Graph knowledge base question and answer optimization method and device, equipment and storage medium

Abstract

The invention provides a graph knowledge base question and answer optimization method and device, equipment and a storage medium, and relates to the technical field of computer systems.The optimization method comprises the steps that when query information is detected, intention recognition is conducted on the query information, a recognition result is obtained, and a corresponding preset graph knowledge base is determined according to the recognition result; when the preset graph knowledge base is a document class graph knowledge base, determining target entity data according to the query information, and inputting the target entity data into the document class graph knowledge base for query to obtain related entity data and a corresponding related entity relationship; and based on a preset large language model, obtaining a final question and answer result according to the related entity data and the corresponding related entity relationship. Through intention recognition and selection of the preset graph knowledge base, the core intention of user query can be accurately positioned, so that the corresponding graph knowledge base is quickly determined, invalid search in irrelevant knowledge bases is avoided, and the query efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer system technology, and more specifically, to a method, apparatus, device, and storage medium for optimizing question answering in a graph knowledge base. Background Technology

[0002] With the development of artificial intelligence technology, graph knowledge base question-answering systems are widely used in enterprise knowledge management, intelligent customer service, and other fields. However, these technologies have the following problems: when processing user queries, they require extensive searching in large-scale general knowledge bases, making it difficult to quickly locate the most relevant part of the query, resulting in low query efficiency; furthermore, they are difficult to adapt to diverse query needs, cannot accurately understand user intent, and tend to search in irrelevant knowledge bases, leading to inaccurate results and requiring users to repeatedly query or obtain information through other means. Summary of the Invention

[0003] The problem solved by this invention is one or more of the aforementioned related technical problems.

[0004] To address the aforementioned problems, this invention provides a method, apparatus, device, and storage medium for optimizing question answering in a graph knowledge base.

[0005] In a first aspect, the present invention provides a graph knowledge base question answering optimization method, comprising: When query information is detected, intent recognition is performed on the query information to obtain recognition results, and the corresponding preset graph knowledge base is determined based on the recognition results; When the preset graph knowledge base is a document class graph knowledge base, the target entity data is determined according to the query information, and the target entity data is input into the document class graph knowledge base for querying to obtain relevant entity data and corresponding relevant entity relationships; Based on a pre-defined large language model, the final question-and-answer result is obtained according to the relevant entity data and the corresponding relevant entity relationships.

[0006] Optionally, the construction process of the document class graph knowledge base includes: Obtain document data and segment the document data to obtain multiple file blocks; Input all the file blocks into the preset extraction model to obtain the corresponding entity data and the corresponding entity relationships; The document class graph knowledge base is constructed based on all the entity data and the corresponding entity relationships.

[0007] Optionally, the step of obtaining the final question-and-answer result based on the preset large language model, according to the relevant entity data and the corresponding relevant entity relationships, includes: Based on the document class graph knowledge base, the relevant entity data and the corresponding relevant entity relationships are obtained according to the target entity data and the preset number of query steps; The relevant entity data and the target entity data are matched with the preset database to obtain the corresponding relevant file blocks; The target entity data, the corresponding related file blocks, the related entity data, and the corresponding related entity relationships are input into the preset large language model to obtain the final question-answering result. The preset database includes multiple file blocks.

[0008] Optionally, the graph knowledge base question answering optimization method further includes: When the preset graph knowledge base is a general graph database, historical query data for the corresponding preset time period is obtained, and the historical query data is analyzed to obtain the analysis results; Based on the analysis results and the query information, the corresponding potential intent information is determined; Based on the general class graph database, a final query statement is generated according to the potential intent information and the query information; The final query statement is input into the large language model to obtain the corresponding final question and answer result; The construction process of the general class graph database includes: The system acquires user-uploaded data, determines node data and corresponding node relationship data based on the uploaded data, and constructs the general class graph database based on the node data and the corresponding node relationship data.

[0009] Optionally, generating the final query statement based on the general class graph database, according to the latent intent information and the query information, includes: Based on the general class graph database, determine the relevant node data and the corresponding relevant node relationship data corresponding to the query information; An initial query statement is generated based on the potential intent information, the relevant node data, and the corresponding relevant node relationship data. The initial query statement is verified and modified to obtain the final query statement.

[0010] Optionally, determining the target entity data based on the query information includes: The query information is then transformed to obtain temporary vector data; The temporary vector data is compared with the entity vectors in the preset vector library to obtain the entity vector with the highest similarity, and the entity vector with the highest similarity is used as the corresponding target entity data.

[0011] Optionally, the graph knowledge base question answering optimization method further includes: The final query statement is evaluated using a pre-defined cache library. Determine whether the question and answer results corresponding to the final query statement exist in the preset cache library; If it exists, the returned question and answer result will be used as the final query statement.

[0012] Secondly, the present invention provides a graph knowledge base question-answering optimization device, comprising: The identification unit is used to identify the intent of the query information when the query information is detected, obtain the identification result, and determine the corresponding preset graph knowledge base based on the identification result. The processing unit is configured to, when the preset graph knowledge base is a document class graph knowledge base, determine the target entity data according to the query information, input the target entity data into the document class graph knowledge base for querying, obtain relevant entity data and corresponding relevant entity relationships; and, based on a preset large language model, obtain the final question-and-answer result according to the relevant entity data and the corresponding relevant entity relationships.

[0013] Thirdly, the present invention provides a graph knowledge base question answering optimization device, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the graph knowledge base question answering optimization method as described in the first aspect when the computer program is executed.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the graph knowledge base question-answering optimization method as described in the first aspect.

[0015] The beneficial effects of the graph knowledge base question-answering optimization method, apparatus, system, and storage medium of the present invention are: When a user's query is received, the first step is to perform an intent recognition process to understand the user's query purpose or need, i.e., the recognition result. Intent recognition is usually achieved through a machine learning model, which may have been trained to recognize different query types and intents.

