Knowledge graph question-answering system based on digital clues

Abstract

The invention relates to a knowledge graph question answering system based on digital clues, and belongs to the field of artificial intelligence and natural language processing. According to the invention, a digital clue platform is utilized to collect and sort multi-source heterogeneous data of a product in a full life cycle, and a related knowledge graph is constructed. And taking the constructed knowledge graph as a database, extracting key information from questions of the user, performing related search in the knowledge graph, and returning a result to the user. The method supports four types of problems, wherein the first type is to query specific information of a certain entity node; the second class is to query entity node information having a corresponding relationship with a given entity node according to a certain determined relationship, and the third class is to query all node information having a relationship with the given entity node; and the fourth class is used for querying a parameter influence domain related to the entity node of a certain'parameter 'type. According to the knowledge question-answering system, the user can quickly and accurately obtain wanted knowledge, and user requirements are greatly facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and natural language processing, and relates to a knowledge graph question-answering system based on digital clues. Background Art

[0002] Digital Thread refers to the creation and use of cross-domain, public, and complex product physical features that support two-way communication of model information in physical and digital spaces (simulation models, test models, process models, and detection models, etc.). On the one hand, it ensures the continuous unification of data, models, and information throughout the entire life cycle from solution analysis, engineering manufacturing and development, production deployment to operational support, thereby achieving dynamic real-time evaluation of current and future product functions and performance; on the other hand, it feeds back information from the physical space to virtual product development, establishes a unified technical framework, supports cross-regional protocol interfaces, integrates with engineering knowledge management systems, provides an integrated perspective on complex organizations, and strengthens the quantitative analysis and confirmation of product performance boundaries and uncertainties.

[0003] Focusing on the entire life cycle of products, the digital clue platform widely collects a large amount of data generated by products from demand, planning, design, production, operation, maintenance, to recycling and scrapping. These data are usually multi-source heterogeneous. According to whether the data is structured, the data can be roughly divided into structured data (such as tables and records stored in relational databases), unstructured data (such as pictures, videos and text data) and semi-structured data (such as data in XML and JSON format). A variety of rich knowledge information exists in massive multi-source heterogeneous data, but most of this knowledge exists in an unstructured form and is difficult for users to use directly. Summary of the invention

[0004] The technical problem solved by the present invention is: to overcome the deficiencies of the prior art and to propose a knowledge graph question-answering system based on digital clues.

[0005] The solution of the present invention is:

[0006] A knowledge graph question-answering system based on digital clues, comprising a knowledge graph construction module, a query statement construction module, a classification analysis module, and a result feedback module;

[0007] Knowledge graph construction module: constructs a knowledge graph based on the multi-source heterogeneous data of the entire product life cycle collected and organized by the digital clue platform;

[0008] Classification and parsing module: classifies the natural language questions input by users to determine their categories and parses them;

[0009] Query statement construction module: uses the knowledge graph constructed by the knowledge graph construction module as a graph database, and constructs query statements based on the classification analysis results of the classification analysis module;

[0010] Result feedback module: obtains query results from the graph database using the query statement and outputs them.

[0011] Preferably, the knowledge graph stores the requirements, architectures, models, data and associations generated during the entire product life cycle in a graph structure;

[0012] In the knowledge graph, different demand items are stored in the form of nodes, and the derived relationships between demands are represented by the edges connecting the nodes.

[0013] Preferably, the knowledge graph contains 8 different entity types, namely: architecture design, design details, manufacturing, requirements, services, simulation, testing, and parameters;

[0014] The system architecture includes the process of the overall structure and organization of the system, design details include the specific details that are concerned and handled during the design process, manufacturing includes the various stages of converting product design into actual products, requirements define the characteristics and functions that the product should have, services include product-related support and after-sales services during the product life cycle, simulation refers to virtual testing and evaluation in the early stages of product design and development, testing refers to the process of verifying and confirming product performance, function and quality during product development and life cycle, and parameters are used to describe and define adjustable or measurable variables of product performance, characteristics and behavior.

[0015] Preferably, the knowledge graph includes five different association relationships: contain, derive, feedback, trace, and verify.

