A method and system for handling data link network assurance requirements

Abstract

The application provides a data chain network guarantee demand processing method and system, and the method comprises the following steps: setting a database and integrating a plurality of AI algorithm models based on service calls, so as to realize knowledge extraction, knowledge recommendation and knowledge query of data chain network guarantee demand; the database comprises a plurality of guarantee schemes and sample knowledge graphs corresponding to the guarantee schemes; the AI algorithm models at least comprise a knowledge extraction model, a knowledge recommendation model and a knowledge query model; each AI algorithm model is encapsulated in the form of a service; a process pool is created, and one process in the process pool corresponds to loading one AI algorithm model; an idle process in the process pool is selected to call the AI algorithm model through a service request, and the calling result of the AI algorithm model is fed back through a service response. The application integrates extraction, recommendation and query, provides intelligent auxiliary capacity for data chain network design and planning personnel when collecting and analyzing data chain network guarantee demand, and effectively improves work efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a data link network guarantee demand processing method and system. BACKGROUND

[0002] Data link is a key system for realizing deep connection of command and control platform, weapon platform and sensor platform, and is a core system for improving the agility and precision of the "detection, control, attack and evaluation" operational loop in joint operations. Unlike general military wireless communication systems, data link is closely related to operational actions, and is a communication system for a specific operational task. In order to provide real-time situation information sharing and rapid operational capabilities from sensors to weapon platforms, communication demand analysis and scientific and fine network design planning need to be performed before the start of operational actions according to specific operational plans. Whether the data link network design planning can be efficiently and accurately completed is the key to realizing data link system networking and operational use, and is the premise for improving the overall combat capability of the operational system and the operational effectiveness of the weapon system by using data link systems. The collection and analysis of data link network guarantee demand as the starting point of data link network design planning is an extremely important link in data link network design planning.

[0003] At present, some experts propose to use LSTM neural networks to automatically extract data link network guarantee demands from operational plan documents, but since data is the main factor affecting the performance of data-driven intelligent algorithms, and the data in the field of intelligent extraction of data link network guarantee demands is relatively small, directly using existing knowledge extraction methods will cause insufficient data and loss of data link network performance.

[0004] Some other experts propose to use a graph similarity calculation method to realize intelligent recommendation of data link network guarantee schemes. This method does not design a corresponding graph similarity calculation method for the characteristics of the data link network guarantee demand knowledge graph, resulting in the problem of low recommendation accuracy when using a general graph similarity calculation method.

[0005] Some other experts propose to use a machine reading model method to provide intelligent query capability of data link network guarantee demands. This method can only find answers in the text through extraction. If the answer does not exist in the text or cannot be directly obtained, the answer can be obtained only after statistical, inductive and reasoning processing based on the text information content. Therefore, the correct answer cannot be obtained by using this method.

[0006] In summary, the existing data link network support demand acquisition and analysis only considers part of the data link network support demand processing, such as single processing mode of intelligent extraction, intelligent recommendation and intelligent query, and lacks system design of integration of intelligent extraction, intelligent recommendation and intelligent query; so that the data link network design planner lacks effective intelligent assistance when collecting and analyzing data link network support demand, and the design planning efficiency is low. SUMMARY

[0007] The present application provides a data link network support demand processing method and system to solve the problem that the existing data link network design planner lacks effective intelligent assistance when collecting and analyzing data link network support demand, and the design planning efficiency is low.

[0008] In one aspect, the present application provides a data link network support demand processing method, which comprises: setting a database and integrating a plurality of AI algorithm models based on service calls to realize knowledge extraction, knowledge recommendation and knowledge query of data link network support demand.

[0009] The database comprises a plurality of support schemes and sample knowledge graphs corresponding to the support schemes; the AI algorithm model at least comprises a knowledge extraction model, a knowledge recommendation model and a knowledge query model; each AI algorithm model is encapsulated in the form of a service;

[0010] A process pool is created, and each process in the process pool corresponds to loading an AI algorithm model; an idle process in the process pool is selected to call the AI algorithm model through a service request, and the calling result of the AI algorithm model is fed back through a service response.

[0011] Further, the knowledge extraction model is used to receive a target document, extract demand elements from the target document, and form a target knowledge graph.

[0012] The construction of the knowledge extraction model comprises: extracting entities from the target document by using an existing extraction model to obtain triple data and generate a triple knowledge base; using a graph representation learning method to generate entity vector encoding corresponding to each entity in the triple knowledge base; using the entity vector encoding as a supplementary enhancement means of entity semantic information to improve the existing extraction model to obtain the knowledge extraction model.