[0016] Based on the obtained intent recognition results, a corresponding preset graph knowledge base is selected for querying. The preset graph knowledge base refers to a pre-built knowledge base used to store specific types of information. The selection of the preset graph knowledge base may be based on the degree of matching between the information types stored in the knowledge base and the user's query intent.

[0017] If the selected preset graph knowledge base is the document class graph knowledge base, the system will determine the target entity data to be queried (such as including nouns, text names, etc.) based on the user's query information, and input this data as query conditions into the document class graph knowledge base for querying.

[0018] After querying the document class diagram knowledge base, you can obtain other entity data related to the target entity data, as well as the relationships between these entities. The query results from the document class diagram knowledge base may include entity attributes, connection relationships between entities, etc.

[0019] Finally, using a pre-defined large language model, the relevant entity data and their relationships obtained from the query are taken as input to generate the final question-and-answer result. The large language model is an artificial intelligence model with language understanding and generation capabilities, capable of generating responses in natural language form based on the input information.

[0020] In summary, this invention, through intent recognition and the selection of a pre-defined graph knowledge base, can accurately pinpoint the core intent of a user's query, thereby quickly determining the corresponding graph knowledge base and avoiding ineffective searches in irrelevant knowledge bases, significantly improving query efficiency. Intent recognition technology ensures the system accurately understands the user's query needs, providing precise guidance for subsequent queries and ensuring the accuracy of query results. Furthermore, a large language model is used to transform the retrieved entity data and relationships into natural and accurate question-and-answer results. The model's powerful language organization capabilities enable it to generate answers that conform to human language habits based on input data, thereby enhancing the user experience. Specifically, the document-type graph knowledge base uses a graph structure to store knowledge, efficiently representing the complex relationships between queried entities. When a document-type graph knowledge base is identified, the system can quickly query relevant entities and their relationships based on the target entity data, comprehensively acquiring information related to the query.

[0021] This method flexibly adapts to different query needs, providing effective query services through intent recognition and knowledge base selection, whether for text-based or project information-based queries. Furthermore, the combination of graph databases and large language models effectively processes and analyzes massive amounts of data, providing users with more accurate decision support. Attached Figure Description

[0022] Figure 1 This is one of the flowcharts illustrating a graph knowledge base question answering optimization method according to an embodiment of the present invention; Figure 2 This is a second flowchart illustrating a graph knowledge base question answering optimization method according to an embodiment of the present invention; Figure 3 This is a third flowchart illustrating a graph knowledge base question-answering optimization method according to an embodiment of the present invention; Figure 4This is a schematic diagram of the structure of a graph knowledge base question answering optimization device according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a graph knowledge base question-answering optimization device according to an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0024] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0025] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0026] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0027] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0028] To address the problems existing in the aforementioned related technologies, embodiments of the present invention provide a graph knowledge base question answering optimization method, apparatus, system, and storage medium.

[0029] like Figure 1 As shown in the figure, an embodiment of the present invention provides a graph knowledge base question answering optimization method, including: Step S100: When query information is detected, the query information is subjected to intent recognition to obtain recognition results, and the corresponding preset graph knowledge base is determined based on the recognition results.

[0030] Specifically, the query information can be a question in natural language, such as "Does employee P belong to department D?" or "Who is the theoretical basis of the GPS system related to?"

[0031] The received query information is preprocessed, including but not limited to word segmentation, stop word removal, and syntactic analysis, in order to better understand the structure and content of the query.

[0032] Pre-trained machine learning or deep learning models (such as Transformer-based models) are used to identify intent from pre-processed query information. The model determines the user's query intent based on features such as keywords, phrases, and context, for example, querying who is related to the theoretical basis of the GPS system.

[0033] Based on the intent recognition results, the system determines which pre-defined graph knowledge base to use. The pre-defined graph knowledge base includes a document-type graph knowledge base and a general-type graph database. The document-type graph knowledge base is built upon document-type data, such as published papers, articles, and internal company documents. This data is processed and analyzed to construct a graph knowledge base containing entities (such as authors, keywords, and journals) and their relationships (such as citation relationships and collaboration relationships).

[0034] General-purpose graph databases: These are built upon structured data, such as organizational structure data and project management data. This data typically exists in tabular or database form, and is transformed to create a graph database containing entities (such as employees, departments, and projects) and their relationships (such as hierarchical relationships and superior-subordinate relationships).

[0035] Based on the intent recognition results, select the most appropriate graph knowledge base type. For example, if the user queries a question related to document content (such as "citation information of a certain paper"), then a document-based graph knowledge base is selected; if the question is related to organizational structure (such as "list of employees in a certain department"), then a general graph database is selected.

[0036] By quickly identifying the user's query intent through intent recognition, the system avoids searching in irrelevant knowledge bases, thereby reducing query time and improving efficiency. For example, if a user queries an organizational structure question, the system will skip the document graph knowledge base and search in a general graph database. Intent recognition accurately understands user needs, enabling the system to select the most relevant graph knowledge base for querying, reducing query errors caused by knowledge base mismatches and improving the accuracy of query results. For instance, for document content queries, the document graph knowledge base can provide a more precise answer. This fast and accurate query process enhances the user experience, allowing users to obtain the information they need more quickly, reducing waiting time and improving work efficiency. The system can flexibly select different graph knowledge bases based on different query intents, adapting to various query needs. Whether it's text-based or structured data queries, the system can effectively handle them, thus expanding the system's application scope. Simultaneously, by selecting the most suitable graph knowledge base for querying, access to irrelevant knowledge bases is avoided, optimizing the utilization of system resources and reducing the waste of computing resources.