[0016] Preferably, the natural language questions input by the user are divided into four categories, namely:

[0017] The first category: query specific information of a certain entity node;

[0018] The second category: querying the entity node information that has a corresponding relationship with a given entity node based on a certain relationship;

[0019] The third category: query all node information that has a relationship with a given entity node;

[0020] The fourth category: querying the parameter impact domain related to a certain entity node of the "parameter" type.

[0021] Preferably, the method for classifying the natural language questions input by the user is as follows:

[0022] Store the name attribute values ​​of all entity nodes of the "parameter" type in a txt file. If the name attribute value of a "parameter" type entity appears in the natural language question input by the user, it means that the question input by the user belongs to the first category.

[0023] Enumerate all possible relationships. If the natural language question input by the user contains any of these relationships, it means that the question input by the user belongs to the second category.

[0024] Determine whether the natural language question input by the user contains the keywords "all" or "all". If so, it means that the question input by the user belongs to the third category;

[0025] If the natural language question input by the user mentions product parameters belonging to different life cycles, and there are different association relationships between the nodes corresponding to these parameters, it means that the question input by the user belongs to the fourth category.

[0026] Preferably, when parsing the natural language input provided by the user, node information and relationship information that may be contained in the natural language question input by the user are extracted.

[0027] Preferably, in the result feedback module, for the fourth type of question, a query statement is used to perform a multi-hop query in the graph database to obtain the query result.

[0028] Preferably, a depth-first search algorithm or a breadth-first search algorithm is used to implement multi-hop query.

[0029] Preferably, the knowledge graph is stored in a neo4j graph database.

[0030] The beneficial effects of the present invention compared with the prior art are:

[0031] The present invention uses multi-source heterogeneous data collected by the digital clue platform to construct a knowledge graph, and creatively divides the natural language questions input by users into four categories according to the scenarios, constructs query statements for each category, and obtains corresponding query results from the knowledge graph. With the present invention, users do not need to analyze the complex relationships hidden in the knowledge graph, but can directly obtain the required relevant knowledge by asking questions in natural language, and feed it back to other life cycles of the product. This allows users to quickly and accurately obtain the knowledge they want, greatly facilitating user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A visual representation of the knowledge graph of the product constructed for the present invention;

[0033] Figure 2 This is a flow chart of the knowledge question answering system of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the accompanying drawings.

[0035] With the rapid development of technologies in related fields such as computers, knowledge graphs provide a way to convert knowledge into manageable data in a unified and structured manner. Although knowledge graphs can be used to convert knowledge into manageable data in a unified and structured manner, it is extremely difficult for users to obtain useful knowledge from massive data. In order to facilitate users and enable them to quickly and accurately obtain the desired knowledge, the present invention constructs a knowledge graph using multi-source heterogeneous data collected by the digital clue platform, and further performs knowledge questions and answers based on the constructed knowledge graph. In this way, users do not need to analyze the complex relationships hidden in the knowledge graph, and can directly obtain the required relevant knowledge by asking questions in natural language, and feed it back to other life cycles of the product.

[0036] The knowledge graph question-answering system based on digital clues of the present invention includes a knowledge graph construction module, a classification and analysis module, a query statement construction module, and a result feedback module.

[0037] The knowledge graph construction module constructs a knowledge graph based on the multi-source heterogeneous data of the product life cycle collected and sorted by the digital clue platform. The classification and analysis module classifies the natural language questions input by the user to determine the category to which they belong, and then analyzes them. The query statement construction module uses the knowledge graph constructed by the knowledge graph construction module as a graph database, and constructs a query statement based on the classification and analysis results of the classification and analysis module. The result feedback module uses the query statement to obtain the query result from the graph database and outputs it.

[0038] The knowledge graph stores the requirements, architectures, models, data, and associations generated throughout the entire product process in a graph structure. Taking the allocation of stakeholder requirements to black box requirements as an example, different requirement items are stored in the form of nodes, and the derived relationships between requirements are represented by the edges connecting the nodes. This solves the problem that only two-dimensional cross-domain relationships can be displayed using tables and matrices throughout the entire process of traditional products. The graph-structured association feedback network realizes the construction of data association feedback links that are traditionally implemented across tables and matrices, and has stronger adaptability when dealing with scenarios such as products that involve a wide range of fields and complex associations.

