A method and system for generating question-and-answer pairs based on large language models and data augmentation

Abstract

The present application provides a method and system for generating question-answer pairs based on a large language model and data enhancement, wherein the text in the target field is processed and decomposed into a set of semantic units, and then candidate questions are iteratively generated through a multi-intent question-answer generation framework; semantic focus drift detection is performed, and questions are adjusted based on two-way matching of context fingerprints to form a core question set; preliminary answers are generated based on this set and semantic divergence indicators are extracted using cross-modal association enhancement logic; a semantic focus track archive is established to map questions with enhanced answer sets and select the best question-answer pair combination; a question-answer sequence containing a context traceability chain is output, integrating semantic tracks, divergence, and cross-modal strength factors. The present application improves the efficiency of identifying and generating high-quality question-answer pairs from complex texts.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical fields of large language models and data enhancement, and in particular to a method and system for generating question-answer pairs based on large language models and data enhancement. Background Art

[0002] In modern knowledge management and intelligent question-answering systems, efficient and accurate question-answering based on large amounts of textual data in specific fields is crucial. This demand is particularly evident in highly specialized and information-intensive fields such as healthcare, legal consulting, and scientific research. These fields require not only systems that can accurately understand complex semantics but also cross-modal data processing capabilities to support knowledge presentation using both text and images, while ensuring the logic and completeness of answers.

[0003] Current technical solutions typically use a large language model based on deep learning combined with data augmentation technology to achieve the automatic generation of question-answer pairs. This type of method first forms a preliminary set of semantic units by decomposing the original text into multi-granular semantic units. Then, using the pre-trained language model framework, a set of candidate questions is generated through multiple iterations of the intent recognition process. In addition, to improve the quality and pertinence of the questions, a probabilistic masking strategy is introduced to dynamically adjust the direction of question generation. For the generated question set, the existing solution also uses a semantic focus drift detection and adaptive reconstruction mechanism to ensure a bidirectional match between the questions and the context fingerprint, thereby screening out the core question set. Ultimately, these questions will be used to generate preliminary answers, and the logical constraints of the answers will be enhanced through cross-modal associations to form the final answer set.

[0004] Although the above methods have improved the performance of question-answering systems to a certain extent, there are still some shortcomings. First, in the semantic unit decomposition stage, although basic semantic information can be captured, the understanding of deep semantic structure is still insufficient, resulting in semantic deviations in the generated questions and answers. Secondly, although the probabilistic masking strategy in the existing scheme helps to adjust the direction of question generation, its flexibility is limited and it is difficult to cope with complex and changeable practical application scenarios. Finally, with regard to the application of cross-modal associations, existing methods focus more on the integration of information at the text level, while there is less exploration of the effective integration of non-text information such as images and audio, which limits the system's expressiveness and accuracy in multimedia environments. Summary of the Invention

[0005] The embodiments of the present application provide a method and system for generating question-answer pairs based on a large language model and data enhancement, so as to solve the problem of low efficiency in identifying and generating high-quality question-answer pairs from complex texts in the prior art.

[0006] In a first aspect, embodiments of the present application provide a method for generating question-answer pairs based on a large language model and data enhancement, comprising:

[0007] Receiving original text input of a target knowledge domain, performing multi-granularity semantic unit decomposition on the original text input, generating a set of semantic units with context fingerprints, wherein each semantic unit in the set of semantic units is bound to an associated adjacent context association tag;

[0008] Input the semantic unit set into a preset multi-intent question-answering generation framework, generate a candidate question set through multiple rounds of intent iteration, and trigger a probability mask strategy to dynamically adjust the question generation direction in each iteration;

[0009] Perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization according to a focus shift threshold, and generate a core question set;

[0010] Generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set and extract a semantic divergence indicator;

[0011] Establishing a semantic focus trajectory archive based on the contextual association tags, performing multi-dimensional topological mapping between the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength;

[0012] Output is a sequence of question-answer pairs containing a context fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index and the strength factor of the cross-modal association to generate a multi-level knowledge topology identification.