[0037] Through the above steps and beneficial effects, step S100 demonstrates how the present invention provides users with specific and accurate query services through intent recognition and selection of a graph knowledge base, while improving query efficiency and user experience.

[0038] Step S200: When the preset graph knowledge base is a document class graph knowledge base, the target entity data is determined according to the query information, and the target entity data is input into the document class graph knowledge base for querying to obtain relevant entity data and corresponding relevant entity relationships.

[0039] Specifically, after determining to use the document class graph knowledge base, the user's original query information is further parsed. The parsing process includes word segmentation of the query statement, removal of stop words, and identification of key entities in the query (such as paper titles, author names, keywords, etc.). For example, for the query "Has author A published any papers on topic X?", "author A" and "topic X" will be identified as key entities.

[0040] Based on the analysis results and the structure of the document class graph knowledge base, the specific target entity data is determined. This may include the entity's name and attributes (such as the publication year of the paper, journal name, etc.). For example, "Author A" is identified as the target entity, and its related attributes and relationships are extracted from the document class graph knowledge base.

[0041] The identified target entity data is used as input to perform a query in the document class graph knowledge base. The query process may use a graph database query language (such as Cypher) to find other entities related to the target entity and their relationships. For example, a query like "MATCH (a:Author {name: 'A'})-[:WROTE]->(p:Paper {topic: 'X'})RETURN a, p" might be executed to find whether author A has published a paper on topic X. Here, MATCH is a keyword in the Cypher query language used to specify the pattern to match. (a:Author {name: 'A'}) defines a node named 'a' of type Author, with the 'name' attribute value being 'A', indicating that the author we are interested in is A. -[:WROTE]-> indicates starting from the author node 'a' and pointing to another node along the relationship named WROTE. In a graph database, a relationship is used to represent the connection between two nodes; here, WROTE represents the action or relationship of an author writing a paper. (p:Paper {topic: 'X'}): This part defines a node named p of type Paper, with the topic attribute value 'X', indicating that the paper we are interested in is on the topic X. RETURN a, p: This is the return part of the query, specifying the nodes to return. Here, it returns the matching author node a and paper node p to check if author A has indeed published a paper on topic X.

[0042] The query results return data on other entities related to the target entity and their relationships. This data may include multiple related entities (such as other papers by author A, co-authors, etc.) and the relationships between them (such as "co-authors," "citation relationships," etc.).

[0043] By identifying target entity data and executing queries, the system can accurately retrieve information relevant to the user's query from the document graph knowledge base, avoiding aimless searches in a large-scale knowledge base and thus improving query accuracy. The query results not only cover information about the target entity but also related entities and their relationships, providing the user with a clear and comprehensive answer. For example, when a user queries whether author A has published papers on topic X, the system can not only return directly related papers but also provide information on other related papers by author A and co-authors, greatly enriching the query results. Leveraging the efficient query capabilities of the graph database, the system can quickly locate and return relevant entity data and their relationships, significantly improving query efficiency and shortening user waiting time. Furthermore, the structural design of the document graph knowledge base allows for flexible expansion of the query scope. If needed, the system can further delve into more attributes and relationships of related entities to meet complex user query needs. This fast and accurate query experience not only improves user satisfaction but also reduces repeated queries caused by inaccurate or incomplete results, effectively improving work efficiency.

[0044] Through the above steps and beneficial effects, step S200 demonstrates how the present invention provides users with more accurate and comprehensive query services through the query of the document class graph knowledge base, while improving system efficiency and user satisfaction.

[0045] Step S300: Based on the preset large language model, the final question-and-answer result is obtained according to the relevant entity data and the corresponding relevant entity relationships.

[0046] Specifically, in step S200, target entity data, related entity data, and relationships relevant to the user's query are obtained from the document class graph knowledge base. This data includes entity attributes, relationships between entities, etc. For example, the query results may include detailed information about author A (such as name and affiliation), detailed information about the paper (such as title, abstract, publication year, and journal name), and co-authorship relationships between author A and other authors.

[0047] Integrate these related entity data and their relationships into a structured dataset. This dataset may include entity attributes, relationships between entities, and the types of these relationships (such as "WROTE", "CO_AUTHOR", etc.).

[0048] For example, the integrated dataset might include the following: author A's name and affiliation; the paper's title, abstract, publication year, and journal name; author A's co-authorship with other authors and their related papers.

[0049] The integrated dataset is used as input and submitted to a pre-defined large language model. The large language model then generates question-and-answer results in natural language based on the input data. The generated results not only directly answer the user's question but also include relevant background information to help the user better understand the query results.

[0050] Large language models transform structured data into natural language, making query results easier for users to read and understand. Users can directly obtain clear and concise answers without having to interpret complex graph database query results. Large language models can combine information from multiple sources to generate richer and more comprehensive answers. Besides directly addressing the user's question, they can also provide background information and relevant details to help users gain a deeper understanding of the query content. This not only allows users to quickly obtain the information they need but also improves user satisfaction because the answers are presented in natural language. At the same time, fast and accurate query results reduce user waiting time and improve work efficiency. Furthermore, large language models can generate more accurate answers based on context and related entity information, reducing the risk of misunderstandings and ambiguities caused by isolated or incomplete data. For complex query needs, it can integrate information from multiple sources to output comprehensive answers that meet diverse user needs.

[0051] In this embodiment, the graph knowledge base question-answering optimization method, upon receiving a user's query information, first performs an intent recognition process to understand the user's query purpose or need, i.e., the recognition result. Intent recognition is typically achieved through a machine learning model, which may have been trained to identify different query types and intents.

[0052] Based on the obtained intent recognition results, a corresponding preset graph knowledge base is selected for querying. A preset graph knowledge base refers to a pre-built knowledge base used to store specific types of information. The selection of the preset graph knowledge base may be based on the degree of matching between the information types stored in the knowledge base and the user's query intent.