[0013] Further, the knowledge extraction method comprises:

[0014] The target document is received, each text paragraph is read, a text paragraph list is generated, and the text paragraph list is divided into a plurality of equal parts to obtain a plurality of text paragraph sub-lists;

[0015] sequentially traversing a plurality of processes in a process pool, passing a text paragraph sub-list to a process, calling the knowledge extraction model by the current process to extract sub-list demand elements from the current text paragraph sub-list;

[0016] After waiting for all processes in the process pool to complete the extraction, splicing the triple data lists of all sub-list demand elements to obtain the demand elements of the target document.

[0017] Further, the knowledge recommendation model is used to calculate the similarity values of the target knowledge graph and each sample knowledge graph in the database, and output the guarantee scheme corresponding to the sample knowledge graph with the highest similarity value.

[0018] Further, calculating the similarity values of the target knowledge graph and each sample knowledge graph in the database includes:

[0019] Based on the basic framework of the SimGNN model, the edges in the knowledge graph are set as special nodes in the knowledge graph; and based on the graph topology difference features of the SimGNN model, combat area difference features and participating force difference features are added.

[0020] Further, adding the combat area difference features and the participating force difference features includes:

[0021] Inputting the target knowledge graph and any sample knowledge graph, performing first path processing, second path processing and third path processing;

[0022] The first path processing includes: edge-to-node processing through the edges in the knowledge graph, edge encoding to add relationship information on the edges, inputting the SimGNN model to obtain the topology structure similarity between the two knowledge graphs based on the SimGNN model;

[0023] The second path processing includes: extracting geographic information in the knowledge graph, setting any geographic location as a reference point, calculating the distance between each geographic location in the knowledge graph and the reference point to obtain the distance vector of the reference point; and inputting the distance vector into the Transformer network to extract the geographic spatial information of the two knowledge graphs to obtain the geographic spatial similarity between the two knowledge graphs;

[0024] The third path processing includes: extracting force information in the knowledge graph, classifying and counting the participating force quantity of each type of force to obtain a force quantity vector; and inputting the force quantity vector into the Transformer network to obtain the force similarity between the two knowledge graphs;

[0025] The topology similarity, the geographical space similarity and the force similarity are input into a Dense network to obtain a similarity value of the target knowledge graph and the sample knowledge graph.

[0026] Further, the knowledge query model is used for receiving a natural language question sentence of a user, converting the natural language question sentence into a query language of the database, executing the query language on the database, generating an answer of the natural language question sentence, and completing the guarantee scheme and the query of the corresponding sample knowledge graph.

[0027] Further, converting the natural language question sentence into the query language of the database comprises:

[0028] A template library is established, and the template library comprises standard question templates of various natural language question sentences.

[0029] Semantic information of the natural language question sentence is processed to obtain question semantic encoding information, and it is determined based on the semantic encoding information that the natural language question sentence belongs to which standard question template.

[0030] The natural language question sentence is extracted to obtain a plurality of key entities.

[0031] The key entities are filled into the standard question template to generate an instantiated query language of the database.

[0032] Further, filling the key entities into the standard question template comprises:

[0033] It is determined whether the key entities exist in the sample knowledge graph.

[0034] If the key entities exist, the key entities are directly filled into the standard question template.

[0035] If the key entities do not exist, an edit distance of the key entities and each sample entity in the sample knowledge graph is calculated, the key entities are replaced with a sample entity with the smallest edit distance in the sample knowledge graph, and the sample entity is filled into the standard question template.

[0036] On the other hand, the application further provides a data link network guarantee demand intelligent processing system, which at least comprises a database, a knowledge extraction module, a knowledge query module and a knowledge recommendation module, and is used for executing steps of any one of the methods.

[0037] Overall, compared with the prior art, the technical solution conceived by the application can achieve the following beneficial effects:

[0038] (1) The application provides a data link network guarantee demand processing method and system, which integrates three modules of intelligent extraction, intelligent recommendation and intelligent query, provides intelligent auxiliary capability for data link network design and planning personnel when collecting and analyzing data link network guarantee demand, and effectively improves the work efficiency of data link network design and planning personnel. In addition, the operation efficiency of the AI model is improved through multi-core concurrency, and the operation efficiency of the system is effectively improved through the shared process pool mode.

[0039] (2) The application provides a data link network guarantee demand processing method and system, which utilizes a knowledge enhanced knowledge extraction model, improves the performance of the subsequent knowledge extraction model through the knowledge formed by the knowledge extraction, and improves the performance of the knowledge extraction model.

[0040] (3) The application provides a data link network guarantee demand processing method and system, which realizes intelligent knowledge recommendation through hybrid graph embedding coding, so that the output guarantee scheme is more rapid and accurate.

