Question and answer data generation method and device based on conceptual graph, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to the medical field and the financial insurance field, and discloses a question and answer data generation method and device based on a conceptual graph, equipment and a storage medium. The method comprises the steps of extracting question and answer pair data from seed data of a preset field to construct a seed data set; extracting a theme set and a knowledge point set from the seed data set, and constructing a mapping relationship between themes in the theme set and knowledge points in the knowledge point set; constructing a conceptual graph based on the topic set, the knowledge point set and the mapping relation, and calculating the weight of edges between nodes in the conceptual graph to obtain a weighted conceptual graph; performing semantic random walk and combined sampling on the weighted conceptual graph to generate a combined path set; and generating a question and answer data sample set based on the combined path set through the large language model, and verifying the question and answer data sample set to generate a target question and answer data sample set. The question and answer data generation accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and can be applied to the medical and financial insurance fields. In particular, it relates to a method, apparatus, device and storage medium for generating question-and-answer data based on concept graphs. Background Technology

[0002] In today's intelligent services, high-quality data is the cornerstone of training reliable artificial intelligence systems. This challenge is particularly prominent in highly specialized fields such as finance, insurance, and healthcare. Existing technologies mainly suffer from the following problems: In the finance and insurance sector, intelligent customer service and knowledge-based question-and-answer systems need to handle complex insurance terms, claims rules, and premium calculation logic. Currently, the training data for these systems primarily relies on manually compiled historical question-and-answer records or rewrites based on fixed templates. This approach has significant limitations: First, the generated data lacks diversity, making it difficult to cover complex reasoning scenarios such as "claims for specific diseases during the waiting period" or "liability determination when multiple insurance types overlap"; second, data updates are lagging, failing to quickly respond to frequent updates to insurance terms, resulting in inaccurate answers and weak generalization ability in actual business operations. More seriously, existing methods lack systematic modeling of the logical relationships between concepts within the insurance field, making it difficult to guarantee the logical consistency of the generated data. In the healthcare field, AI models used for patient consultation, assisted diagnosis, or health management also face the dilemma of insufficient training data. The medical field involves complex logical relationships between symptoms, drugs, examination indicators, and treatment plans. Existing technologies struggle to systematically generate question-and-answer data that comprehensively infers information based on patient history, current symptoms, and medication contraindications. Simple content rewriting cannot guarantee the accuracy and logical consistency of medical knowledge, posing a serious risk of generating erroneous medical advice. Furthermore, the rapid updating and iteration of medical knowledge makes it difficult for traditional data generation methods to keep up in a timely manner.

[0003] The core problem shared by these fields is that existing technologies lack structured representations and reasoning mechanisms for domain-specific knowledge. The data generation process fails to effectively model the semantic relationships between concepts within the domain, resulting in incomplete coverage, logical inconsistencies, and low accuracy in the generated question-and-answer data. Summary of the Invention

[0004] The purpose of this application is to provide a question-and-answer data generation method, apparatus, device, and storage medium based on a concept graph, so as to improve the accuracy of question-and-answer data generation.

[0005] To address the aforementioned technical problems, embodiments of this application provide a question-and-answer data generation method based on a concept graph, comprising: Question-answer pairs are extracted from seed data in a predefined domain to construct a seed dataset; Extract a topic set and a knowledge point set from the seed dataset, and construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set; A concept graph is constructed based on the topic set, the knowledge point set, and the mapping relationship, and the weights of the edges between nodes in the concept graph are calculated to obtain a weighted concept graph. Perform semantic random walks and combined sampling on the weighted concept graph to generate a set of combined paths; A question-and-answer data sample set is generated based on the combined path set using a large language model, and the question-and-answer data sample set is validated to generate a target question-and-answer data sample set.

[0006] To address the aforementioned technical problems, embodiments of this application provide a question-and-answer data generation apparatus based on a concept graph, comprising: The seed dataset extraction module is used to extract question-answer pair data from seed data in a preset domain to construct a seed dataset. The mapping relationship construction module is used to extract the topic set and the knowledge point set from the seed dataset, and construct the mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set; The concept graph construction module is used to construct a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and to calculate the weight of the edges between nodes in the concept graph to obtain a weighted concept graph; A combined path set generation module is used to perform semantic random walks and combined sampling on the weighted concept graph to generate a combined path set; The question-and-answer data sample generation module is used to generate a question-and-answer data sample set based on the combined path set using a large language model, and to verify the question-and-answer data sample set to generate a target question-and-answer data sample set.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a computer device, including one or more processors; and a memory for storing one or more programs, such that the one or more processors implement the question-and-answer data generation method based on the concept graph described in any one of the above-mentioned methods.

[0008] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is: a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the question-and-answer data generation method based on the concept graph described above.

[0009] This invention provides a method, apparatus, device, and storage medium for generating question-and-answer data based on a concept graph. The method includes: extracting question-and-answer pairs from seed data in a preset domain to construct a seed dataset; extracting a topic set and a knowledge point set from the seed dataset, and constructing a mapping relationship between topics in the topic set and knowledge points in the knowledge point set; constructing a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and calculating the weights of edges between nodes in the concept graph to obtain a weighted concept graph; performing semantic random walks and combined sampling on the weighted concept graph to generate a combined path set; generating a question-and-answer data sample set based on the combined path set using a large language model, and validating the question-and-answer data sample set to generate a target question-and-answer data sample set. This invention, by constructing a concept graph model of domain knowledge and combining semantic random walks and large language model generation techniques, systematically solves the problems of incomplete coverage and logical confusion in professional domain question-and-answer data. It has the advantages of improving data generation quality, ensuring knowledge relevance, and improving the accuracy of question-and-answer data generation. Attached Figure Description