[0053] If the selected preset graph knowledge base is the document class graph knowledge base, the system will determine the target entity data to be queried (such as including nouns, text names, etc.) based on the user's query information, and input this data as query conditions into the document class graph knowledge base for querying.

[0054] After querying the document class diagram knowledge base, you can obtain other entity data related to the target entity data, as well as the relationships between these entities. The query results from the document class diagram knowledge base may include entity attributes, connection relationships between entities, etc.

[0055] Finally, using a pre-defined large language model, the relevant entity data and their relationships obtained from the query are taken as input to generate the final question-and-answer result. The large language model is an artificial intelligence model with language understanding and generation capabilities, capable of generating responses in natural language form based on the input information.

[0056] In summary, this invention, through intent recognition and the selection of a pre-defined graph knowledge base, can accurately pinpoint the core intent of a user's query, thereby quickly determining the corresponding graph knowledge base and avoiding ineffective searches in irrelevant knowledge bases, significantly improving query efficiency. Intent recognition technology ensures the system accurately understands the user's query needs, providing precise guidance for subsequent queries and ensuring the accuracy of query results. Furthermore, a large language model is used to transform the retrieved entity data and relationships into natural and accurate question-and-answer results. The model's powerful language organization capabilities enable it to generate answers that conform to human language habits based on input data, thereby enhancing the user experience. Specifically, the document-type graph knowledge base uses a graph structure to store knowledge, efficiently representing the complex relationships between queried entities. When a document-type graph knowledge base is identified, the system can quickly query relevant entities and their relationships based on the target entity data, comprehensively acquiring information related to the query.

[0057] This method flexibly adapts to different query needs, providing effective query services through intent recognition and knowledge base selection, whether for text-based or project information-based queries. Furthermore, the combination of graph databases and large language models effectively processes and analyzes massive amounts of data, providing users with more accurate decision support.

[0058] Optionally, the construction process of the document class graph knowledge base includes: Obtain document data and segment the document data to obtain multiple file blocks; Input all the file blocks into the preset extraction model to obtain the corresponding entity data and the corresponding entity relationships; The document class graph knowledge base is constructed based on all the entity data and the corresponding entity relationships.

[0059] Specifically, the document data sources for building the graph knowledge base are collected. This document data can be text files in various formats, such as TXT, PDF, and Word documents, and the content may include academic papers, news articles, and internal company reports. The acquired document data is then segmented into multiple file blocks. Segmentation can be based on the document's physical structure (e.g., chapters, paragraphs) or content relevance (e.g., topics, keywords). This segmentation aims to make the document data easier to process and analyze, facilitating subsequent entity extraction and relationship identification.

[0060] The segmented file blocks are input into a pre-defined extraction model. This extraction model can be an NLP model based on machine learning or deep learning, used to identify entities (such as names of people, places, organizations, keywords, etc.) and their relationships within the file blocks. In other words, it extracts useful information from the document data, including entity data and the relationships between entities, providing structured data for building a graph knowledge base.

[0061] Based on the extracted entity data and entity relationships, a document-level graph knowledge base is constructed. This process includes creating nodes (representing entities) and edges (representing relationships between entities) in the graph database and defining the attributes of the nodes and edges. In other words, the extracted entity data and relationships are organized in the form of a graph, forming a structured graph knowledge base that facilitates subsequent querying and analysis.

[0062] By constructing a document-based graph knowledge base, a large amount of unstructured document data can be transformed into structured knowledge, significantly improving data utilization and value. Graph structures possess powerful relational expression capabilities, naturally revealing complex relationships between entities, thereby enhancing knowledge representation efficiency. Based on the complex query functions of the graph knowledge base, the system can effectively handle question-and-answer requests involving multiple entities and relationships, greatly improving the flexibility and accuracy of the question-and-answer system. Simultaneously, with the deep analysis and exploration functions of the graph knowledge base, patterns and relationships hidden in document data are revealed, powerfully promoting the discovery of new knowledge. Furthermore, the graph knowledge base provides decision-makers with comprehensive and accurate information support, helping them make better decisions.

[0063] Through the above steps, the construction process of the document class graph knowledge base provides a structured and efficient foundation for knowledge storage and retrieval, thereby improving the performance and user experience of the question-answering system.

[0064] Optionally, the step of obtaining the final question-and-answer result based on the preset large language model, according to the relevant entity data and the corresponding relevant entity relationships, includes: Based on the document class graph knowledge base, the relevant entity data and the corresponding relevant entity relationships are obtained according to the target entity data and the preset number of query steps; The relevant entity data and the target entity data are matched with the preset database to obtain the corresponding relevant file blocks; The target entity data, the corresponding related file blocks, the related entity data, and the corresponding related entity relationships are input into the preset large language model to obtain the final question-answering result. The preset database includes multiple file blocks.

[0065] Specifically, based on the document class graph knowledge base, and combined with the target entity data and the preset number of query steps, the relevant entity data and their corresponding relationships are retrieved.

[0066] For example: If the target entity is an author A, and the preset query steps are 2, then first find the papers directly associated with A and the co-authors (step 1); then further find the citation information of these papers, other works of the co-authors, etc. (step 2).

[0067] In the document class graph knowledge base, the query scope is expanded according to the preset number of query steps, with the target entity as the center, to obtain entity data and relationships that are directly and indirectly related to the target entity, providing a data foundation for generating comprehensive and accurate question and answer results.

[0068] The obtained relevant entity data and target entity data are matched with a preset database to filter out file blocks containing these entity data.