[0039] At present, the constructed knowledge graph includes 8 different entity types, as shown in Table 1:

[0040] Table 1 Description of knowledge graph entity nodes

[0041]

[0042] In addition, the knowledge graph currently constructed includes five different association relationships: contain, derive, feedback, trace, and verify. Figure 1 As shown. The knowledge graph stores multi-source heterogeneous data of the entire life cycle of the product collected using the digital clue platform. During the entire life cycle of the product, data collection is mainly carried out around the three main life cycles of demand, design and simulation. Therefore, there are a large number of "parameter" type entity nodes in the knowledge graph that have inclusion relationships with entity nodes of types such as architectureDesign, detailDesign and requirement. Intuitively, it means "the product needs to meet n different parameter requirements in terms of structural design, design details and demand relationships".

[0043] Although deep learning technology can effectively improve the efficiency of knowledge question answering and may mine more hidden knowledge that is not stored in the original knowledge graph, this type of method is not applicable in specific fields such as aerospace. On the one hand, many fields, including aerospace, have high requirements for the reliability of knowledge and need to ensure absolute correctness. Some unverified hidden information should not be used. On the other hand, the training of deep learning models requires a large number of samples, which means that a lot of resources are needed to collect, clean and label data. Considering that the amount of data in the knowledge graph currently constructed is not sufficient, the use of a deep learning model with a complex structure is prone to overfitting problems, which ultimately affects the accuracy of knowledge question answering and knowledge mining. Based on the above considerations, the present invention takes the accuracy of knowledge question answering as the primary purpose, and uses the query of the graph database to ensure the reliability of the answers given.

[0044] Consider the following scenarios:

[0045] Scenario 1: When using or designing a product, a user does not know the specific requirements for a certain parameter. In this case, the user can use the knowledge question-answering system to describe his or her question in natural language and obtain the specific requirements for a certain parameter.

[0046] Scenario 2: Assume that the user knows that node a has a certain relationship r with a certain node, but does not know which node the corresponding relationship r is with. If the user wants to know the information of the entity node that has a corresponding certain relationship with a, the user can use the knowledge question answering system to ask questions in natural language and obtain the information of the entity node that has a specific relationship with a.

[0047] Scenario 3: Assuming that a user wants to fully understand a certain entity node, the user can use the knowledge question-answering system to ask questions in natural language and obtain information about all other entity nodes that are related to the entity node.

[0048] In addition, considering more complex multi-hop queries, the knowledge question answering system can answer the parameter impact domain related to a certain "parameter" type entity node.

[0049] In summary, the problems can be classified into 4 different categories:

[0050] 1. Answer specific information of a certain entity node;

[0051] 2. According to a certain relationship, answer the entity node information that has a corresponding relationship with the given entity node;

[0052] 3. Answer all node information related to the given entity node;

[0053] 4. Answer the parameter impact domain related to a certain entity node of type "parameter".

[0054] Considering the four types of questions mentioned above, for the natural language questions entered by the user, we first need to classify them according to the questions to clarify which category they belong to. Then parse the question and extract the node information and relationship information that may be contained in the question. Then convert the question into the corresponding graph database query statement. Since the constructed knowledge graph is stored in the neo4j graph database, it is converted into the corresponding cypher statement, and the query result is returned as the final return value. The results obtained based on the graph database query can guarantee accuracy. For knowledge that does not exist in the database, other data and knowledge will not be returned at will. The flow chart of the knowledge question and answer system is as follows: Figure 2 shown.

[0055] Regarding the first type of question, that is, answering specific information about a certain entity node. Some entity nodes contain a description attribute, which is a specific description of the entity. When answering this type of question, the knowledge question answering system directly uses the value of the description attribute as the answer based on the user's question. In order to determine whether the user's question belongs to the first type of question, the following approach is adopted: the name attribute values ​​of all "parameter" type entity nodes are stored in the keywords4SearchParameters.txt file. If the name attribute value of a "parameter" type entity appears in the user's question, it means that the question raised by the user belongs to the first category. It is worth noting that since entity nodes of different "parameter" types contain different attributes, for example, one node contains name and description attributes, while another node contains Name and Text attributes. When converting into a cypher statement, all situations need to be considered, so Name and name will be used as query constraints for querying.