[0013] Optionally, the step of performing semantic focus drift detection on the candidate question set, performing bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, and dynamically triggering question reorganization according to a focus displacement threshold to generate a core question set includes:

[0014] Constructing a context fingerprint trajectory file for each semantic unit in the semantic unit set, and converting the context association mark into a multi-dimensional semantic focus vector through vector space projection;

[0015] Performing reverse semantic trajectory analysis on each question in the candidate question set, extracting the latent semantic focus vector of the question, and performing bidirectional dynamic alignment between the latent semantic focus vector and the multidimensional semantic focus vector, and calculating a trajectory offset index between each question semantic unit pair;

[0016] Based on a preset displacement threshold interval, the trajectory offset indicator is divided into a repairable offset interval and an unrepairable offset interval. An adaptive reorganization rule is triggered for problems in the repairable offset interval, and alternative candidate problems are generated through semantic focus vector interpolation technology. An elimination marking operation is performed on the problems in the unrepairable offset interval, and the alternative candidate problems are merged with the non-eliminated problems in the candidate problem set to generate a reorganized candidate problem set.

[0017] The trajectory offset index is recalculated based on the reorganized candidate question set. When the trajectory offset index converges to a preset stability threshold range, a core question set is generated. Otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed.

[0018] Optionally, the trajectory offset indicator is divided into a repairable offset interval and an unrepairable offset interval according to a preset displacement threshold interval, an adaptive reorganization rule is triggered for problems in the repairable offset interval, alternative candidate problems are generated through semantic focus vector interpolation technology, an elimination marking operation is performed on the problems in the unrepairable offset interval, and the alternative candidate problems are merged with the non-eliminated problems in the candidate problem set to generate a reorganized candidate problem set, including:

[0019] Based on the context fingerprint trajectory archive, a historical semantic focus vector library based on the context association mark is constructed, and a sliding window aggregation analysis is performed on the context association mark to generate a spatiotemporal distribution map of the historical semantic focus vector;

[0020] Calculating the normalized weight value of the trajectory offset indicator for each problem in the candidate problem set, and combining it with the evolution trend parameter of the corresponding semantic unit in the spatiotemporal distribution map to determine whether the problem in the candidate problem set falls into the repairable offset interval;

[0021] For the problem within the repairable offset interval, extract a set of adjacent semantic focus vectors from the historical semantic focus vector library, calculate an interpolation weight coefficient based on the evolution trend parameter, perform linear interpolation in the vector space to generate an alternative candidate problem, and attach a reorganization identifier to the alternative candidate problem;

[0022] The alternative candidate questions are mixed and sorted with the questions that have not been eliminated in the candidate question set, and stability verification is performed on the mixed question set based on the reorganization identifier. If the decrease in the trajectory offset indicator of the alternative candidate question does not reach a preset repair gain threshold, the interpolation weight coefficient is adjusted to generate a reorganized candidate question set.

[0023] Optionally, for the problem within the repairable offset interval, extracting a set of adjacent semantic focus vectors from the historical semantic focus vector library, calculating an interpolation weight coefficient based on the evolution trend parameter, performing linear interpolation in the vector space to generate an alternative candidate problem, and adding a reorganization identifier to the alternative candidate problem includes:

[0024] Based on the spatiotemporal distribution map, an interpolation path topology map of the vectors in the historical semantic focus vector library is constructed, and a set of evolutionary trend parameters corresponding to each semantic unit is extracted through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool;

[0025] Analyzing the position coordinates of the semantic focus vector of the problem within the repairable offset interval in the interpolation path topology graph, calculating the semantic continuity index between adjacent vectors based on the evolution trend parameter, and screening the candidate interpolation vector group that meets the path coherence constraint;

[0026] A dynamic interpolation weight matrix is constructed based on the candidate interpolation vector group, an interpolation scale coefficient is generated through the semantic density distribution and evolution direction parameters between nodes in the interpolation path topology graph, a constrained linear interpolation operation is performed in the vector space to generate alternative candidate questions, and a dynamic reorganization identifier is attached to the alternative candidate questions.