[0010] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for a question-and-answer data generation method based on a concept map according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the implementation of the question-and-answer data generation method based on a concept graph provided in this application embodiment; Figure 3 yes Figure 2 A flowchart illustrating a specific implementation method of step S2; Figure 4 yes Figure 2 A flowchart illustrating a specific implementation of step S3; Figure 5 yes Figure 4 A flowchart illustrating a specific implementation of step S33; Figure 6 yes Figure 2 A flowchart illustrating a specific implementation of step S4; Figure 7 yes Figure 2 A flowchart illustrating a specific implementation of step S5; Figure 8 yes Figure 2A flowchart illustrating another specific implementation method following step S5; Figure 9 This is a schematic diagram of a question-and-answer data generation device based on a concept map provided in an embodiment of this application; Figure 10 This is a schematic diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0013] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0015] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] It should be noted that the question-and-answer data generation method based on concept graphs provided in this application is generally executed by a server, and correspondingly, the question-and-answer data generation device based on concept graphs is generally configured in the server.

[0017] The concept map-based question-and-answer data generation method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server receives the client's question and generates answer data. In this invention, the server sends the answer data to the client. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0018] The question-and-answer data generation method based on concept graphs provided in this application can be applied to intelligent customer service and knowledge question-and-answer systems in the financial and insurance industry, or to intelligent consultation and triage scenarios in the medical field.

[0019] In existing technologies, intelligent services in specialized fields face the core challenge of insufficient data quality. Intelligent customer service systems in the financial and insurance sector rely on manually compiled question-and-answer records, making it difficult to cover complex scenarios such as claims for specific diseases during waiting periods and cross-liability determinations across multiple insurance types. In the healthcare sector, auxiliary diagnostic models are limited by template-based generation methods, failing to accurately express the logical connections between medical history, symptoms, and medication contraindications. Traditional data generation methods suffer from fragmented knowledge structures, lack of logical connection modeling, and insufficient verification of content accuracy, resulting in limited data diversity and the risk of knowledge errors.

[0020] To address these issues, researchers observed that domain-specific knowledge possesses networked interconnected characteristics, rendering single knowledge points insufficient to support complex reasoning needs. Analysis of cases where outdated insurance policy updates led to model failure revealed the need for a dynamic knowledge association system. In studies of medical misdiagnosis, it was found that the multi-path relationships between symptoms and treatment plans were not effectively modeled. This led to a design approach combining knowledge graph technology with generative models: first, constructing a conceptual network reflecting the topological relationships of domain knowledge; then, generating diverse data based on these network paths; and finally, performing semantic verification using a large language model.

[0021] Therefore, this application proposes to construct a seed dataset by extracting question-and-answer pairs from seed data in a predefined domain; extracting a set of topics and a set of knowledge points from the seed dataset and constructing a mapping relationship; constructing a concept graph based on the set of topics, the set of knowledge points, and the mapping relationship, and calculating the weights of the edges between nodes to obtain a weighted concept graph; performing semantic random walks and combinatorial sampling on the weighted concept graph to generate a set of combinatorial paths; and generating a set of question-and-answer data samples based on the set of combinatorial paths using a large language model and verifying it to generate a set of target question-and-answer data samples. Specifically, an initial dataset is constructed by parsing an insurance claims case database, and knowledge points such as insurance types and claims conditions are extracted using entity recognition technology. Historical consultation records are clustered by topic to form topic sets including car insurance claims and health insurance renewals. The co-occurrence frequency of knowledge points under each topic is statistically analyzed to establish a probability mapping between topics and knowledge points such as deductible calculation rules. A concept graph containing premium calculation nodes and claims process nodes is constructed, and the weights between nodes are calculated using a point mutual information algorithm. Path traversal based on transition probability is performed on the weighted concept graph to generate a combinatorial path containing waiting period disease diagnosis and insurance liability determination. The path is input into a large language model to generate specific question-answer pairs. The accuracy of the use of medical terminology is verified by a rule engine, and finally, question-answer data that conforms to clinical guidelines is output.

[0022] Please see Figure 2 , Figure 2 This paper illustrates a specific implementation of a question-and-answer data generation method based on concept maps.

[0023] It should be noted that if substantially the same result is obtained, the method of this invention is not based on... Figure 2 Limited to the order of the processes shown, this method includes the following steps: S1: Extract question-answer pairs from seed data in a predefined domain to construct a seed dataset.

[0024] Specifically, the seed dataset refers to an initial data set formed by filtering authoritative documents and historical question-and-answer records within the domain. This can be achieved by extracting question-and-answer fragments from insurance policy documents using natural language processing techniques, ensuring the domain suitability of the generated data. The preset neighborhood is set according to actual needs and is not limited here. In a specific embodiment, question-and-answer pairs can be extracted from seed data in a question-and-answer system in the financial insurance field, or from seed data in an intelligent consultation system in the medical field.

[0025] S2: Extract a topic set and a knowledge point set from the seed dataset, and construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set.

[0026] Specifically, a topic set refers to the core question categories abstracted from question-and-answer data. For example, it can be achieved by identifying high-frequency symptom topics in medical consultation records using clustering algorithms, and is used to establish the logical framework for knowledge organization. A knowledge point set refers to the smallest information unit that constitutes professional domain knowledge. For example, it can be achieved by using named entity recognition technology to extract entities such as drug names and examination indicators, serving as the basic nodes for building a knowledge network. A mapping relationship refers to the logical association between topics and knowledge points. Specifically, the association strength can be calculated by statistically analyzing the frequency of occurrence of knowledge points under a specific topic, and is used to construct the topological structure of the knowledge network.

[0027] Please see Figure 3 , Figure 3 A specific implementation of step S2 is shown below: S21: Input each question-answer pair data in the seed dataset into the large language model to parse the question-answer pair data through prompting engineering, generating the topic set and the knowledge point set. S22: Calculate the probability of each knowledge point in the knowledge point set appearing under a topic in the topic set, to construct a mapping relationship between topics in the topic set and knowledge points in the knowledge point set.