[0069] For example, suppose a pre-defined database contains a large number of paper file blocks, each with a unique ID and containing entity information such as author and paper title. By matching the IDs of target author A and its related papers and co-authors, a specific paper file block containing this information can be quickly identified.

[0070] Using a pre-defined database, specific file blocks containing relevant entity data are located to provide detailed background information and contextual support for subsequent question-and-answer results.

[0071] The target entity data, related file blocks, related entity data and their relationships are taken as input and submitted to a pre-defined large language model to generate the final question-and-answer results.

[0072] For example, after inputting data including detailed information about author A and co-author B, as well as relevant paper file blocks, the large language model generates results similar to the following: "Author A and co-author B co-authored a paper titled 'Research Methods' in 2022, published in the journal Science, which mainly explored important issues in the relevant field. In addition, the paper has been cited by many subsequent studies and has had a certain impact on the development of the field." Leveraging the powerful language generation and understanding capabilities of large language models, based on relevant input information, we generate logically coherent, content-rich, and user-relevant question-and-answer results that meet user needs and query intent.

[0073] By setting a preset number of query steps in the graph knowledge base, we can obtain entity data and their relationships that are directly and indirectly related to the target entity, thus ensuring the accuracy and comprehensiveness of the question-and-answer results. Simultaneously, by matching with a pre-defined database to obtain corresponding file blocks, we can provide results containing detailed background and contextual information, avoiding isolated and incomplete information. The large language model can generate logically coherent and fluent question-and-answer results that conform to human language habits, allowing users to obtain clear answers directly without having to interpret complex graph database query results, thereby improving user satisfaction. For complex queries, the model can integrate information from multiple sources to generate comprehensive answers that meet diverse user needs. Setting a preset number of query steps helps control the query scope, preventing information overload or irrelevant information due to overly broad queries. Furthermore, matching file blocks in the pre-defined database can accurately locate files, reduce processing of irrelevant data, improve query efficiency, and shorten user waiting time.

[0074] Optionally, the graph knowledge base question answering optimization method further includes: When the preset graph knowledge base is a general graph database, historical query data for the corresponding preset time period is obtained, and the historical query data is analyzed to obtain the analysis results; Based on the analysis results and the query information, the corresponding potential intent information is determined; Based on the general class graph database, a final query statement is generated according to the potential intent information and the query information; The final query statement is input into the large language model to obtain the corresponding final question and answer result; The construction process of the general class graph database includes: The system acquires user-uploaded data, determines node data and corresponding node relationship data based on the uploaded data, and constructs the general class graph database based on the node data and the corresponding node relationship data.

[0075] Specifically, when the preset graph knowledge base is determined to be a general class graph database, historical query data within a preset time period is retrieved. This historical query data contains all query records of users on the general class graph database within that time period.

[0076] For example, retrieve all user query records for the organizational chart knowledge base within the past month, including query time, query content, and query results. Analyzing historical query data helps understand users' query habits and common query patterns, providing a basis for subsequently determining potential intent information.

[0077] Based on the analysis results and the current user's query information, we can determine the user's potential intent. Potential intent may include the user's query preferences for specific entities and their query needs regarding entity relationships. For example, by analyzing historical query data, we can discover that users frequently query employee information and their hierarchical relationships within a particular department. When the user queries that department again, we can infer that their potential intent is to understand the department's latest employee composition and hierarchical relationships. In other words, by understanding the user's potential query intent, we can support the generation of more accurate query statements.

[0078] Based on a standard class graph database, the final query statement is generated according to potential intent information and user query information. The query statement will explicitly specify the entities, relationships, and query conditions to be queried.

[0079] For example: Based on the user's query "Query employee information for department D" and potential intent information "Understand the latest employee composition and hierarchical relationships in department D", generate the following query statement: For example, "MATCH (d:Department {name: 'D'})" <-[:belongsTo]-(e:Employee)-[:reportsTo]->(m:Employee) RETURN e, m".

[0080] MATCH is used to find nodes and relationships that match a specified pattern.

[0081] (d:Department {name: 'D'}): This operation matches a node of type Department with its name attribute set to 'D'. This means we want to find department nodes with the name 'D'. For example, department D could be "Sales" or "Marketing", etc.

[0082] <- [:belongsTo]-: This operation matches a relationship of type "belongsTo" pointing to the department node d. This means we want to find employee nodes that have a "belongsTo" relationship with department D. The "belongsTo" relationship indicates that the employee belongs to that department.

[0083] (e:Employee): This operation matches a node of type Employee. This means we are looking for an employee node. For example, an employee could be "Zhang San" or "Li Si", etc.

[0084] -[:reportsTo]->: This operation matches a relationship of type reportsTo, pointing to another employee node m. This means we want to find the parent employee node m that has a reportsTo relationship with employee e. The reportsTo relationship indicates that employee e reports to employee m.

[0085] (m:Employee): This operation matches a node of type Employee. This means we are looking for the parent employee node. For example, the parent employee might be "Wang Wu" or "Zhao Liu", etc.

[0086] RETURN e, m: This operation returns the matched employee node e and its parent employee node m. This means the query results will include all employees related to department D and their direct superiors.

[0087] The final query is input into a pre-defined large language model, which generates the final question-and-answer results. The large language model then answers the user's question in natural language based on the query results.

[0088] The construction process of a standard class graph database involves: acquiring user-uploaded data and determining node data and corresponding node relationship data based on this data; then, constructing the standard class graph database based on the node data and node relationship data.

[0089] For example, a user uploads organizational structure data for a company, including employee information, department information, and hierarchical relationships. Based on this data, nodes (such as employee nodes and department nodes) and relationships (such as belonging relationships and reporting relationships) are created, constructing a general-purpose graph database. This provides a structured data foundation for the question-and-answer system, supporting the querying and analysis of structured information such as organizational structure.