[0041] (4) The application provides a data link network guarantee demand processing method and system, which realizes knowledge query of data link network guarantee demand based on template matching and key information extraction, can convert user query into database query statements, and generates answers to user queries after executing the database query statements, not only solves the phenomenon that the answer cannot be given when the answer is not in the text, but also can use the database query statement to explain and trace the generated answer, so as to have good explainability and improve the credibility of the user to the answer. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0043] Figure 1 is a construction schematic diagram of a knowledge extraction model of the data link network guarantee demand processing method and system provided by the application;

[0044] Figure 2 is a principle schematic diagram of an existing relationship classification model of the data link network guarantee demand processing method and system provided by the application;

[0045] Figure 3 is a principle schematic diagram of a relationship classification model of the data link network guarantee demand processing method and system provided by the application;

[0046] Figure 4is a knowledge extraction process schematic diagram of a data link network guarantee demand processing method and system provided by the application;

[0047] Figure 5 is a knowledge recommendation process schematic diagram of a data link network guarantee demand processing method and system provided by the application;

[0048] Figure 6 is a similarity calculation schematic diagram of a data link network guarantee demand processing method and system provided by the application;

[0049] Figure 7 is a specific embodiment schematic diagram of a knowledge query model of a data link network guarantee demand processing method and system provided by the application;

[0050] Figure 8 is a system structure schematic diagram of a data link network guarantee demand processing method and system provided by the application. DETAILED DESCRIPTION

[0051] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings and embodiments in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] It should be noted that in the description of the embodiments of the present application, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method or system. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of another identical element in the process, method or system comprising the element.

[0053] The present application provides a data link network guarantee demand processing method, which comprises the following steps: setting a database and integrating a plurality of AI algorithm models based on service calls to realize knowledge extraction, knowledge recommendation and knowledge query of data link network guarantee demand.

[0054] The database comprises a plurality of guarantee schemes and sample knowledge graphs corresponding to the guarantee schemes. In the formation process of each guarantee scheme in the database, the guarantee scheme is extracted to generate a graph, preferably based on the knowledge extraction model mentioned later.

[0055] The AI algorithm model at least includes a knowledge extraction model, a knowledge recommendation model and a knowledge query model; and each AI algorithm model is encapsulated in the form of a service.

[0056] It should be noted that the data chain network support demand knowledge extraction tool automatically extracts demand elements from the target document by calling the knowledge extraction model, and constructs a data chain network support demand knowledge graph; wherein the demand elements include eight types of data chain support demand elements, such as task type, force allocation, command relationship, coordination relationship, force deployment, combat area, battlefield environment and action plan. The data chain network support demand knowledge recommendation tool automatically calculates the matching degree of the target document and each data chain network support demand knowledge graph in the database by calling the knowledge recommendation model, and realizes intelligent recommendation of the best data chain network support scheme in the database. The data chain network support demand knowledge query tool realizes intelligent question and answer on the data chain network support demand knowledge graph by calling the knowledge query model, and provides flexible database query and management capabilities for data chain network design and planning personnel.

[0057] In the data chain network support demand intelligent processing method, multiple AI algorithm models are used to realize data chain network support demand knowledge extraction, knowledge query and knowledge recommendation. Since the underlying libraries relied on by each AI algorithm model are more and different, integrating them together may cause conflicts of underlying neural network modules, so in the integration design, the deployment process is complex, and the AI algorithm model integration based on service calling is adopted, each AI algorithm model is encapsulated in the form of a service, the running environment of each AI algorithm model is isolated, and each AI algorithm model runs in its own service process space, thereby avoiding various conflict problems that may occur in the same process space.

[0058] In addition, as an embodiment of the present application, a B / S mode structure design is adopted to simplify the software system deployment difficulty; the system interface is presented in the form of relying on a browser, further simplifying the difficulty of interface development work.

[0059] A process pool is created, and one process in the process pool corresponds to loading one AI algorithm model; the AI algorithm model is called by selecting an idle process from the process pool through a service request, and the calling result of the AI algorithm model is fed back through a service response.

[0060] The data chain network support demand intelligent processing method needs to use multiple AI algorithm models based on deep neural networks, resulting in large amount of calculation and long time consumption. As an embodiment of the present application, the multi-core concurrency mode can be used to improve the running efficiency of the AI model; for example, a 64-core domestic Feiteng 2000+ CPU is used to improve the performance through the advantage of CPU multi-core concurrency.

[0061] In addition, the initialization of the neural network and the import of the network parameter file are required in the loading process of the AI algorithm model, and the process requires a time consumption of seconds. Therefore, the shared process pool is adopted, and the time consumption problem of the initialization of the AI algorithm model is eliminated when the AI algorithm model is called. Specifically, the shared process pool is created at the start, each process in the process pool loads an AI algorithm model, and the initialization of the AI algorithm model is completed. Thereafter, each time the AI algorithm model is called, only the idle process in the shared process pool is selected, so that the running efficiency is effectively improved.

[0062] The knowledge extraction model is configured to receive a target document, extract a requirement element from the target document, and form a target knowledge graph.