[0028] The "cue engineering" refers to the optimized cue strategy. Specifically, it can use structured instruction templates to guide large language models in performing semantic parsing tasks. By designing cue words that include domain terminology and relation types, it guides the model to identify implicit topic hierarchies and knowledge point elements from question-answer pairs. The "probabilistic statistical method" refers to association modeling based on conditional probability. Specifically, it can use Bayes' theorem to calculate the frequency of knowledge points appearing under specific topics, and quantify the strength of the association between knowledge points and topics by statistically analyzing the co-occurrence frequency and distribution ratio of knowledge points and topics in question-answer data.

[0029] Specifically, in the financial insurance scenario, when processing question-and-answer pairs involving "waiting period disease claims," ​​the prompting system guides the large language model to identify the "insurance liability determination" topic and related knowledge points such as "waiting period calculation" and "disease classification" through a pre-set insurance clause parsing template. Subsequently, the frequency of the "waiting period calculation" knowledge point under the "insurance liability determination" topic in all question-and-answer data is statistically analyzed, and the conditional probability value of this knowledge point under the corresponding topic is calculated to be 0.83, thus establishing a strong correlation mapping. In the healthcare scenario, for the "drug interactions" topic, statistical analysis reveals that the conditional probability of the "liver and kidney function indicators" knowledge point under related topics reaches 0.76, revealing the core position of this knowledge point in medication safety assessment.

[0030] Traditional methods rely on manual labeling of topics or rule-based keyword matching, which suffers from incomplete topic coverage and coarse-grained knowledge point extraction. For example, in the insurance field, manual labeling may miss the topic of "cross-liability of multiple insurance types," while rule-based methods cannot identify the implicit relationship between the knowledge point of "deductible adjustment" and the topic of "renewal clauses." The embodiments of this application, through the semantic understanding capabilities of a large language model, can automatically discover the logical relationship between the knowledge point of "genetic testing results" and the topic of "early cancer screening," a relationship that is not explicitly labeled in manually compiled datasets.

[0031] S3: Construct a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and calculate the weights of the edges between nodes in the concept graph to obtain a weighted concept graph.

[0032] Specifically, a weighted concept graph refers to a knowledge network that includes node weights and edge weights. For example, it can be implemented by using a point mutual information algorithm to calculate the correlation between diseases and examination items, and is used to guide the selection of data generation paths.

[0033] Please see Figure 4 , Figure 4 A specific implementation of step S3 is shown below: S31: All elements in the topic set and the knowledge point set are used as nodes to form the node set of the concept graph. S32: Based on the mapping relationship, the connection relationships between the topics in the topic set and the knowledge points in the knowledge point set are constructed to build the edges of the concept graph. S33: The weights of the edges between nodes in the concept graph are calculated to obtain the weighted concept graph.

[0034] In this context, a node set refers to a heterogeneous network node group composed of topics and knowledge points. This can be achieved by uniformly encoding both types of elements as vertices in a graph data structure, forming a network topology that covers the knowledge hierarchy. Edge weights are numerical indicators that quantify the strength of semantic associations between nodes. They can be implemented using a weighted fusion calculation of mutual information, co-occurrence strength, and embedding similarity, used to characterize the logical tightness between knowledge elements.

[0035] Specifically, in the implementation process in the healthcare field, themes such as "cardiovascular disease" and "diabetes management" are first used together with knowledge points such as "ECG abnormality range" and "insulin injection dosage calculation" as nodes. Based on the statistical results of the frequent occurrence of the knowledge point "ECG abnormality range" under the "cardiovascular disease" theme in electronic medical record question-and-answer data, a mapping relationship between the two is established as a connection edge. When calculating the weight of this edge, both the co-occurrence frequency of the two in historical question-and-answers and the vector similarity generated by the pre-trained model are considered, ultimately obtaining a quantitative value reflecting the strength of knowledge association. This construction method enables the hypertension diagnostic criteria to be strongly associated with the cardiovascular disease theme and to establish a secondary association with renal function test indicators, forming a multi-dimensional knowledge network structure.

[0036] Traditional methods typically construct knowledge graphs using single-dimensional keyword co-occurrence, such as using only disease names as nodes and symptoms as edge attributes. This results in a flattened knowledge hierarchy and unquantifiable association strength. This application introduces a two-tiered node structure of topic and knowledge point, preserving the hierarchical nature of the knowledge system while achieving cross-level semantic association modeling. A weighted calculation mechanism using multi-dimensional feature fusion overcomes the shortcomings of traditional TF-IDF methods that ignore semantic similarity, enabling effective connections to be established between semantically related but literally different knowledge points such as "drug interactions" and "contraindications for combined drug use." This application can accurately construct a network topology reflecting the professional knowledge structure. In medical applications, the generated question-and-answer data can cover both strongly correlated combinations like "hereditary disease screening" and "genetic testing indicators," as well as weakly correlated but reasonable knowledge combinations like "postoperative rehabilitation" and "nutritional supplement selection." In the financial and insurance field, weighted concept graphs can ensure that the knowledge points of "accident liability determination" and "insurance law clauses" under the theme of "auto insurance claims" are linked in a way that conforms to business logic. At the same time, weight control can avoid incorrect associations with "life insurance premium calculation", thereby improving the domain adaptability and logical rigor of the generated data.

[0037] Please see Figure 5 , Figure 5 A specific implementation of step S33 is shown below: S331: Collect co-occurrence data from the seed dataset and input node embedding vectors based on the concept graph using a pre-trained model. S332: Calculate the point mutual information and co-occurrence strength of each edge in the concept graph based on the co-occurrence data. S333: Calculate the embedding similarity between different node embedding vectors. S334: Calculate the weights of edges between nodes in the concept graph based on the point mutual information, the co-occurrence strength, and the embedding similarity to obtain the weighted concept graph.