[0056] For the second type of questions: Based on a certain relationship, answer the entity node information that has a corresponding relationship with the given entity node. In order to determine whether the question belongs to the second type of question, the method adopted is to enumerate all possible relationships, such as inclusion, derivation, feedback, etc. If the question contains a certain specific relationship, the given node information and the given relationship information are extracted for query. The parsed specific entity nodes and specific associations are directly converted into corresponding cypher statements, and the knowledge graph queries and returns the results based on the cypher query statements. In addition, in order to facilitate users to ask questions, fuzzy query statements are used when converting to cypher statements.

[0057] For the third type of question: answer all node information that has a relationship with a given entity node. In order to determine whether a question belongs to the third type of question, the method used is to determine whether the user's question contains keywords such as "all" and "all". And it can be regarded as an extension of the second type of question. It is necessary to query according to all possible associations and add the results of each time to the set. Similarly, for the convenience of users, the question is mapped to a cypher statement of fuzzy query.

[0058] For the fourth type of problem, it involves complex multi-hop queries. For example, if the user's question mentions the parameters of products belonging to different life cycles, and there are complex relationships between these nodes, then when querying, the nodes involved in the name value will derive other nodes, or the user wants to obtain all nodes related to the name value node, or wants to obtain the relevant parameter impact domain, then a multi-hop query is required. The method used here is the classic depth-first search (DepthFirst Search) or breadth-first search algorithm (Breadth First Search). Taking the breadth-first algorithm as an example, a queue is maintained to store nodes that are in the parameter impact domain but have not been queried. Each time for an unqueried node, it is converted into the third type of problem, that is, querying all nodes related to the entity node, and entering the unqueried nodes into the queue.

[0059] According to the natural language question input provided by the user, the question needs to be classified to determine the category it belongs to. Then the question is parsed to extract the node information and relationship information that may be involved. Subsequently, the question is converted into a corresponding graph database query statement. The knowledge graph of the present invention is built based on the Neo4j graph database, so the query statement needs to be converted into a corresponding Cypher query statement, and the query result is used as the final return value. In the specific implementation process, the questions are divided into four categories, and examples of these four categories of questions are as follows:

[0060] The first type of question is to query specific information of a certain entity node. The knowledge question answering system extracts key information from the user's question and can give an accurate answer response by processing the key information. The cypher query statement is as follows: cypher1 = "MATCH(n:`parameter`) where n.Name = ~\".*{0}.*\"RETURN n.Text".format(line).

[0061] The second type of question is to query the entity node information that has a corresponding relationship with a given entity node based on a certain relationship. The system extracts the certain relationship in the question raised by the user, queries based on the relationship, and returns the result to the user. The cypher query statement is as follows: cypher = "MATCH (m) - [r: satisfy] -> (n) where m.name = ~\".*{0}\"RETURN n.Name".format(node_name).

[0062] The third type of problem is to query all node information that has a relationship with a given entity node. This type of problem can be seen as an extension of the second type of problem. In this type of problem, the system needs to query all possible associations of key nodes. In this way, the user will get all the node information that has a relationship with the target node. The cypher query statement is as follows: cypher1 = "MATCH (m) - [r] -> (n) where m.name = ~

[0063] \".*{0}\"RETURN n.name".format(node_name).

[0064] The fourth type of question is to query the parameter impact domain related to a certain "parameter" type entity node. This question involves complex multi-hop queries. The system will query related entity nodes and nodes related to them based on keywords. These relationships include derivation, tracing and verification, and will display the parameter impact domains of these related nodes. The Cypher query statement is as follows: cypher = "MATCH (m) - [r] -> (n) where m.name = ~\". * {0} \ "RETURN n.name, TYPE (r)". format (current_node_name).