[0090] By analyzing historical query data, the system accurately grasps users' potential intentions, generating more precise query statements and improving the accuracy and relevance of query results. The large language model transforms query results into natural language, allowing users to intuitively obtain the information they need without deciphering complex graph database queries, thereby improving user satisfaction. Based on user query habits and historical data, the system dynamically adjusts query strategies to meet the diverse needs of different users. The construction of a general-purpose graph database makes structured data queries more efficient; the system can quickly locate and return relevant entity data and their relationships, reducing user waiting time. Furthermore, the construction and use of general-purpose graph databases helps enterprises optimize the management and utilization of structured data, further improving knowledge management efficiency.

[0091] It should be noted that in practical applications, the preset graph knowledge base (including the document graph knowledge base and the general graph database) needs to be updated in real time based on the constantly updated text data or architectural data to ensure the accuracy and timeliness of the information.

[0092] In this context, "query rewriting" refers to the intelligent modification and optimization process performed by the system on the user's original query (especially a fuzzy query). The following is a detailed explanation: Optionally, generating the final query statement based on the general class graph database, according to the latent intent information and the query information, includes: Based on the general class graph database, determine the relevant node data and the corresponding relevant node relationship data corresponding to the query information; An initial query statement is generated based on the potential intent information, the relevant node data, and the corresponding relevant node relationship data. The initial query statement is verified and modified to obtain the final query statement.

[0093] Specifically, based on a standard class graph database, the relevant node data and corresponding node relationship data corresponding to the user's query information are determined. That is, in the standard class graph database, the nodes and relationships directly related to the user's query are found, providing the basic data for generating the query statement.

[0094] By combining potential intent information, relevant node data, and corresponding relationship data, an initial query statement is generated. This statement explicitly specifies the entities, relationships, and query conditions to be queried. A preliminary query statement expressing the user's query intent is thus formed, but this statement may require further verification and refinement.

[0095] The initial query statement is validated to check its syntactic correctness, logical rationality, and accurate expression of the user's intent. Based on the validation results, the statement is modified as necessary, such as adjusting query conditions and correcting syntax errors, to obtain the final query statement. This process ensures that the query statement can be executed accurately and returns results that match the user's intent.

[0096] The validation and modification process ensures that the query statement accurately reflects the user's intent, improving the accuracy and relevance of the query results. The optimized final query statement can quickly locate and return relevant entity data and their relationships, effectively reducing user waiting time. The validation process guarantees the syntactic correctness and logical rationality of the query statement, preventing erroneous statements from causing system anomalies or incorrect results. Furthermore, generating query statements by combining potential intent information can better meet the diverse query needs of different users, enhancing the system's flexibility and adaptability.

[0097] Optionally, determining the target entity data based on the query information includes: The query information is then transformed to obtain temporary vector data; The temporary vector data is compared with the entity vectors in the preset vector library to obtain the entity vector with the highest similarity, and the entity vector with the highest similarity is used as the corresponding target entity data.

[0098] Specifically, the query information is input into a pre-trained language model or vectorization tool to generate temporary vector data that can represent the semantics of the query. For example, models such as BERT, Word2Vec, and Embedding models can be used to convert the query text into fixed-dimensional vectors.

[0099] Similarity calculations are performed between temporary vector data and entity vectors in a pre-defined vector library. Common methods include cosine similarity and Euclidean distance. For example, the cosine similarity between a temporary vector and an entity vector can be calculated to find the entity vector with the highest similarity.

[0100] Based on the similarity calculation results, the entity vector with the highest similarity is selected as the corresponding target entity data. For example, if the query information is "XX mobile phone", after vectorization and similarity calculation, the entity vector with the highest similarity is determined to be "XXPhone".

[0101] Vectorization captures the semantic features of query information, helping the system understand user intent at the semantic level and improving query accuracy. By leveraging vector similarity calculation, the system can quickly locate the most relevant entity in a large-scale entity vector database, improving query efficiency. The vector similarity-based matching mechanism can intelligently identify key entities in the query information, achieving accurate matching even with ambiguous query expressions.

[0102] Optionally, determining the relevant node data and corresponding relevant node relationship data corresponding to the query information based on the general class graph database includes: Based on the query information, search the graph database for the top K most relevant names (nodes) (topK, where K can be set according to specific circumstances), and obtain the relationship data of the K names (nodes). Based on the obtained nodes and their corresponding relationship data, an initial query statement is generated in conjunction with the potential intent information.

[0103] The above process is mainly due to the fact that users' original queries may not be precise or clear enough for various reasons, such as vague wording, typos, or incomplete expression, making it difficult to accurately understand their true intentions. For example, a user might enter "query information about XX mobile phone," which is a broad query and it may be difficult to determine whether the user wants to know about the price, performance, new model releases, or specific information such as user reviews of XX mobile phone.

[0104] Based on the user's original query (query information), the top K most relevant names are searched in the graph database. These names are considered to be the entities or concepts most closely related to the user's query intent. For example, in the above example, the top K names might be "XX mobile phone (a model of a certain brand of mobile phone) price", "iXX mobile phone performance", and "XX mobile phone user reviews".

[0105] Then, the ambiguous or inaccurate parts of the original query are replaced with these topK names that are consistent with the data in the graph database, making the original query more accurate and standardized, and better able to match the entities and relationships in the graph database.

[0106] It retrieves information from the user's long-term memory to understand their historical query behavior, interests, and preferences. Simultaneously, it further personalizes the original query by incorporating the current context (such as relevant content mentioned by the user in previous conversations). For example, if the user has consistently searched for the performance parameters of a certain mobile phone, the original query might be rewritten to include performance-related content, such as "the processor performance of the XX mobile phone," making the rewritten query more relevant to the user's individual needs and the current query context, thus further improving the accuracy and relevance of the query.