[0038] Co-occurrence data refers to the frequency statistics of knowledge points and topics appearing together in the seed data. This can be achieved by using text matching algorithms to perform pattern recognition and frequency statistics on question-answer pairs, reflecting the explicit association between knowledge points and topics. Node embedding vectors refer to the vectorized representation of nodes in the concept graph using a pre-trained language model. This can be achieved by encoding node text using BERT or GraphSAGE models, capturing the distribution characteristics of nodes in the semantic space. Node mutual information is a measure of the statistical independence of two nodes in the co-occurrence data. This can be calculated by the logarithm ratio of the product of the joint probability and the marginal probability, quantifying the association strength between knowledge points and topics. Co-occurrence strength is a normalized index of the actual co-occurrence frequency of two nodes in the seed data. This can be calculated by the ratio of the co-occurrence frequency to the global average frequency, characterizing the actual observed strength of the association between nodes. Embedding similarity refers to the proximity of node embedding vectors in the semantic space. This can be calculated using cosine similarity or Euclidean distance, measuring the potential semantic association between nodes.

[0039] Specifically, the co-occurrence frequency of question-answer pairs in the seed dataset is first statistically analyzed to identify co-occurrence patterns between knowledge points and topics. The pre-trained model converts each node into a high-dimensional vector, preserving semantic information. Then, mutual information and co-occurrence strength are calculated separately; the former reveals statistical independence, while the latter reflects the actual co-occurrence level. Simultaneously, the similarity of node vectors is calculated to capture semantic associations. Finally, the statistical and semantic indicators are weighted and fused. For example, a linear weighting method can be used to combine mutual information, co-occurrence strength, and embedding similarity according to a preset ratio to generate a comprehensive weight value. This process, through multi-dimensional feature fusion, ensures that the edge weights simultaneously reflect explicit co-occurrence patterns and potential semantic associations.

[0040] Traditional methods typically rely solely on co-occurrence frequency or single semantic similarity to calculate edge weights, easily overlooking the difference between statistical independence and actual observations, and failing to effectively integrate semantic information. For example, simply using point mutual information may overestimate low-frequency but strongly correlated node pairs, while using only embedding similarity may introduce semantic drift. This application's embodiments, by integrating statistical and semantic features, preserve the co-occurrence characteristics of the data itself while leveraging the semantic understanding capabilities of pre-trained models, effectively balancing the weight contributions of explicit and implicit associations. This application's embodiments can accurately quantify the association strength between nodes in the concept graph, overcoming the limitations of single-indicator calculations and providing a reliable path selection basis for subsequent semantic random walks. By integrating multi-dimensional association features, it ensures that the generated weighted concept graph reflects both the actual data distribution patterns and captures deep-level semantic relationships, thereby improving the logical consistency and knowledge coverage of subsequent question-and-answer data generation.

[0041] S4: Perform semantic random walks and combined sampling on the weighted concept graph to generate a set of combined paths.

[0042] Specifically, semantic random walk refers to a graph traversal algorithm based on semantic associations, which can be implemented by adjusting the weight parameters in the transition probability matrix to discover potential knowledge combination paths. Combinatorial sampling refers to a multi-path collaborative generation mechanism, which can be implemented using the Monte Carlo method to sample medical treatment plan paths, thereby improving the diversity of data generation.

[0043] Please see Figure 6 , Figure 6 A specific implementation of step S4 is shown below: S41: Randomly select a topic node in the weighted concept graph as the starting point for the walk. S42: Calculate the probability of moving from the current node to the next stage, obtaining the walk probability. S43: Perform semantic random walks and combined sampling on the weighted concept graph based on the starting point, the walk probability, and a preset step size to generate the combined path set.

[0044] In this context, a theme node refers to an entity unit in the concept graph that contains a theme or knowledge point. The starting point of the path generation is the initial node, which can be implemented using a uniformly distributed random sampling algorithm to ensure equal coverage probability for nodes in different regions. The walk probability is a parameter indicating the likelihood of node transitions, which can be implemented using edge weight normalization to convert the weights of all outgoing edges from the current node into a probability distribution. The preset step size is the length threshold for path generation, which can be implemented using a fixed value or a dynamic adjustment strategy, such as setting the maximum step size as a percentage of the concept graph's diameter.

[0045] Specifically, in the node selection phase of the weighted concept graph, an arbitrary topic or knowledge point is selected as the starting point of the path through a uniform random sampling mechanism, breaking the path pattern problem caused by fixed starting points. During node transfer, the transfer probability distribution of adjacent nodes is calculated based on edge weights, prioritizing strongly related nodes corresponding to high-weight edges, while retaining access opportunities for weakly related nodes with low probability. When the length of the traversal path reaches the preset step size, the current path generation is terminated. Multiple traversal paths are cross-recombined through a combined sampling strategy to form a composite path set that includes core knowledge point associations and cross-topic extensions. For example, when generating a path in the medical field, it may start from the topic of "diabetes," pass through the knowledge point of "insulin resistance," and then jump to the topic of "cardiovascular complications," forming a cross-disciplinary knowledge association path.

