can be used again to score the obtained entity attributes, and the final score is obtained by weighted summation, so as to obtain attributes with higher fit with the corresponding entity and discard attributes with lower scores.

[0058] S4, link and fuse the entity-relationship-entity of the triple database in step S2 and the attribute values ​​of the integrated entity to obtain the civil aviation knowledge graph (see Figure 3 , intercept and list some entity relationship annotations for relationship display). See Figure 4 , intercepting a part of the civil aviation knowledge graph, which contains nearly 20,000 entities and multiple inter-entity links (as well as annotation information, etc.), which facilitates knowledge sharing among civil aviation personnel or other personnel in other fields and efficient information integration or retrieval processing of civil aviation data (such as Figure 6As shown, the civil aviation knowledge graph of the present invention is used for rapid retrieval of civil aviation databases), which is convenient for users to quickly know the relationship between civil aviation entity nodes and links. The civil aviation knowledge graph not only provides civil aviation managers and passengers with references to civil aviation knowledge relationships and information summaries, which is convenient for assisting scientific decision-making, but also can discover hidden knowledge and laws, and promote the innovative development of the industry; at the same time, it can also be used as a training and education resource to provide practitioners with systematic learning materials and improve the overall service level. The present invention not only enhances the efficiency of information utilization, but also promotes the development of various industries and lays a solid foundation for continuous innovation.

[0059] The present invention constructs entity-relationship-entity triples through multi-level integration and entity linking of entities, and further constructs a knowledge graph using the Neo4j graph database. Through the Cypher query language, we have successfully imported nearly 20,000 entity nodes and the relationships between them. Compared with traditional relational databases, Neo4j is more suitable for representing complex entity (node) and relationship (edge) structures. Cypher, as the query language of Neo4j, is designed specifically for interacting with graph data and can better express queries on graph structures.

[0060] In some embodiments, the present invention further comprises the following method:

[0061] S5. Continuously update and collect civil aviation knowledge data and build a civil aviation knowledge database. The user inputs keywords or entity words corresponding to the entity, and obtains the civil aviation knowledge graph of each entity in the civil aviation knowledge database according to steps S1 to S4. The civil aviation knowledge graph is used to perform search and analysis on the civil aviation knowledge database for keywords or entity words corresponding to the entity, and obtains the search results corresponding to the entity. Figure 6 Taking the entity word of ticket as an example, the search results are partially intercepted as follows Figure 6 shown.

[0062] like Figure 5 As shown, a knowledge graph construction system for the field of civil aviation services includes a BERT-BiLSTM-CRF algorithm model, a convolutional neural network model, a conditional random field entity node integration model, and a civil aviation knowledge graph model structure. The BERT-BiLSTM-CRF algorithm model has a civil aviation knowledge text database inside. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation fit from the civil aviation knowledge text database and obtains interrelated entity vector sequences, feature vector sequences, and annotation sequences. The convolutional neural network model is constructed based on an attention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract the local feature matrix C of the entity correspondence relationship in the sentence vector = (c1, c2, ..., c n), construct an entity-relationship-entity triple database; the conditional random field entity node integration model is constructed based on semi-supervised learning and / or weakly supervised learning methods, and the conditional random field entity node integration model integrates the annotation information in the annotation sequence. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents the feature function that extracts features from the input labeled sequence, λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, and j represents the position of the input annotation sequence; the conditional random field entity node integration model stores the annotation information in the integrated annotation sequence in the entity and uses it as the attribute value of the entity; the civil aviation knowledge graph model structure uses the entity-relationship-entity of the triple database and the attribute value of the integrated entity to link and fuse and obtain the civil aviation knowledge graph.

[0063] A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for constructing a knowledge graph in the field of civil aviation services of the present invention.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a knowledge graph in the field of civil aviation services, characterized by: The methods include: S1. Construct a BERT-BiLSTM-CRF algorithm model and a civil aviation knowledge text database. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation fit from the civil aviation knowledge text database and obtains interrelated entity vector sequences, feature vector sequences, and annotation sequences. S2. Construct a convolutional neural network model based on the attention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract the local feature matrix C = (c1, c2, ..., c n ), construct a triple database of entity-relationship-entity; S3. Construct a conditional random field entity node integration model based on semi-supervised learning or / and weakly supervised learning. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents the feature function that extracts features from the input labeled sequence, λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, where j represents the position of the input annotation sequence; the annotation information in the integrated annotation sequence is stored in the entity and serves as the attribute value of the entity; S4. Use the entity-relationship-entity of the triple database in step S2 and the attribute values ​​of the integrated entity to link and fuse to obtain a civil aviation knowledge graph.

2. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1, characterized in that: Also includes the following methods: A. Use the cosine similarity method to calculate the similarity of entities in different sentence vectors: Where similarity(e, c) represents the judgment result of the similarity between e and c, e represents the selected entity, and c represents the entity used for similarity comparison, that is, the candidate entity. represents the vector of entity e, A vector representing entity C; B. Entity c with a similarity greater than the similarity threshold or / and the highest similarity is merged with entity e in terms of attribute values ​​and relationships.

3. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1, characterized in that: The sentence vectors and the entity vectors contained in the civil aviation knowledge text database are extracted using sliding windows respectively, and the entity vectors contained in the sentence vectors are extracted using n sliding windows corresponding to the filters.

4. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1 or 3, characterized in that: The local feature matrix also includes the corresponding weight matrix W∈R w*d , where R represents the real number system, w represents the sliding window size, and d represents the dimension after embedding concatenation; the sentence vector is recorded as a sequence (q1, q2, ...q i , ..., q m ), q i The data representing position i in the sentence is convolved according to the following formula to obtain the sequence: c i =Wq i-w+1：i +b,Wq i-w+1 :i represents q i-w+1 to q i The weight of b represents the bias vector, C i Represents the local features of position i in the sentence, And taking the entities in the sentence vector as the core, a local feature matrix of the entity correspondence in the sentence vector is extracted.

5. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1, characterized in that: The model update expression using the semi-supervised learning method in the conditional random field entity node integration model is as follows: P(y|x; θ) = Mode1(x; θ), where θ represents the model parameters; θ new =argmaxlogP(Y|X;θ1),θ new represents the model parameters after iterative update, θ1 represents the model parameters before iterative update, X represents the original labeled data set, and Y represents the integrated labeled data set.

6. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1, characterized in that: The entity vector sequence of the BERT-BiLSTM-CRF algorithm model is arranged in sequence according to the entity vectors. The feature vectors in the feature vector sequence are features corresponding to the entity vectors extracted from the text data and arranged in sequence according to the feature vectors. The annotation sequence includes semantic dependency information between entities.

7. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1 or 6, characterized in that: In step S1, the BERT-BiLSTM-CRF algorithm model includes a BERT masked language model, a BiLSTM network and a conditional random field CRF module. The BERT masked language model uses a masked language method to generate a deep bidirectional language representation for the civil aviation knowledge text data in the civil aviation knowledge text database. The BiLSTM network uses forward and reverse methods to capture bidirectional long-distance semantic dependencies of information. The conditional random field CRF module extracts entity nodes to obtain mutually related entity vector sequences, and uses a set of feature functions to obtain mutually related feature vector sequences and annotation sequences.

8. The method for constructing a knowledge graph in the field of civil aviation services according to claim 1, characterized in that: Also includes the following methods: S5. Continuously update and collect civil aviation knowledge data and build a civil aviation knowledge database. The user inputs keywords or entity words corresponding to the entity, and obtains the civil aviation knowledge graph of each entity in the civil aviation knowledge database according to steps S1 to S4. The civil aviation knowledge graph is used to search and analyze the keywords or entity words corresponding to the entity in the civil aviation knowledge database, and the search results corresponding to the entity are obtained.

9. A knowledge graph construction system for civil aviation service field, characterized by: The invention comprises a BERT-BiLSTM-CRF algorithm model, a convolutional neural network model, a conditional random field entity node integration model and a civil aviation knowledge graph model structure. The BERT-BiLSTM-CRF algorithm model has a civil aviation knowledge text database. The BERT-BiLSTM-CRF algorithm model extracts entities with high civil aviation fit from the civil aviation knowledge text database and obtains mutually related entity vector sequences, feature vector sequences and annotation sequences. The convolutional neural network model is constructed based on the intention mechanism. The convolutional neural network model extracts sentence vectors and entity vectors contained in the civil aviation knowledge text database and uses n filters to identify and extract the local feature matrix C=(c1, c2, ..., c n ), construct an entity-relationship-entity triple database; the conditional random field entity node integration model is constructed based on semi-supervised learning and / or weakly supervised learning methods, and the conditional random field entity node integration model integrates the annotation information in the annotation sequence. The formula is as follows: Where x represents the labeled sequence taken as input in step S1, y represents the integrated labeled sequence, Z(x) represents the normalization factor, and f k represents the feature function that extracts features from the input labeled sequence, λ k Denotes the characteristic function f k The weight of y j-1 Indicates the j-1th position labeled data, y j represents the j-th position annotation data, and j represents the position of the input annotation sequence; the conditional random field entity node integration model stores the annotation information in the integrated annotation sequence in the entity and is the attribute value of the entity; the civil aviation knowledge graph model structure uses the entity-relationship-entity of the triple database and the attribute value of the integrated entity to link and fuse and obtain the civil aviation knowledge graph.

10. A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.

Images