[0107] Optionally, the graph knowledge base question answering optimization method further includes: The final query statement is evaluated using a pre-defined cache library. Determine whether the question and answer results corresponding to the final query statement exist in the preset cache library; If it exists, the returned question and answer result will be used as the final query statement.

[0108] Specifically, the generated final query statement is sent to a pre-defined cache for querying. This is to quickly determine whether the query statement has been executed and whether corresponding question-and-answer results exist.

[0109] If a matching query result exists in the cache, it will be retrieved and returned directly. This avoids repeatedly executing the same query and saves query time. If no matching query result exists in the cache, the step of inputting the final query statement into the large language model will be executed.

[0110] Leveraging the fast query capabilities of a caching library, the existence of readily available answers can be determined instantly, avoiding the need to execute the entire query process each time and significantly improving query efficiency. This not only reduces redundant processing of the same queries, effectively lowering system resource consumption, but also shortens user wait times and improves system response speed. Users can obtain the information they need more quickly, thereby increasing user satisfaction. Furthermore, the caching mechanism helps balance system load, especially in high-concurrency query scenarios, effectively reducing database pressure and ensuring stable system operation.

[0111] In some embodiments, such as Figure 2 As shown, the process of the graph knowledge base question answering optimization method is as follows: Figure 2 The overall process can be divided into two main stages: the construction of the document class diagram knowledge base and the question-and-answer process.

[0112] Map construction phase: Obtain the text content of the original files, which can be various types of documents, such as papers, articles, reports, etc.

[0113] Chunking (file block splitting): The original file text content is split into multiple file blocks. This facilitates subsequent processing and analysis, making the data easier to manage and index.

[0114] input (input large language model): Inputs the file block into the preset large language model.

[0115] extract-entity and extract-relation: These methods use a large language model to extract entities and relations from file blocks, resulting in entities and entity-relation relationships.

[0116] The construction of the vector database, knowledge graph (document class graph knowledge base), and SQLite database (pre-defined database) includes: The extracted entity and relationship information is stored in a vector database, a knowledge graph, and an SQLite database, respectively. These databases will be used for subsequent queries and question answering. The vector database stores vector representations of entities for fast similarity searching. The knowledge graph stores entities and their relationships, organizing the data in a graph structure. The SQLite database stores structured data such as file blocks for easy retrieval and management.

[0117] Question and Answer Phase: Obtaining the user's query question (query information); search (Search related entities, relationships, and file blocks): Search for related entities, entity relationships, and file blocks in vector databases, knowledge graphs, and SQLite databases based on the user's question.

[0118] merge (merge context): merges the searched related entities, entity relationships and file blocks to form complete context information.

[0119] input (input large language model): Input the merged context information into the preset large language model.

[0120] Summarize (Generate Final Answer): The large language model generates the final question-and-answer result based on the contextual information of the input and presents it to the user in natural language.

[0121] Where n_hops: the number of steps to move in the graph, which is the preset number of query steps.

[0122] Neighbors: The neighboring nodes are the nodes that can be reached from the current node in n_hops steps.

[0123] has-edge: Whether there is an edge between nodes, that is, to determine whether there is a relationship between entities.

[0124] merge: Combines related entities, relationships, and text blocks into a single source of data for the problem.

[0125] source_id: The source ID of the entity, i.e., the ID of the file block where the entity is located.

[0126] By following the steps above, a document-based knowledge graph can be constructed, and an efficient question-and-answer function can be implemented based on this graph.

[0127] In other embodiments, such as Figure 3 As shown, the process of the graph knowledge base question answering optimization method is as follows: The question-and-answer process in a general class diagram database; Upload JSON data: Users upload data in JSON format to the system. Example: A user uploads a JSON file containing employee information, department information, and their relationships.

[0128] Parse the uploaded JSON data and extract the entity and relationship information (nodes and node relationships).

[0129] Create nodes representing employees in a graph database (GraphDB - a general-purpose graph database). For example, create an Employee node with attributes including name and position.

[0130] Create department nodes: Create nodes representing departments in the graph database. Based on the parsed data, construct relationships between entities in the graph database. Common relationship types include belongsTo, supervises, and containsSubDepartment.

[0131] Build other custom relationships as needed, such as `inChargeOf` (responsible for). For example, build the relationship `inChargeOf` to indicate that an employee is responsible for a certain project or task.

[0132] The graph construction is complete, and the graph database now contains all the necessary nodes and relationships.

[0133] Question and answer phase: Obtaining user queries.

[0134] Search the graph database for the top K names related to the query question.

[0135] Based on the top K names found and their corresponding historical query data, rewrite the user's original query. Check the cache for question-and-answer results corresponding to the rewritten query.

[0136] If no results are found in the cache, the generated graph query is sent to the Large Language Model (LLM) to obtain the final query results. The query results are also cached so that future identical queries can directly use the cached results. Finally, the query results are returned to the user in natural language.

[0137] Caching mechanisms effectively avoid duplicate queries while reducing resource consumption during the query process, significantly improving query efficiency. Leveraging the efficient query capabilities of graph databases ensures accurate and relevant query results. Caching mechanisms also balance system load, especially in high-concurrency query scenarios, strongly guaranteeing stable query operation while allowing users to quickly obtain the information they need, thus improving user satisfaction.

[0138] like Figure 4 As shown in the figure, an embodiment of the present invention provides a graph knowledge base question answering optimization device, comprising: The identification unit is used to identify the intent of the query information when the query information is detected, obtain the identification result, and determine the corresponding preset graph knowledge base based on the identification result. The processing unit is configured to, when the preset graph knowledge base is a document class graph knowledge base, determine the target entity data according to the query information, input the target entity data into the document class graph knowledge base for querying, obtain relevant entity data and corresponding relevant entity relationships; and, based on a preset large language model, obtain the final question-and-answer result according to the relevant entity data and the corresponding relevant entity relationships.