[0046] Traditional methods generate paths using fixed-rule traversal or simple random walks, resulting in high path repetition rates and broken logical connections. For example, fixed-rule traversal algorithms can only generate linear paths in a preset order, failing to capture cross-topic knowledge connections; simple random walks ignore edge weight differences, leading to numerous invalid jumps in the generated paths. This application's embodiments introduce controllable randomness through a weighted probability transfer mechanism, preserving the strength of knowledge connections while ensuring the generated paths conform to domain knowledge logic and cover potential connection combinations. The combined sampling strategy further overcomes the limitations of single paths, constructing more complex knowledge network topologies through multi-path fragment recombination. This application effectively solves the problems of insufficient path diversity and broken logical connections in the generation of professional domain question-and-answer data. In the financial insurance field, it can generate cross-business paths covering "policy loan overdue processing" and "credit rating impact," ensuring that question-and-answer data includes multi-factor connection scenarios. In the healthcare field, it can form composite paths of "chronic disease management" and "drug interactions," supporting the generation of complex question-and-answer pairs requiring comprehensive medical history and medication knowledge. The path generation mechanism maintains the core knowledge logic through weighted transition probabilities, while expanding path diversity through combined sampling. The resulting question-and-answer data significantly improves the breadth of scenario coverage while maintaining professional accuracy.

[0047] S5: Generate a question-and-answer data sample set based on the combined path set using a large language model, and verify the question-and-answer data sample set to generate a target question-and-answer data sample set.

[0048] Specifically, a large language model refers to a pre-trained model with natural language generation capabilities. It can be implemented by fine-tuning a model based on the Transformer architecture, and is used to transform abstract paths into natural language question-answer pairs.

[0049] Please see Figure 7 , Figure 7 A specific implementation of step S5 is shown below: S51: Extract the topic and knowledge points of each combined path in the combined path set to obtain the target topic and target knowledge points. S52: Construct a question-and-answer template based on the target topic and target knowledge points, and generate content based on the question-and-answer target using the large language model to obtain the question-and-answer data sample set. S53: Perform consistency verification and constraint verification on the question-and-answer data sample set to generate the target question-and-answer data sample set.

[0050] The combination path refers to the sequence of nodes sampled from the weighted concept graph through semantic random walks, which can be implemented using graph traversal algorithms to ensure the logical connection between the generated content and the original knowledge system. The question-answering template is a structured generation framework containing the target topic and knowledge point elements, which can be implemented using natural language prompt templates to constrain the generation direction of the large language model. Consistency verification is the process of checking the internal logical consistency of the question-answer, which can be implemented using logical reasoning rule matching to eliminate self-contradictory erroneous content. Constraint verification is the process of checking whether the generated content conforms to domain knowledge specifications, which can be implemented using knowledge graph relationship verification to filter out erroneous expressions that violate professional rules.

[0051] Specifically, after the target topic and knowledge points are extracted from the combination path, they are transformed into question generation instructions and answer constraints through a pre-defined template structure. When the large language model generates candidate question-answer pairs within the template framework, its free generation space is limited to the knowledge points associated with the topic, avoiding the generation of content that deviates from the professional domain. In the verification phase, the consistency verification module checks whether the answer contradicts the knowledge points in the question through a logical rule engine, such as detecting the violation statement "Amoxicillin is recommended for patients with penicillin allergy" in a medical scenario; the constraint verification module verifies whether the generated content conforms to established medical guidelines or insurance terms by querying the knowledge graph, such as confirming the consistency between the statement "Strenuous exercise is prohibited during the acute phase of myocardial infarction" and authoritative medical literature.

[0052] Traditional question-and-answer data generation methods rely on manually written templates or simple content rewriting, resulting in a lack of systematic logical correlation verification of the generated content. For example, when directly using large language models to freely generate question-and-answer pairs in existing technologies, errors in knowledge point combination or statements that violate domain constraints may occur, such as incorrectly associating "insulin treatment for diabetes" with "contraindications for medication in patients with renal failure." This application's embodiment, however, achieves both accuracy and compliance control of professional knowledge by using structured path element extraction and a dual verification mechanism. This application solves the problem of broken logical correlations between topics and knowledge points when generating question-and-answer data. Path element extraction ensures the topological correlation between the generated content and the knowledge system; it overcomes the deficiency of insufficient accuracy in generated content by eliminating knowledge errors and logical contradictions through a dual verification mechanism; and it avoids the risk of generated content violating professional domain constraints by ensuring data compliance with industry standards through knowledge graph verification. In the healthcare field, this solution can accurately generate question-and-answer data that conforms to clinical guidelines, such as "target blood pressure control values ​​for patients with stage III chronic kidney disease." In the financial insurance field, it can generate professional content that conforms to the terms and conditions, such as "calculation of deductible for auto insurance and relationship between accident liability ratio."

[0053] Please see Figure 8 , Figure 8 A specific implementation method following step S5 is shown below: S5A: Calculate the quality score of each question-and-answer data sample in the question-and-answer data sample set. S5B: Filter the question-and-answer data sample set according to the quality score to obtain a filtered question-and-answer data sample set. S5C: Calculate the semantic similarity between the samples in the filtered question-and-answer data sample set and the samples in the knowledge base. S5D: Deduplicate the filtered question-and-answer data sample set based on the semantic similarity to determine a new question-and-answer data sample set. S5E: Update the samples in the knowledge base based on the new question-and-answer data sample set.

[0054] The quality score refers to a numerical evaluation of the accuracy, logical coherence, and professional relevance of question-and-answer samples using preset indicators. This can be achieved through a combination of rule-based matching and model prediction. For example, the BERT model can be used to score the semantic reasonableness of answers, combined with manually annotated keyword coverage to calculate a comprehensive score. Semantic similarity measures the semantic closeness of two text segments using a vector space model. This can be achieved by generating sentence vectors using a pre-trained language model and then calculating cosine similarity. For example, the Sentence-BERT model can be used to map question-and-answer pairs to 768-dimensional vectors before similarity matching. Knowledge base updates refer to the process of incrementally merging new data into the storage system based on semantic deduplication results. This can be achieved using version control and transaction processing mechanisms, such as setting version snapshots to ensure the atomicity of data updates and building inverted indexes to accelerate retrieval efficiency.