[0027] Optionally, the trajectory offset index is recalculated based on the reorganized candidate question set, and when the trajectory offset index converges to a preset stability threshold range, a core question set is generated; otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed, including:

[0028] Constructing a dynamic convergence judgment model for the trajectory offset indicator, and generating stability baseline parameters and confidence intervals by statistically analyzing the trajectory offset indicators during the historical reorganization process through a sliding window;

[0029] Inputting the reorganized candidate problem set into the dynamic convergence determination model, calculating the convergence difference between the trajectory offset index and the stability baseline parameter, and triggering an iteration termination condition if the convergence difference satisfies a preset stability threshold range;

[0030] When the convergence difference exceeds a preset stability threshold range, adjusting the weight coefficient of the semantic focus vector interpolation according to the gradient direction of the convergence difference, and generating an incremental candidate question set based on the unused semantic focus vectors in the context fingerprint trajectory archive;

[0031] performing cross-round trajectory offset verification on the incremental candidate problem set, extracting trajectory offset fluctuation parameters generated during the verification process, and feeding the trajectory offset fluctuation parameters back to the dynamic convergence determination model to update the stability baseline parameters;

[0032] If the consecutive iterations reach the preset maximum reorganization depth threshold, the iteration termination condition is forcibly triggered, and a core problem set carrying a convergence difference parameter and an incremental reorganization identifier is output.

[0033] Optionally, generating a preliminary answer set based on the core question set, enhancing answer logic constraints through cross-modal association, generating an enhanced answer set, and extracting a semantic divergence index include:

[0034] Constructing a multimodal semantic space mapping network to convert each question in the core question set into a semantic query vector, and retrieving a set of candidate answer fragments from a multimodal data source in the target knowledge domain through a distributed reasoning engine;

[0035] Performing a semantic alignment operation on the candidate answer segment set, calculating a semantic matching index between the semantic query vector and the candidate answer segment set through a cross-modal association matrix, and screening a set of answer segments that meet preliminary logical constraints;

[0036] The logical enhancement weight coefficient of the answer fragment set is generated based on the cross-modal association matrix, the answer fragment set is converted into a preliminary answer set through a multimodal semantic fusion algorithm, and the semantic divergence index of the preliminary answer set is extracted.

[0037] Optionally, the step of establishing a semantic focus trajectory archive based on the contextual association tag, performing multi-dimensional topological mapping between the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength includes:

[0038] Constructing a multidimensional topological structure of a semantic focus trajectory archive, mapping the semantic unit set, the core question set, and the enhanced answer set into a node set in a topological graph through the context association mark, and generating initial connectivity strength weights between nodes based on the semantic divergence index and the logical consistency scoring parameter;

[0039] Performing semantic focus trajectory analysis on the node set, extracting the context fingerprint trajectory code of each node, and generating a semantic correlation matrix between nodes by calculating the similarity of the context fingerprint trajectory codes;

[0040] A dynamic connectivity strength calculation model is constructed based on the semantic relevance matrix and the initial connectivity strength weights, and the connectivity strength weights between nodes are optimized through multiple rounds of iterations to generate a topological connectivity graph in a stable state;

[0041] Performing cluster analysis on the topological connectivity graph to divide the node set into a plurality of semantic focus clusters according to a connectivity strength threshold, wherein each cluster contains a strongly associated combination of question nodes and answer nodes;

[0042] Perform logical consistency verification on the question-answer pair combination within the semantic focus cluster, extract the logical conflict parameters generated during the verification process, feed the logical conflict parameters back to the dynamic connectivity strength calculation model to update the connectivity strength weight, and screen out the target question-answer pair combination, wherein the target question-answer pair combination includes the optimized values of the logical consistency scoring parameters and the semantic divergence index.