[0063] In the field of data link network support demand knowledge extraction, there is a problem of insufficient data, which leads to the loss of accuracy of the existing data-driven knowledge extraction model. In order to solve this problem, the knowledge extraction model is improved by using the knowledge extraction model, and the entity pairs and their relationships in the triple knowledge base formed after knowledge extraction have objective meaning.

[0064] Specifically, the construction of the knowledge extraction model is as shown in Figure 1 Firstly, an existing extraction model is used to extract entities from a target document to obtain triple data and generate a triple knowledge base. Then, a graph representation learning method such as TransE is used to generate entity vector encoding corresponding to each entity in the triple knowledge base. Finally, the entity vector encoding is used as a supplementary enhancement means of entity semantic information to improve the existing extraction model and obtain a knowledge extraction model.

[0065] It should be noted that the existing extraction model usually adopts two steps of named entity recognition and relationship classification.

[0066] The named entity recognition model is responsible for extracting an entity pair, i.e. a subject word and an object word, from the text, and then using the relationship classification model to determine the relationship between the entity pair in the current text environment.

[0067] For example, in the text “the air force early warning aircraft and the fighter aircraft, and the naval missile destroyer and the mine sweeper form command relationship and cooperative relationship respectively”, after using the named entity recognition model, two entity pairs are obtained, which are [early warning aircraft, fighter aircraft] and [missile destroyer, mine sweeper]; then, the relationship classification model is used to determine the relationship between the two entity pairs.

[0068] As shown in Figure 2As shown, the existing relationship classification model mainly determines the relationship of an entity pair according to the semantic information of the text where the entity pair is located, the semantic information of the subject word and the semantic information of the object word when determining the relationship of an entity pair. Among them, the semantic information of the text can be obtained by splicing the word vector, character vector and part-of-speech vector in the text, or using pre-trained language models such as Google BERT and Baidu ERNIE to extract the semantic information of the text; and then using CNN convolutional neural network and LSTM long short-term memory neural network processing. The semantic information of the subject word and the semantic information of the object word are usually directly expressed by the semantic information of the text at the position of the subject word and the object word in the text.

[0069] However, in the text, the air force early warning aircraft and the fighter aircraft, and the naval missile destroyer and the mine sweeper form a command relationship and a cooperative relationship respectively, and the existing relationship classification model is difficult to accurately determine the relationship between the two entity pairs [early warning aircraft, fighter aircraft] and [missile destroyer, mine sweeper], because the two relationship words "command relationship" and "cooperative relationship" are closely together, and only the word "separately" indicates that the two relationships correspond to the two entity pairs. But if the knowledge in the knowledge base formed before is combined, a common sense knowledge is obtained, that is, the early warning aircraft and the fighter aircraft cannot form a cooperative relationship, and the missile destroyer and the mine sweeper cannot form a command relationship.

[0070] Therefore, the present application proposes to improve the relationship classification model, such as Figure 3 As shown, using TransE and other graph representation learning methods, the knowledge vectors of each entity in the knowledge base are encoded, and the entity vector encoding information is used as a supplement to the semantic information of the entity to realize the knowledge-enhanced relationship classification model. In this way, it can be easily determined that the early warning aircraft and the fighter aircraft are in a command relationship, and the missile destroyer and the mine sweeper are in a cooperative relationship.

[0071] In addition, in order to fully utilize the multi-core concurrency performance and improve the performance of the data link network support demand knowledge extraction model, the method for knowledge extraction comprises the following steps: receiving a target document, reading each text paragraph, generating a text paragraph list, and dividing the text paragraph list into multiple equal parts to obtain multiple text paragraph sub-lists; sequentially traversing multiple processes in a process pool, passing a text paragraph sub-list to a process, and calling a knowledge extraction model by the current process to extract a sub-list demand element from the current text paragraph sub-list; after all processes in the process pool complete the extraction, splicing the triple data list of all sub-list demand elements to obtain the demand elements of the target document.

[0072] It should be noted that the text paragraph list is divided into equal parts consistent with the number of processes in the process pool.

[0073] As a specific embodiment of the present application, the multiprocessing library of Python is called to create a process pool containing 8 processes; each process in the process pool loads a knowledge extraction model respectively, and completes the initialization work of loading related deep learning model and network parameters. Then, as shown in Figure 4 The request of the user for calling the knowledge extraction model is received, the Python-docx library is called to read each text paragraph from the docx format target document imported by the user to generate a text paragraph list; and the text paragraph list read from the target document is divided into 8 equal parts, i.e. 8 text paragraph sublists are formed; the 8 processes in the process pool are traversed in turn, one of the 8 text paragraph sublists is passed to one process, and the process calls the knowledge extraction model to extract the requirement elements from the text paragraph sublist; after waiting for all the processes in the process pool to complete the extraction task, the triple data lists of the requirement elements extracted by each process are spliced together and returned to the knowledge extraction module.