[0055] Specifically, the question-and-answer data samples first undergo a quality score calculation module, which sets three dimensions: grammatical correctness, knowledge accuracy, and logical consistency. Each dimension is assigned a weight coefficient, and the results are then summed using a weighted average. During the filtering stage, a dynamic threshold is set based on the application scenario; for example, 0.7 is used as the cutoff value in the medical field, retaining only samples with a value higher than this. In semantic similarity calculation, existing data in the knowledge base and the filtered samples are uniformly encoded into vectors, and an approximate nearest neighbor index structure is established to accelerate the matching process. The deduplication operation employs a dual verification mechanism; for example, data with a similarity exceeding 0.85 is considered duplicate data, but the version with the higher quality score is retained. The knowledge base update uses a rolling window strategy; for example, version iteration is triggered each time the amount of new data reaches a set threshold, while historical versions are retained for backtracking queries.

[0056] Traditional methods rely on manual review for quality screening, which is inefficient and lacks standardized criteria. For example, in the insurance field, it may take three experts two days to review 500 data entries. Existing semantic deduplication methods often use keyword matching, which cannot distinguish between synonyms such as "waiting period" and "observation period," leading to the accidental deletion of valid data. Knowledge base updates often use a full-coverage approach, which carries the risk of losing historical valid data. This application's embodiments improve data processing efficiency through automated quality grading and dynamic threshold adjustment; accurately identify substantially duplicate content by combining deep semantic representation similarity calculation; and reduce storage overhead while ensuring data integrity through an incremental update mechanism. This application effectively solves the problem of model training bias caused by unstable generated data quality.

[0057] In one data instance, question-answer pairs are automatically extracted from a small number of high-quality seed financial insurance Q&As or documents to obtain a seed dataset. Each seed dataset contains a question and an answer. High-level semantic extraction is performed on the seed dataset using GPT-5 to obtain a topic set. With knowledge point set The theme represents the macro-level business direction, such as "health insurance," "auto insurance," "annuity insurance," and "claims policy." The knowledge point represents the micro-level terms or formulas, such as "waiting period calculation," "deductible rules," and "cash value surrender."

[0058] Then, a mapping relationship is constructed between the topics in the topic set and the knowledge points in the knowledge point set. This mapping relationship is as follows: This refers to the probability of a knowledge point appearing under certain conditions within a topic. A concept map is constructed; this concept map is an undirected weighted graph. In this concept graph, the node set V = T∪K, and the edge set E includes topic-topic edges, topic-knowledge point edges, and knowledge point-knowledge point edges. Topic-topic edges represent business relationships (e.g., "critical illness insurance" and "medical insurance" share common claims logic); topic-knowledge point edges represent knowledge attribution; and knowledge point-knowledge point edges represent co-occurrence logic or dependency relationships. W refers to the set of weights for all edges in the concept graph. The weights of edges between nodes in the concept graph are calculated using the following formula: ; In this diagram, PMI represents point mutual information, reflecting co-occurrence strength; Sim represents embedding similarity; and α and β are balance coefficients. This weighted concept graph reflects the structured relationships between financial knowledge points, providing a semantic space for the subsequent generation of diverse data.

[0059] A random walk is performed on the weighted concept graph to sample diverse paths. Initial nodes are randomly selected from the topic set T, and the walk strategy is as follows: Let the step size be L, and each combined path P=[v1,v2,...,vL] represent a potential topic-knowledge point combination. Through multiple samplings, a large number of combined paths {P1,P2,...,PM} can be obtained.

[0060] Input samples from the question-and-answer data sample set into the GPT-5 generation template, for example: "Based on topic {t_i} and knowledge point {k_j}, generate a reasoning-based question-and-answer sample involving insurance claim judgment or risk assessment, requiring clear calculation or logical steps." Model output: Each sample includes a reasoning process, for example: Question: If the insured is hospitalized due to an accident during the waiting period, can they claim compensation? Answer: As this is an accidental event within the waiting period, compensation is available; however, it will not be paid out for illness. When the question-and-answer data sample set is generated, automatic verification and quality control of the samples within the set are required. This application performs two levels of verification on the samples: Consistency verification: Detect the internal logical consistency of questions and answers, through GPT-5 self-test or symbol rule comparison; Constraint verification: Determine the legality of the content based on the legal knowledge base (e.g., determine the authenticity of insurance clauses).

[0061] Finally, the validated samples are filtered using the following formula: in, It is the logical consistency score (0-1). It is the legality score (0-1). , These are weighting coefficients (e.g., all set to 0.5). Set a threshold δ (e.g., 0.7) to retain only... Samples with a value greater than δ are used to form a filtered question-and-answer data sample set. Finally, the samples in the knowledge base are updated using the filtered question-and-answer data sample set.

[0062] In this embodiment, question-and-answer pairs are extracted from seed data in a preset domain to construct a seed dataset. A topic set and a knowledge point set are extracted from the seed dataset, and a mapping relationship is constructed between topics in the topic set and knowledge points in the knowledge point set. A concept graph is constructed based on the topic set, the knowledge point set, and the mapping relationship, and the weights of edges between nodes in the concept graph are calculated to obtain a weighted concept graph. Semantic random walks and combined sampling are performed on the weighted concept graph to generate a set of combined paths. A question-and-answer data sample set is generated based on the combined path set using a large language model, and the question-and-answer data sample set is validated to generate a target question-and-answer data sample set. This embodiment of the invention, by constructing a concept graph model of domain knowledge and combining semantic random walks and large language model generation techniques, systematically solves the problems of incomplete coverage and logical confusion in professional domain question-and-answer data. It has the advantages of improving data generation quality and ensuring knowledge relevance, and is conducive to improving the accuracy of question-and-answer data generation.