[0043] In a second aspect, an embodiment of the present application provides a question-answer pair generation system based on a large language model and data enhancement, comprising:

[0044] A receiving module is configured to receive an original text input of a target knowledge domain, perform multi-granularity semantic unit decomposition on the original text input, and generate a set of semantic units with context fingerprints, wherein each semantic unit in the set of semantic units is bound to an associated adjacent context association tag;

[0045] An input module is used to input the semantic unit set into a preset multi-intent question-answering generation framework, generate a set of candidate questions through multiple rounds of intent iteration, and trigger a probabilistic masking strategy to dynamically adjust the question generation direction at each iteration;

[0046] a matching module configured to perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization based on a focus shift threshold, and generate a core question set;

[0047] A generation module, configured to generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set, and extract a semantic divergence indicator;

[0048] An establishment module is used to establish a semantic focus trajectory archive based on the contextual association mark, perform multi-dimensional topological mapping between the core question set and the enhanced answer set, drive aggregation through the trajectory archive, and screen target question-answer pair combinations based on connectivity strength;

[0049] The output module is used to output a sequence of question-answer pairs containing a context fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index and the strength factor of the cross-modal association to generate a multi-level knowledge topology identification.

[0050] In a third aspect, an embodiment of the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a question-answer pair generation method based on a large language model and data enhancement as described in any one of the first aspects.

[0051] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon, which, when executed by a processor, implements a question-answer pair generation method based on a large language model and data enhancement as described in any one of the first aspects.

[0052] In an embodiment of the present application, an original text input of a target knowledge domain is received, the original text input is decomposed into multi-granularity semantic units to generate a set of semantic units with context fingerprints, each semantic unit in the set of semantic units is bound to an associated adjacent context association tag; the set of semantic units is input into a preset multi-intention question-answering generation framework, a set of candidate questions is generated through multiple rounds of intention iteration, and each iteration triggers a probability mask strategy to dynamically adjust the direction of question generation; semantic focus drift detection is performed on the candidate question set, and a two-way trajectory matching is performed between the question and the context fingerprint of the set of semantic units through an adaptive reconstruction mechanism, and the focus displacement threshold is used to determine the candidate question. The value dynamically triggers question reorganization to generate a core question set; a preliminary answer set is generated based on the core question set, and the answer logic constraints are enhanced through cross-modal association to generate an enhanced answer set and extract the semantic divergence index; a semantic focus trajectory archive is established based on the context association mark, and the core question set and the enhanced answer set are multi-dimensionally topologically mapped, and the aggregation is driven by the trajectory archive, and the target question and answer pair combination is screened according to the connectivity strength; a question and answer pair sequence containing a context fingerprint traceability chain is output, wherein each question and answer pair integrates the semantic focus trajectory archive, the semantic divergence index and the strength factor of the cross-modal association to generate a multi-level knowledge topology identification.

[0053] The technical solution of this application has the following beneficial effects:

[0054] This application not only improves the efficiency of identifying and generating high-quality question-answer pairs from complex texts, but also significantly enhances the accuracy and depth of answers through technical means such as semantic analysis, adaptive reconstruction, and cross-modal association. At the same time, through contextual association tags and semantic focus track archives, a multi-level topological identification of knowledge is achieved, which helps to better understand and apply information and greatly optimizes the knowledge management and retrieval process.

[0055] Furthermore, the embodiment of the present application also performs a process of semantic focus drift detection on the candidate question set, by constructing a context fingerprint trajectory file for each semantic unit, and using vector space projection technology to convert the context association tag into a multidimensional semantic focus vector. Then, a reverse semantic trajectory analysis is performed on each question in the candidate question set, its potential semantic focus vector is extracted, and it is dynamically aligned with the multidimensional semantic focus vector in both directions to calculate the trajectory offset index between the pairs of question semantic units. According to the preset displacement threshold interval, the repairable and unrepairable offset intervals are distinguished, and the adaptive reorganization rules are applied to the former and alternative candidate questions are generated by semantic focus vector interpolation, while the latter are eliminated. Finally, the trajectory offset is re-evaluated based on the reorganized candidate question set until it converges within the stability threshold range, and the core question set is output.