[0074] Further, an idle process is selected from the process pool through a service request. Specifically:

[0075] It is determined whether there is an idle process; if yes, the current idle process is called to complete the extraction of requirement elements from one text paragraph sublist; if no, it is determined whether there is a task of incomplete extraction, if there is a task, the waiting continues until there is no task; if there is no task, the triple data lists of the requirement elements extracted by each process are spliced together and fed back to the knowledge extraction module.

[0076] The knowledge recommendation model is used to calculate the similarity value of the target knowledge graph and each sample knowledge graph in the database, and output the guarantee scheme corresponding to the sample knowledge graph with the highest similarity value. In this way, the automatic calculation of the matching degree of the target document and each sample knowledge graph in the database is completed, and the intelligent recommendation of the best guarantee scheme in the database is realized.

[0077] Specifically, as shown in Figure 5 After the user inputs the target document, the target document is first processed by the knowledge extraction model to generate a target knowledge graph A; then each guarantee scheme in the database is traversed to obtain a sample knowledge graph B corresponding to the current guarantee scheme; then the similarity between the target knowledge graph A and the sample knowledge graph B is calculated by calling the knowledge recommendation algorithm; it should be noted that if the similarity between the current target knowledge graph A and the sample knowledge graph B is higher than the previous calculated similarity, it means that a more matched guarantee scheme is obtained, and based on the most matched similarity value, the guarantee scheme most matched with the target document is updated until each guarantee scheme in the database is traversed to obtain the guarantee scheme most matched with the current target document.

[0078] Currently, the calculation of the similarity of two knowledge graphs is based on a SimGNN model, and the graph embedding coding is realized through the graph topological structure information, so as to calculate the graph similarity. However, when the SimGNN model performs graph embedding coding, only the nodes in the graph are considered, and the information on the edges is not considered, so that the information is lost when the graph embedding coding is performed on the knowledge graph, and the accuracy of the graph similarity calculation is affected.

[0079] In order to more accurately calculate the similarity of two knowledge graphs, the application calculates based on a hybrid graph embedding coding knowledge recommendation model.

[0080] On the one hand, the edge coding method in the SimGNN model is improved, and specifically, based on the basic framework of the SimGNN model, the edges in the knowledge graph are set as special nodes in the knowledge graph.

[0081] In the prior art, when processing the edges in the graph, a edge is encoded by the information of two nodes on the edge, but the specific relationship information represented by the edge is discarded. For example, there is a triple <escort task, start time, January 1, 2023> in the target knowledge graph A, and there is a triple <escort task, end time, January 1, 2023> in the sample knowledge graph B. Obviously, the target knowledge graph A and the sample knowledge graph B have certain differences, one start time is January 1, 2023, and the other end time is January 1, 2023. However, when using the SimGNN model to perform graph embedding coding, the edge coding input of the target knowledge graph A is [<escort task, January 1, 2023>], and the edge coding input of the sample knowledge graph B is also [<escort task, January 1, 2023>], which will cause the coding information of the target knowledge graph A and the sample knowledge graph B to be completely consistent, and thus the similarity value calculated is also completely the same. The reason for this problem is that the SimGNN model mainly relies on node coding information, and the coding information of an edge is only composed of the coding information of the two nodes associated on the edge, and the specific relationship information on the edge is discarded, so that the above two knowledge graphs become indistinguishable.

[0082] In order to solve the problem that the target knowledge graph A and the sample knowledge graph B have differences but the coding information of the two is completely consistent, the application regards the edge as a special node in the graph on the basis of the basic framework of the SimGNN model.

[0083] For example, in the triple <escort task, end time, January 1, 2023> of the sample knowledge graph B, the relationship name "end time" is also regarded as a node in the graph, so the edge encoding represented by this triple changes from the original [<escort task, January 1, 2023>] to [<escort task, end time>, <end time, January 1, 2023>, <escort task, January 1, 2023>]; Similarly, the triple <escort task, start time, January 1, 2023> of the target knowledge graph A is encoded into [<escort task, start time>, <start time, January 1, 2023>, <escort task, January 1, 2023>]; Thus, the encoding inputs of the two knowledge graphs become different, which is more in line with the actual situation.

[0084] That is, change the edge encoding in SimGNN from [<node 1 on the edge, node 2 on the edge>] to [<node 1 on the edge, relationship name on the edge>, <relationship name on the edge, node 2 on the edge>, <node 1 on the edge, node 2 on the edge>]. It should be noted that the present application only regards the relationship name on the edge as a special node on the graph, and the encoding information is still consistent with the existing node encoding method in the graph.