[0063] This application's embodiments effectively improve the scenario coverage of professional domain question-and-answer data, enabling the generation of complex cases involving the correlation between gene testing results and exclusion clauses in critical illness insurance within the insurance field. It enhances the logical coherence between knowledge points, ensuring that symptom descriptions and medication recommendations in generated medical Q&As conform to clinical pathway standards. A dual verification mechanism eliminates factual errors, automatically excluding liability determination schemes that violate insurance law when generating health insurance claim cases. It automates the data generation process, rapidly generating treatment plan Q&A data compliant with new standards by adjusting concept graph nodes when medical guidelines are updated.

[0064] Please refer to Figure 9 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of a concept map-based question-and-answer data generation device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0065] like Figure 9 As shown, the concept graph-based question-and-answer data generation device of this embodiment includes: a seed dataset extraction module 61, a mapping relationship construction module 62, a concept graph construction module 63, a combined path set generation module 64, and a question-and-answer data sample generation module 65, wherein: The seed dataset extraction module 61 is used to extract question-answer pair data from seed data in a preset domain to construct a seed dataset; The mapping relationship construction module 62 is used to extract a topic set and a knowledge point set from the seed dataset, and construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set; The concept graph construction module 63 is used to construct a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and to calculate the weight of the edges between nodes in the concept graph to obtain a weighted concept graph. The combined path set generation module 64 is used to perform semantic random walks and combined sampling on the weighted concept graph to generate a combined path set; The question-and-answer data sample generation module 65 is used to generate a question-and-answer data sample set based on the combined path set using a large language model, and to verify the question-and-answer data sample set to generate a target question-and-answer data sample set.

[0066] Furthermore, the mapping relationship construction module 62 includes: The data parsing unit is used to input each question-answer pair data in the seed dataset into the large language model, so as to parse the question-answer pair data through prompting engineering and generate the topic set and the knowledge point set; The probability calculation unit is used to calculate the probability of a knowledge point in the knowledge point set appearing under a topic in the topic set, so as to construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set.

[0067] Furthermore, the concept map construction module 63 includes: A node determination unit is used to take all elements in the topic set and the knowledge point set as nodes to jointly form the node set of the concept graph; A connection relationship construction unit is used to construct connection relationships between topics in the topic set and knowledge points in the knowledge point set based on the mapping relationship, so as to construct the edges of the concept graph; The weight calculation unit is used to calculate the weights of the edges between nodes in the concept graph to obtain the weighted concept graph.

[0068] Furthermore, the weight calculation unit includes: The co-occurrence data statistics subunit is used to statistically analyze co-occurrence data from the seed dataset and embed the node embedding vectors of the concept graph based on the pre-trained model. A point mutual information calculation subunit is used to calculate the point mutual information of each edge in the concept graph and reflect the co-occurrence intensity based on the co-occurrence data; An embedding similarity calculation subunit is used to calculate the embedding similarity between different node embedding vectors; A weighted concept graph generation subunit is used to calculate the weights of edges between nodes in the concept graph based on the point mutual information, the reflection co-occurrence strength, and the embedding similarity, so as to obtain the weighted concept graph.

[0069] Furthermore, the combined path set generation module 64 includes: The starting point determination unit is used to randomly select a theme node in the weighted concept graph as the starting point for the walk; The walk probability calculation unit is used to calculate the probability of moving from the current node to the next stage, and obtain the walk probability; A random walk unit is used to perform semantic random walks and combined sampling on the weighted concept graph according to the walk start point, the walk probability and the preset step size, to generate the combined path set.

[0070] Furthermore, the question-and-answer data sample generation module 65 includes: The target topic extraction unit is used to extract the topic and knowledge points of each combined path in the combined path set to obtain the target topic and target knowledge points. The content generation unit is used to construct a question-and-answer template based on the target topic and target knowledge points, and to generate content based on the question-and-answer target through the large language model to obtain the question-and-answer data sample set. The sample verification unit is used to perform consistency verification and constraint verification on the question-and-answer data sample set to generate the target question-and-answer data sample set.

[0071] Furthermore, the question-and-answer data sample generation module 65 also includes: A quality score calculation unit is used to calculate the quality score of each question-and-answer data sample in the question-and-answer data sample set. A sample filtering unit is used to filter the question-and-answer data sample set according to the quality score to obtain a filtered question-and-answer data sample set. A semantic similarity calculation unit is used to calculate the semantic similarity between samples in the filtered question-and-answer data sample set and samples in the knowledge base; The sample deduplication unit is used to deduplicat the filtered question-and-answer data sample set based on the semantic similarity to determine the new question-and-answer data sample set; The sample update unit is used to update the samples in the knowledge base based on the newly added question and answer data sample set.

[0072] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 10 , Figure 10 This is a basic structural block diagram of the computer device in this embodiment.

[0073] Computer device 7 includes a memory 71, a processor 72, and a network interface 73 that are interconnected via a system bus. It should be noted that... Figure 10Only a computer device 7 with three components—memory 71, processor 72, and network interface 73—is shown. It should be understood that implementing all shown components is not required; more or fewer components may be implemented alternatively. Those skilled in the art will understand that this computer device is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and embedded devices.

[0074] Computer devices can include desktop computers, laptops, handheld computers, and cloud servers. These devices allow for human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0075] The memory 71 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 71 may be an internal storage unit of the computer device 7, such as the hard disk or memory of the computer device 7. In other embodiments, the memory 71 may also be an external storage device of the computer device 7, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, flash card, etc., equipped on the computer device 7. Of course, the memory 71 may also include both internal storage units and external storage devices of the computer device 7. In this embodiment, the memory 71 is typically used to store the operating system and various application software installed on the computer device 7, such as the program code of a question-and-answer data generation method based on concept maps. In addition, the memory 71 can also be used to temporarily store various types of data that have been output or will be output.