[0056] This method significantly improves the quality of candidate questions through precise semantic focus drift detection and an adaptive reconstruction mechanism, ensuring that the generated set of core questions more closely matches the semantic structure of the original text. The application of semantic focus vector interpolation technology not only effectively fixes some of the offset issues, but also improves its accuracy and pertinence while maintaining the intent of the questions. Furthermore, the iterative execution of semantic focus vector interpolation and trajectory offset verification further enhances the robustness and stability of the system, making the final output set of core questions more reflective of the user's real needs and the deep structure of the knowledge domain. This approach significantly improves the question-answering system's ability to handle complex text and enhances the user experience.

[0057] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0059] Figure 1 A flowchart of a method for generating question-answer pairs based on a large language model and data enhancement provided in an embodiment of the present application;

[0060] Figure 2 A schematic diagram of the structure of a question-answer pair generation system based on a large language model and data enhancement provided in an embodiment of the present application;

[0061] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to enable people skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0063] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0064] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0065] Figure 1 A flowchart of a method for generating question-answer pairs based on a large language model and data enhancement is provided for the embodiment of this application. Figure 1 As shown, the method includes:

[0066] Step 101: receiving original text input in a target knowledge domain, performing multi-granularity semantic unit decomposition on the original text input, and generating a semantic unit set with context fingerprints;

[0067] In this step, multi-granularity semantic unit decomposition is a process that parses the raw text input into semantic units at different levels (such as sentences, phrases, or words). Each semantic unit is then assigned a context fingerprint and adjacent context tags. The context fingerprint helps determine the unique position of each semantic unit within its surroundings, while the context tags identify adjacent semantic units. This approach helps to more accurately understand the text content and its context.

[0068] In practice, the system first receives raw text input from the target knowledge domain. It then uses natural language processing techniques to perform multi-level parsing of the text, generating a collection of semantic units with contextual information. During this process, each semantic unit is analyzed in relation to its surrounding context and associated tags are added to facilitate subsequent understanding and processing of this information.

[0069] For example, in an intelligent customer service scenario, suppose a user asks for specific details about the return policy for an electronic product. After receiving the query, the system first decomposes the text into multi-granular semantic units. For example, "I want to know about the return policy for this electronic product" is decomposed into multiple semantic units: "I want to know," "this electronic product," and "return policy." The system adds contextual fingerprints and association tags to each semantic unit. For example, "I want to know" may be bound to the relevant context of the user's demand expression, while "return policy" is associated with specific terms and conditions. This step ensures the accuracy of subsequent question generation and answer extraction.

[0070] Step 102: Input the semantic unit set into a preset multi-intent question-answering generation framework, generate a candidate question set through multiple rounds of intent iteration, and trigger a probability mask strategy to dynamically adjust the question generation direction in each iteration;

[0071] In this step, the multi-intent question-answering generation framework, a deep learning-based system, generates a series of possible questions based on a set of semantic units. With each iteration, a probabilistic masking strategy dynamically adjusts the direction of question generation to increase the diversity and specificity of questions, ensuring that the generated questions are both comprehensive and relevant to actual needs.

[0072] In practice, the semantic unit set generated in step 101 is fed into a pre-configured question-answering framework, and a set of candidate questions is generated through multiple iterations. During this process, a probabilistic masking strategy is employed to increase the flexibility and specificity of question generation, ensuring that the generated questions are both diverse and highly practical.