[0085] On the other hand, since the existing SimGNN model only uses the graph topology structure in the knowledge graph to calculate the graph similarity. When directly applied to knowledge graph similarity calculation, it does not involve domain knowledge, which will lead to poor accuracy of similarity calculation.

[0086] Therefore, the present application adds the combat area difference feature and the participating force difference feature on the basis of the graph topology structure difference feature of the SimGNN model.

[0087] Specifically, as shown in Figure 6 The addition of the combat area difference feature and the participating force difference feature includes:

[0088] Input the target knowledge graph and any sample knowledge graph, and perform first, second and third processing;

[0089] The first processing includes: edge-to-node processing through the edges in the knowledge graph, edge encoding to add relationship information on the edge, inputting into the SimGNN model, and obtaining the topology structure similarity between the two knowledge graphs based on the SimGNN model calculation;

[0090] The second processing includes: extracting geographic information in the knowledge graph, setting any geographic location as a reference point, calculating the distance between each geographic location in the knowledge graph and the reference point respectively to obtain the distance vector of the reference point; and inputting the distance vector into the Transformer network to extract the geographic spatial information of the two knowledge graphs to obtain the geographic spatial similarity between the two knowledge graphs.

[0091] The third processing includes: extracting force information in the knowledge graph, classifying and counting the number of participating forces of each type of force to obtain a force quantity vector; and inputting the force quantity vector into the Transformer network to obtain the force similarity between the two knowledge graphs.

[0092] The topology similarity, the geographic spatial similarity and the force similarity are spliced and input into the Dense network to obtain the similarity value of the target knowledge graph and the sample knowledge graph.

[0093] In the test, the security scheme recommended by the method of the application has higher accuracy, reaching 92.37%, while the accuracy of the security scheme recommended in the prior art is only 83.54%.

[0094] The knowledge query model is used to receive a natural language question of a user, convert the natural language question into a query language of a database, execute the query language on the database, generate an answer to the natural language question, and complete the query of the security scheme and the corresponding sample knowledge graph. Thus, intelligent question answering on the knowledge graph of the data link network security requirement is realized, and flexible database query and management capabilities are provided for data link network design and planning personnel.

[0095] In the prior art, the model based on machine reading comprehension can only find the answer in the text in an extraction manner, and if the answer does not exist in the text, the correct answer cannot be given.

[0096] For example, given a text and a question of a user; the text is: “On January 1, 2023, the first ship, the second ship and the third ship form the 44th escort formation from a certain city and a certain military port, and sail to the first location and the second location to replace the 43rd escort formation to perform the escort task. Among them, the first ship is an X-type a ship with hull number 1 and a service time of January 1, 2021; the second ship is a Y-type b ship with hull number 2 and a service time of January 1, 2020; and the third ship is a Z-type c ship with hull number 3 and a service time of January 1, 2010.” The question of the user is: “What is the longest service time of the ship in the 44th escort formation?” Since there is no such description in the text, the model based on machine reading comprehension cannot accurately answer such a question.

[0097] Therefore, the application is based on a knowledge query model of template matching and key information extraction, that is, converting a natural language question of a user into a query language of a database, and then executing the query language on the database to generate an answer to the question of the user.

[0098] For example, after completing the data chain network support demand knowledge extraction from the text, a knowledge graph containing the following triple information is obtained: <43rd convoy formation, formation relationship, first ship>, <43rd convoy formation, formation relationship, second ship>, <nth convoy formation, formation relationship, third ship>, <first ship, hull number, 1>, <first ship, model, X type a ship>, <first ship, service time, January 1, 2021>, <second ship, hull number, 2>, <second ship, model, Y type b ship>, <second ship, service time, January 1, 2020>, <third ship, hull number, 3>, <third ship, model, Z type c ship>, and <third ship, service time, January 1, 2010>. By executing the query statement "MATCH (n)-[:formation relationship]->(m)-[:service time]->(q) WHERE n.name='convoy formation' RETURN q.name ORDER BY q.name LIMIT 1" in the knowledge graph, the answer to the question of the user can be obtained, and the problem in the prior art is solved.

[0099] Since the number of data samples in the field of intelligent processing of data link network support demand is small, it is difficult to train a neural network model. Therefore, according to the actual use of the user, the question of the user is pre-set as a question template.

[0100] Specifically, converting the natural language question into the query language of the database includes: establishing a template library, the template library including standard question templates of a plurality of natural language questions; processing semantic information of the natural language question to obtain question semantic coding information, and determining which standard question template the natural language question belongs to based on the question semantic coding information; extracting the natural language question to obtain a plurality of key entities; and filling the key entities into the standard question template to generate an instantiated query language of the database.