[0076] In some embodiments, processor 72 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. Processor 72 is typically used to control the overall operation of computer device 7. In this embodiment, processor 72 is used to run program code stored in memory 71 or process data, for example, to run the program code of the concept map-based question-and-answer data generation method described above, to implement various embodiments of the concept map-based question-and-answer data generation method.

[0077] The network interface 73 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 7 and other electronic devices.

[0078] This application also provides another embodiment, namely, providing a computer-readable storage medium storing a computer program that can be executed by at least one processor to cause the at least one processor to perform the steps of the above-described concept map-based question-and-answer data generation method.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of this application.

[0080] Obviously, the embodiments described above are merely some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the scope of this application. This application can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of protection of this application.

Claims

1. A question-and-answer data generation method based on concept maps, characterized in that, include: Question-answer pairs are extracted from seed data in a predefined domain to construct a seed dataset; Extract a topic set and a knowledge point set from the seed dataset, and construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set; A concept graph is constructed based on the topic set, the knowledge point set, and the mapping relationship, and the weights of the edges between nodes in the concept graph are calculated to obtain a weighted concept graph. Perform semantic random walks and combined sampling on the weighted concept graph to generate a set of combined paths; A question-and-answer data sample set is generated based on the combined path set using a large language model, and the question-and-answer data sample set is validated to generate a target question-and-answer data sample set.

2. The question-and-answer data generation method based on concept graphs according to claim 1, characterized in that, The step of extracting a topic set and a knowledge point set from the seed dataset, and constructing a mapping relationship between topics in the topic set and knowledge points in the knowledge point set, includes: Each question-answer pair in the seed dataset is input into the large language model to parse the question-answer pair data through prompting engineering, thereby generating the topic set and the knowledge point set. Calculate the probability of a knowledge point in the knowledge point set appearing under a topic in the topic set, so as to construct a mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set.

3. The question-and-answer data generation method based on concept graphs according to claim 1, characterized in that, The process of constructing a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and calculating the weights of the edges between nodes in the concept graph to obtain a weighted concept graph, includes: All elements in the topic set and the knowledge point set are used as nodes to form the node set of the concept graph; Based on the mapping relationship, the connection relationship between the topics in the topic set and the knowledge points in the knowledge point set is constructed to build the edges of the concept graph; The weights of the edges between nodes in the concept graph are calculated to obtain the weighted concept graph.

4. The question-and-answer data generation method based on concept graphs according to claim 3, characterized in that, The calculation of the weights of the edges between nodes in the concept graph to obtain the weighted concept graph includes: Co-occurrence data are statistically analyzed from the seed dataset, and the node embedding vectors are input to the concept graph using a pre-trained model. Based on the co-occurrence data, the point mutual information and the co-occurrence intensity of each edge in the concept graph are calculated; Calculate the embedding similarity between the different node embedding vectors; The weighted concept graph is obtained by calculating the weights of the edges between nodes in the concept graph based on the point mutual information, the reflection co-occurrence strength, and the embedding similarity.

5. The question-and-answer data generation method based on concept graphs according to claim 1, characterized in that, The step of performing semantic random walks and combined sampling on the weighted concept graph to generate a set of combined paths includes: Randomly select a theme node in the weighted concept graph as the starting point for the walk; Calculate the probability of moving from the current node to the next stage to obtain the walk probability; Based on the starting point, the probability of walking, and the preset step size, a semantic random walk and combined sampling are performed on the weighted concept graph to generate the set of combined paths.

6. The question-answering data generation method based on concept graphs according to any one of claims 1 to 5, characterized in that, The process of generating a question-and-answer data sample set based on the combined path set using a large language model, validating the question-and-answer data sample set, and generating a target question-and-answer data sample set includes: Extract the topic and knowledge points of each combined path from the set of combined paths to obtain the target topic and target knowledge points; Based on the target topic and target knowledge points, a question-and-answer template is constructed, and content is generated based on the question-and-answer target using the large language model to obtain the question-and-answer data sample set. Consistency and constraint verifications are performed on the question-and-answer data sample set to generate the target question-and-answer data sample set.

7. The question-answering data generation method based on concept graphs according to any one of claims 1 to 5, characterized in that, After generating a question-and-answer data sample set based on the combined path set using a large language model, and validating the question-and-answer data sample set to generate a target question-and-answer data sample set, the method further includes: Calculate the quality score of each question-and-answer data sample in the question-and-answer data sample set; The question-and-answer data sample set is filtered according to the quality score to obtain the filtered question-and-answer data sample set. Calculate the semantic similarity between samples in the filtered question-and-answer data sample set and samples in the knowledge base; Based on the semantic similarity, the filtered question-and-answer data sample set is deduplicated to determine the new question-and-answer data sample set. The samples in the knowledge base are updated based on the newly added question-and-answer data sample set.

8. A question-and-answer data generation device based on concept maps, characterized in that, include: The seed dataset extraction module is used to extract question-answer pair data from seed data in a preset domain to construct a seed dataset. The mapping relationship construction module is used to extract the topic set and the knowledge point set from the seed dataset, and construct the mapping relationship between the topics in the topic set and the knowledge points in the knowledge point set; The concept graph construction module is used to construct a concept graph based on the topic set, the knowledge point set, and the mapping relationship, and to calculate the weight of the edges between nodes in the concept graph to obtain a weighted concept graph; A combined path set generation module is used to perform semantic random walks and combined sampling on the weighted concept graph to generate a combined path set; The question-and-answer data sample generation module is used to generate a question-and-answer data sample set based on the combined path set using a large language model, and to verify the question-and-answer data sample set to generate a target question-and-answer data sample set.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the question-and-answer data generation method based on the concept map as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the question-and-answer data generation method based on a concept graph as described in any one of claims 1 to 7.

Images