[0073] For example, in the aforementioned intelligent customer service scenario, the system uses a multi-intent question-answering framework to generate multiple relevant questions for the query "I want to know the return policy for this electronic product," such as "What is the return period for this electronic product?", "What documentation is required for a return?", and "Can I return a damaged product?" These questions cover different aspects that users may be interested in, providing comprehensive information support.

[0074] Step 103: Perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization according to a focus shift threshold, and generate a core question set;

[0075] In this step, semantic focus drift detection aims to assess the degree of match between candidate questions and the original semantic units. It uses an adaptive reconstruction mechanism to achieve bidirectional trajectory matching and decides whether to restructure the questions based on a displacement threshold to generate a more accurate set of core questions. This method can effectively reduce the problem of question deviation caused by semantic drift.

[0076] In practice, semantic focus drift detection is performed on candidate question sets, calculating the trajectory offset between each question and the set of semantic units. Based on a preset displacement threshold, question reorganization rules are triggered to generate a core set of questions. This process ensures the accuracy and relevance of the questions.

[0077] For example, continuing with the above example, the system detected that certain questions, such as "How do I contact after-sales service?" while related to returns, deviated from the primary focus of "return policy." Therefore, the system restructured these questions to generate a core set of questions more focused on the return process, such as "How long is the return policy valid?" and "What documents are required for a return?" These questions directly addressed the user's initial query, improving the user experience.

[0078] Step 104: Generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set, and extract a semantic divergence index;

[0079] In this step, cross-modal association strengthens answer logic constraints by combining information from multiple data formats, such as text and images, to optimize the quality of preliminary answers and extract a semantic divergence metric to measure answer diversity. This approach improves the completeness and logic of answers by integrating multiple information sources.

[0080] In practice, we generate a preliminary set of answers based on a core set of questions. We introduce cross-modal connections (such as charts and videos) to enhance the logic and completeness of the answers, while also extracting a semantic divergence metric to assess the diversity of the answers. This step ensures that the answers are both comprehensive and easy to understand.

[0081] For example, in response to the core question, "What documents are required for a return?", the system not only provides a detailed list of documents but also includes an image of a sample form, demonstrating the correct format. Furthermore, the system calculates a semantic divergence index to ensure that the information provided is comprehensive and non-redundant, avoiding information overload. This multimodal response allows users to quickly obtain the information they need and take the correct action.

[0082] Step 105: Establishing a semantic focus trajectory archive based on the contextual association tags, performing multi-dimensional topological mapping on the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength;

[0083] In this step, based on the semantic focus trajectory archive created with contextual tags, a multidimensional topological mapping is performed between the core question and the enhanced answer set. The trajectory archive is then used to drive aggregation and select the optimal question-answer combination. This approach, through contextual tags and trajectory archives, enhances the coherence and consistency of question-answer pairs.

[0084] In practice, we build a semantic focus trajectory profile based on contextual tags, match the core question set with the enhanced answer set, and select the optimal question-answer pair combination based on connectivity strength. This process ensures the best possible match between question and answer pairs and improves the overall performance of the system.

[0085] For example, in the intelligent customer service system, the system uses contextual tags to match the question "What documents are required for a return?" with a detailed list of documents and images of sample forms, generating high-quality question-answer pairs. For example, the system displays a text description of the required return documents, along with a clear sample form to guide users through filling them out. This approach not only improves information readability but also strengthens user trust.

[0086] Step 106: Output a sequence of question-answer pairs containing a context fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier.

[0087] This step outputs a sequence of question-answer pairs containing a contextual fingerprint traceability chain. Each question-answer pair integrates a semantic focus trajectory profile, semantic divergence indicators, and cross-modal association strength factors to form a multi-level knowledge topology. This method enhances the user experience by integrating rich contextual information and data types.

[0088] In practice, the final output is a sequence of question-answer pairs supported by a contextual fingerprint traceability chain. These question-answer pairs not only contain rich background information but also enhance user understanding and experience by integrating multiple data types. This process ensures high usability and user satisfaction of the question-answer pairs.