[0065] The present invention uses the constructed knowledge graph as a database, extracts key information from the user's questions, performs relevant searches in the knowledge graph, and returns the results to the user. The present invention supports four types of questions. The first type is to query the specific information of a certain entity node; the second type is to query the entity node information that has a corresponding relationship with a given entity node based on a certain determined relationship; the third type is to query all node information that has a relationship with a given entity node; the fourth type is to query the parameter influence domain related to a certain "parameter" type entity node. Based on digital clues, the present invention integrates knowledge engineering and natural language processing, and uses the constructed knowledge graph to form a knowledge base to form a clear and reliable knowledge question and answer system.

[0066] The contents not described in detail in the specification of the present invention belong to the common knowledge of the professionals in this field.

Claims

1. A knowledge graph question answering system based on digital clues, characterized by: It includes knowledge graph construction module, query statement construction module, classification analysis module, and result feedback module; Knowledge graph construction module: constructs a knowledge graph based on the multi-source heterogeneous data of the entire product life cycle collected and organized by the digital clue platform; Classification and parsing module: classifies the natural language questions input by users to determine their categories and parses them; Query statement construction module: uses the knowledge graph constructed by the knowledge graph construction module as a graph database, and constructs query statements based on the classification analysis results of the classification analysis module; Result feedback module: obtains query results from the graph database using the query statement and outputs them.

2. According to claim 1, a knowledge graph question-answering system based on digital clues is characterized by: The knowledge graph stores the requirements, architectures, models, data and associations generated throughout the product life cycle in a graph structure; In the knowledge graph, different demand items are stored in the form of nodes, and the derived relationships between demands are represented by the edges connecting the nodes.

3. A knowledge graph question-answering system based on digital clues according to claim 2, characterized in that: The knowledge graph contains 8 different entity types, namely: architecture design, design details, manufacturing, requirements, services, simulation, testing, and parameters; The system architecture includes the process of the overall structure and organization of the system, design details include the specific details that are concerned and handled during the design process, manufacturing includes the various stages of converting product design into actual products, requirements define the characteristics and functions that the product should have, services include product-related support and after-sales services during the product life cycle, simulation refers to virtual testing and evaluation in the early stages of product design and development, testing refers to the process of verifying and confirming product performance, function and quality during product development and life cycle, and parameters are used to describe and define adjustable or measurable variables of product performance, characteristics and behavior.

4. A knowledge graph question-answering system based on digital clues according to claim 2, characterized in that: The knowledge graph includes five different relationships: contain, derive, feedback, trace, and verify.

5. The knowledge graph question-answering system based on digital clues according to claim 1, characterized in that: The natural language questions entered by users are divided into four categories: The first category: query specific information of a certain entity node; The second category: querying the entity node information that has a corresponding relationship with a given entity node based on a certain relationship; The third category: query all node information that has a relationship with a given entity node; The fourth category: query the parameter impact domain related to a certain "parameter" type entity node.

6. A knowledge graph question-answering system based on digital clues according to claim 5, characterized in that: The method for classifying natural language questions input by users is as follows: Store the name attribute values ​​of all entity nodes of the "parameter" type in a txt file. If the name attribute value of a "parameter" type entity appears in the natural language question input by the user, it means that the question input by the user belongs to the first category. Enumerate all possible relationships. If the natural language question input by the user contains any of these relationships, it means that the question input by the user belongs to the second category. Determine whether the natural language question input by the user contains the keywords "all" and "all". If so, it means that the question input by the user belongs to the third category; If the natural language question input by the user mentions product parameters belonging to different life cycles, and there are different association relationships between the nodes corresponding to these parameters, it means that the question input by the user belongs to the fourth category.

7. The knowledge graph question-answering system based on digital clues according to claim 1, characterized in that: When parsing the natural language input provided by the user, the node information and relationship information that may be contained in the natural language question input by the user are extracted.

8. The knowledge graph question-answering system based on digital clues according to claim 1, characterized in that: In the result feedback module, for the fourth type of questions, query statements are used to perform multi-hop queries in the graph database to obtain query results.

9. A knowledge graph question-answering system based on digital clues according to claim 8, characterized in that: Use depth-first search algorithm or breadth-first search algorithm to implement multi-hop query.

10. The knowledge graph question-answering system based on digital clues according to claim 1, characterized in that: The knowledge graph is stored in the neo4j graph database.

Images