[0139] like Figure 5 As shown in the figure, an embodiment of the present invention provides a graph knowledge base question answering optimization device, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the graph knowledge base question answering optimization method as described above when the computer program is executed.

[0140] Alternatively, a graph knowledge base question-answering optimization device includes a memory and a processor coupled to the memory; the memory is configured to store a computer program; the processor is configured to perform the following operations when the computer program is executed: When query information is detected, intent recognition is performed on the query information to obtain recognition results, and the corresponding preset graph knowledge base is determined based on the recognition results; When the preset graph knowledge base is a document class graph knowledge base, the target entity data is determined according to the query information, and the target entity data is input into the document class graph knowledge base for querying to obtain relevant entity data and corresponding relevant entity relationships; Based on a pre-defined large language model, the final question-and-answer result is obtained according to the relevant entity data and the corresponding relevant entity relationships.

[0141] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the graph knowledge base question-answering optimization method as described above.

[0142] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations: When query information is detected, intent recognition is performed on the query information to obtain recognition results, and the corresponding preset graph knowledge base is determined based on the recognition results; When the preset graph knowledge base is a document class graph knowledge base, the target entity data is determined according to the query information, and the target entity data is input into the document class graph knowledge base for querying to obtain relevant entity data and corresponding relevant entity relationships; Based on a pre-defined large language model, the final question-and-answer result is obtained according to the relevant entity data and the corresponding relevant entity relationships.

[0143] The following describes a graph knowledge base question-answering optimization device that can serve as a server or client of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The graph knowledge base question-answering optimization device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The graph knowledge base question-answering optimization device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0144] The graph knowledge base question-answering optimization device includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0145] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0146] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for optimizing a graph knowledge base question answering, characterized in that, The method comprises the following steps: When the query information is detected, the query information is subjected to intention recognition to obtain an identification result, and a corresponding preset graph knowledge base is determined according to the identification result; When the preset graph knowledge base is a document type graph knowledge base, target entity data is determined according to the query information, and the target entity data is input into the document type graph knowledge base for query to obtain relevant entity data and corresponding relevant entity relationships; Based on a preset large language model, a final question and answer result is obtained according to the relevant entity data and the corresponding relevant entity relationships.

2. The graph knowledge base question answering optimization method of claim 1, wherein, The construction process of the document type graph knowledge base comprises: Obtaining document data and segmenting the document data to obtain a plurality of file blocks; All the file blocks are input into a preset extraction model to obtain corresponding entity data and corresponding entity relationships; The document type graph knowledge base is constructed according to all the entity data and the corresponding entity relationships.

3. The graph knowledge base question answering optimization method according to claim 2, characterized in that, Based on the document type graph knowledge base, the target entity data and a preset query step number are used to obtain the relevant entity data and the corresponding relevant entity relationships; The relevant entity data and the target entity data are matched with the preset database to obtain corresponding relevant file blocks; The target entity data, the corresponding relevant file blocks, the relevant entity data and the corresponding relevant entity relationships are input into the preset large language model to obtain the final question and answer result; The preset database comprises a plurality of file blocks. The graph knowledge base question and answer optimization method further comprises:

4. The graph knowledge base question answering optimization method of claim 1, wherein, When the preset graph knowledge base is a general type graph database, historical query data of a corresponding preset time period is obtained, and the historical query data is analyzed to obtain an analysis result; According to the analysis result and the query information, corresponding potential intention information is determined; Based on the general type graph database, a final query statement is generated according to the potential intention information and the query information; The final query statement is input into the large language model to obtain the corresponding final question and answer result; The construction process of the general type graph database comprises: Obtaining user uploaded data, determining node data and corresponding node relationship data according to the uploaded data, and constructing the general type graph database according to the node data and the corresponding node relationship data. Based on the general type graph database, a final query statement is generated according to the potential intention information and the query information, which comprises:

5. The graph knowledge base question answering optimization method according to claim 4, characterized in that, According to the general type graph database, relevant node data corresponding to the query information and corresponding relevant node relationship data are determined; According to the potential intention information, the relevant node data and the corresponding relevant node relationship data, an initial query statement is generated; The initial query statement is verified and modified to obtain the final query statement. The target entity data is determined according to the query information, which comprises:

6. The graph knowledge base question answering optimization method according to claim 1, characterized in that, The query information is converted to obtain temporary vector data; ​ The temporary vector data is subjected to similarity calculation with entity vectors in a preset vector library, to obtain an entity vector with the highest similarity, and the entity vector with the highest similarity is taken as corresponding target entity data.

7. The graph knowledge base question answering optimization method according to claim 5, characterized in that, The graph knowledge base question answering optimization method further includes: The final query statement is judged by a preset cache library, It is judged whether the question and answer result corresponding to the final query statement exists in the preset cache library; If so, the returned question and answer result is taken as the final query statement.

8. A graph knowledge base question-answering optimization device, characterized in that, It includes: An identification unit is configured to, when detecting query information, perform intent identification on the query information to obtain an identification result, and determine a corresponding preset graph knowledge base according to the identification result; A processing unit is configured to, when the preset graph knowledge base is a document type graph knowledge base, determine target entity data according to the query information, input the target entity data into the document type graph knowledge base for query, and obtain related entity data and corresponding related entity relationships; Based on a preset large language model, a final question and answer result is obtained according to the related entity data and the corresponding related entity relationships.

9. A graph knowledge base question answering optimization device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the graph knowledge base question answering optimization method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and when the computer program is executed by the processor, the graph knowledge base question answering optimization method of any one of claims 1 to 7 is implemented.

Images