[0089] For example, the intelligent customer service system provides users with a complete question-and-answer sequence, covering all steps from asking a question to receiving a detailed answer, along with diagrams and instructions for the return process. For example, in response to the question, "What documents are required for a return?" the system not only provides a detailed document list and sample forms, but also includes a video tutorial on the return process, helping users more intuitively understand the entire process. This multi-layered knowledge presentation significantly improves user satisfaction and problem-solving efficiency.

[0090] Through the six-step approach described above, the intelligent customer service system is able to efficiently and accurately extract and generate high-quality question-answer pairs from complex text. This approach not only improves the relevance of questions and the accuracy of answers, but also enhances user understanding and experience through the integration of cross-modal data. Furthermore, the application of semantic focus trajectory archives and contextual fingerprint traceability chains further enhances the robustness and reliability of the system, enabling intelligent customer service to better serve diverse and complex user needs, significantly improving user satisfaction and system effectiveness.

[0091] To further improve the accuracy and flexibility of semantic focus drift detection, in some embodiments, step 103 performs semantic focus drift detection on the candidate question set, performs bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, and dynamically triggers question reorganization according to a focus shift threshold to generate a core question set, including:

[0092] A context fingerprint trajectory file is constructed for each semantic unit in the semantic unit set, and the context association mark is converted into a multidimensional semantic focus vector through vector space projection; a reverse semantic trajectory analysis is performed on each question in the candidate question set to extract the potential semantic focus vector of the question, and the potential semantic focus vector is dynamically aligned with the multidimensional semantic focus vector in both directions to calculate the trajectory offset index between each pair of question semantic units; according to a preset displacement threshold interval, the trajectory offset index is divided into a repairable offset interval and an unrepairable offset interval, and an adaptive reorganization rule is triggered for the questions in the repairable offset interval, and alternative candidate questions are generated through the semantic focus vector interpolation technology, and an elimination marking operation is performed on the questions in the unrepairable offset interval, and the alternative candidate questions are merged with the non-eliminated questions in the candidate question set to generate a reorganized candidate question set; the trajectory offset index is recalculated based on the reorganized candidate question set, and when the trajectory offset index converges to a preset stability threshold range, a core question set is generated, otherwise semantic focus vector interpolation and trajectory offset verification are iteratively performed.

[0093] In this embodiment, the context fingerprint trajectory archive refers to constructing a time series archive containing the context information of each semantic unit. Through vector space projection, the system converts these context-related tags into multi-dimensional semantic focus vectors to facilitate subsequent comparison and analysis. This method enables the system to more accurately capture the position of each semantic unit in its context and its relationship with other units. Reverse semantic trajectory analysis is used to start from the problem, trace back its potential semantic focus vector, and compare it with the multi-dimensional semantic focus vector of the original semantic unit to evaluate its degree of deviation.

[0094] In an embodiment of the present application, first, the system constructs a context fingerprint trajectory file for each semantic unit and converts it into a multidimensional semantic focus vector. Then, a reverse semantic trajectory analysis is performed on each question in the candidate question set, its potential semantic focus vector is extracted, and it is dynamically aligned with the multidimensional semantic focus vector in both directions to calculate the trajectory offset index. Next, according to the preset displacement threshold interval, the repairable and unrepairable offset intervals are distinguished. For problems within the repairable offset interval, the semantic focus vector interpolation technology is used to generate alternative candidate questions; for problems within the unrepairable offset interval, they are directly eliminated. Finally, the trajectory offset index is recalculated based on the reorganized candidate question set until it converges within the preset stability threshold range, and the core question set is output.

[0095] The following is a specific embodiment:

[0096] In an intelligent customer service system, a user inquires about the warranty policy for a particular smartphone. The system first decomposes the query text into multi-granular semantic units, identifying key phrases such as "smartphone" and "warranty policy." Next, the system constructs a contextual fingerprint trajectory profile for each semantic unit and converts the contextual association tags into a multidimensional semantic focus vector through vector space projection. For example, "warranty policy" might be tagged as being associated with "product model" and "purchase date."