[0101] As a specific embodiment of the application, as shown in Figure 7 After inputting the natural language question of the user:

[0102] (1) Extract the semantic information of the user to obtain the semantic encoding information of the question. As an implementation of the present invention, the semantic extraction layer 1 is used to extract the semantic information of the user's question for the first time, for example, using the BERT pre-trained language model; then the semantic extraction layer 2 is used to extract the semantic information of the user's question for the second time to obtain the semantic encoding information of the question; wherein, the semantic extraction layer 2 includes a BiLSTM (Bidirectional Long Short-Term Neural Network) with 300 hidden neurons.

[0103] (2) In the semantic encoding information of the user's question, perform question template classification operation on the output [CLS] identifier position to obtain the standard question template corresponding to the user's natural language question. For example, for the user question "What is the action route of the escort formation?", the matching Cypher graph database query statement question template is "MATCH (n)-[:formation relationship]->(m)-[:service time]->(q) WHERE n.name='key entity' RETURN q.name ORDER BYq.name LIMIT 1"; it should be noted that all template libraries need to be predefined, and then the corresponding standard question template is searched from the list of template libraries according to the classification results.

[0104] (3) Using the key entity start position classification layer, the semantic encoding information of the user's question is serialized and classified to obtain the key entity start position; using the key entity end position classification layer, the semantic encoding information of the user's question is serialized and classified to obtain the key entity end position; combining the key entity start position and the key entity end position, the key entity is obtained from the user's question, such as Figure 7 The "escort fleet" in the middle.

[0105] (4) Fill the key entities into the standard question template to generate an instantiated database query language. That is: assemble the key entities and the question template to obtain the database query statement corresponding to the user's question; such as... Figure 7 The code snippet shows the following: "MATCH(n)-[:formation relationship]->(m)-[:service time]->(q) WHERE n.name='escort formation' RETURN q.nameORDER BY q.name LIMIT 1".

[0106] (5) Finally, execute the query statement in the database to get the answer to the user's question.

[0107] It should be noted that if the extracted key entities do not exist in the knowledge graph, the generated query may not produce any results after execution on the graph database.

[0108] Therefore, the application further comprises the following steps of judging whether the key entity exists in the sample knowledge graph, directly filling the key entity into the standard question template if the key entity exists in the sample knowledge graph, calculating the edit distance between the key entity and each sample entity in the sample knowledge graph, replacing the key entity with the sample entity with the minimum edit distance in the sample knowledge graph, and filling the sample entity into the standard question template if the key entity does not exist in the sample knowledge graph. In this way, the generated answer can be explained and traced using the database query statement, thereby having good explainability and improving the credibility of the answer.

[0109] As a preferred embodiment of the application, the system uses the open source graph database Neo4j and the graph database query language Cypher in the selection of the query language of the database. Specifically, the specific process of obtaining the query result of the user by executing the Cypher graph database query statement on the Neo4j graph database is as follows:

[0110] determining whether each triple data in the triple data list of the data chain network design planning guarantee requirement is traversed:

[0111] If yes, the py2neo library is called to add the subject and object in the current knowledge graph triple data as nodes to the Neo4j graph database, and then the py2neo library is called to add the current knowledge graph triple data as an edge to the Neo4j graph database, until each triple data in the triple data list is traversed;

[0112] If no, the py2neo library is called to input the Cypher graph database query statement corresponding to the natural language question of the user into the Neo4j graph database, and execute the Cypher graph database query statement in the Neo4j graph database to obtain the query result of the user.

[0113] In addition, in the knowledge query module, the calling of the knowledge query model is realized in a service calling manner, so as to avoid the conflict of the underlying modules caused by the simultaneous loading of the AI algorithm models.

[0114] As a preferred embodiment of the application, the RESTFUL-based service calling is realized by calling the flask library, and the specific process is as follows: before the knowledge query module runs, the flask library is called to start the data chain network guarantee requirement knowledge query service and listen to the service request from the user in real time; after receiving the knowledge query request from the user, the natural language question of the current query is obtained from the user request data, and the knowledge query model is called to convert the natural language question of the user into the Cypher graph database query statement; finally, the flask library is called to return the obtained Cypher graph database query statement to the user.

[0115] In another aspect, the present application also provides a data link network guarantee requirement intelligent processing system, which comprises at least a database, a knowledge extraction module, a knowledge query module and a knowledge recommendation module, and is used for executing the steps of any one of the above methods. Figure 8 For a specific embodiment of the present application, the technical features consistent with the above method are not described again.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of the various embodiments or some parts of the embodiments.

[0117] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for handling data link network assurance requirements, characterized in that, The method comprises: setting a database and integrating a plurality of AI algorithm models based on service calls to realize knowledge extraction, knowledge recommendation and knowledge query of data chain network guarantee requirements; The database comprises a plurality of guarantee schemes and sample knowledge graphs corresponding to the guarantee schemes; the AI algorithm models at least comprise a knowledge extraction model, a knowledge recommendation model and a knowledge query model; each AI algorithm model is encapsulated in the form of a service; A process pool is created, one process in the process pool corresponds to loading one AI algorithm model; an idle process in the process pool is selected by a service request to call the AI algorithm model, and a calling result of the AI algorithm model is fed back by a service response; The knowledge query model is used to receive a natural language question sentence of a user, convert the natural language question sentence into a query language of the database, and specifically comprises: A template library is established, the template library comprises a plurality of standard question templates of natural language questions; Semantic information of the natural language question is processed to obtain question semantic coding information, and it is judged that the natural language question belongs to which standard question template based on the semantic coding information; A plurality of key entities are extracted from the natural language question; The key entities are filled into the standard question template to generate an instantiated query language of the database.

2. The method of claim 1, wherein, The knowledge extraction model is used to receive a target document, extract demand elements from the target document to form a target knowledge graph; The construction of the knowledge extraction model comprises: extracting entities from the target document by using an existing extraction model to obtain triple data and generate a triple knowledge base; using a graph representation learning method to generate entity vector coding corresponding to each entity in the triple knowledge base; using the entity vector coding as a supplementary and enhancement means of entity semantic information to improve the existing extraction model to obtain the knowledge extraction model.

3. The method of claim 1 or 2, wherein The method for knowledge extraction comprises: Receiving a target document, reading each text paragraph, generating a text paragraph list, and dividing the text paragraph list into a plurality of equal parts to obtain a plurality of text paragraph sublists; Iterating through a plurality of processes in the process pool in turn, passing a text paragraph sublist to a process, and extracting sublist demand elements from the current text paragraph sublist by calling the knowledge extraction model by the current process; After all processes in the process pool complete the extraction, the triple data lists of all sublist demand elements are spliced to obtain the demand elements of the target document.

4. The method of claim 2, wherein the network security requirement is determined based on a type of the data link network. The knowledge recommendation model is used to calculate a similarity value of the target knowledge graph and each sample knowledge graph in the database, and output a guarantee scheme corresponding to a sample knowledge graph with the highest similarity value. ​ 5. The method of claim 4, wherein the network security requirement is determined by the network security requirement determination module based on the network security requirement information and the network security requirement information is received from the network security requirement information database. 5 Calculating the similarity value of the target knowledge graph and each sample knowledge graph in the database comprises: Based on the basic framework of the SimGNN model, the edges in the knowledge graph are set as special nodes in the knowledge graph; and based on the difference characteristics of the graph topology structure of the SimGNN model, combat area difference characteristics and combat force difference characteristics are added.

6. The method of claim 5, wherein the network security requirement is determined by the network security requirement determination module based on the network security requirement information and the network security requirement information is received from the network security requirement information database. 5 The increasing of the combat area difference feature and the combat force difference feature comprises: inputting the target knowledge graph and any sample knowledge graph, performing first, second and third processing; the first processing comprises: edge-to-node processing, edge encoding to increase the relationship information on the edge, and inputting into the SimGNN model to obtain the topological structure similarity between the two knowledge graphs based on the SimGNN model; the second processing comprises: extracting geographical information from the knowledge graph, setting any geographical position as a reference point, calculating the distance between each geographical position in the knowledge graph and the reference point to obtain the distance vector of the reference point; and inputting the distance vector into the Transformer network to extract the geographical space information of the two knowledge graphs and obtain the geographical space similarity between the two knowledge graphs; the third processing comprises: extracting the force information in the knowledge graph, classifying and counting the number of combat forces of each type of force to obtain the force quantity vector; and inputting the force quantity vector into the Transformer network to obtain the force similarity between the two knowledge graphs; splicing the topological structure similarity, the geographical space similarity and the force similarity, and inputting into the Dense network to obtain the similarity value of the target knowledge graph and the sample knowledge graph.

7. The method of claim 1, wherein the network security requirements are determined by a network security policy. The knowledge query model is used to receive a user's natural language question, convert the natural language question into a query language of the database, execute the query language on the database, generate an answer to the natural language question, and complete the query of the support scheme and the corresponding sample knowledge graph. ​ 8. The data link network assurance requirement processing method as described in claim 1, characterized in that, The filling of the key entity into the standard question template comprises: determining whether the key entity exists in the sample knowledge graph; if it exists, directly filling the key entity into the standard question template; if it does not exist, calculating the edit distance between the key entity and each sample entity in the sample knowledge graph, replacing the key entity with the sample entity with the smallest edit distance in the sample knowledge graph, and filling the sample entity into the standard question template.

9. A data link network demand assurance intelligent processing system, characterized in that, The system at least comprises a database, a knowledge extraction module, a knowledge query module and a knowledge recommendation module, which are used to perform the steps of the method in any one of claims 1-8.

Images