[0097] The system then generates multiple candidate questions, such as: "How long is the warranty period for this smartphone?", "What supporting documents are required during the warranty period?", etc. By performing reverse semantic trajectory analysis on these questions, the system extracts the latent semantic focus vector of each question, and performs bidirectional dynamic alignment with the multi-dimensional semantic focus vector of the original semantic unit to calculate the trajectory offset index. Assuming that the system finds that some questions (such as "How do I contact after-sales service?") have a large trajectory offset and belong to the unrepairable offset interval, they are marked for elimination. For other questions within the repairable offset interval, the system uses semantic focus vector interpolation technology to generate alternative candidate questions, such as: "Can the battery be replaced during the warranty period?"

[0098] Finally, the system recalculates the trajectory offset index based on the reorganized set of candidate questions. After multiple iterations, the system adjusts until the trajectory offset index for all questions converges within the preset stability threshold, outputting a set of core questions. This allows the intelligent customer service system to not only more accurately understand user query intent but also provide answers that better meet actual needs, significantly improving user experience and service quality.

[0099] To further improve the accuracy and flexibility of semantic focus drift detection, in certain embodiments, the method of dividing the trajectory offset indicator into a repairable offset interval and an unrepairable offset interval according to a preset displacement threshold interval in step 103, triggering an adaptive reorganization rule for problems within the repairable offset interval, generating alternative candidate questions using semantic focus vector interpolation technology, performing an elimination marking operation on problems within the unrepairable offset interval, and merging the alternative candidate questions with the non-eliminated questions in the candidate question set to generate a reorganized candidate question set further includes:

[0100] Based on the context fingerprint trajectory archive, a historical semantic focus vector library based on the context association mark is constructed, and a sliding window aggregation analysis is performed on the context association mark to generate a spatiotemporal distribution map of the historical semantic focus vectors; a normalized weight value of the trajectory offset index is calculated for each question in the candidate question set, and combined with the evolution trend parameters of the corresponding semantic unit in the spatiotemporal distribution map, it is determined whether the questions in the candidate question set fall into a repairable offset interval; for questions within the repairable offset interval, a set of adjacent semantic focus vectors is extracted from the historical semantic focus vector library, an interpolation weight coefficient is calculated according to the evolution trend parameter, linear interpolation is performed in the vector space to generate alternative candidate questions, and a reorganization identifier is attached to the alternative candidate questions; the alternative candidate questions are mixed and sorted with the non-eliminated questions in the candidate question set, and stability verification is performed on the mixed question set based on the reorganization identifier. If the decrease in the trajectory offset index of the alternative candidate question does not reach a preset repair gain threshold, the interpolation weight coefficient is adjusted to generate a reorganized candidate question set.

[0101] In this embodiment, a historical semantic focus vector library is constructed based on contextual fingerprint trajectory archives. It contains the semantic focus vectors of each semantic unit at different time points. These vectors are aggregated and analyzed through a sliding window to form a spatiotemporal distribution map, showing the trend of semantic units changing over time and context. The normalized weight value is used to measure the degree of deviation between each candidate question and its original semantic unit, while the evolution trend parameter helps predict the possible direction of future changes.

[0102] In an embodiment of the present application, first, the system constructs a historical semantic focus vector library based on the context fingerprint trajectory archive, and generates a spatiotemporal distribution map by performing a sliding window aggregation analysis on the context association tags. Next, the system calculates the normalized weight value of the trajectory offset index of each question in the candidate question set, and combines the evolution trend parameters of the corresponding semantic unit in the spatiotemporal distribution map to determine whether the problem belongs to a repairable offset interval. For repairable problems, the system extracts a set of adjacent semantic focus vectors from the historical semantic focus vector library, calculates the interpolation weight coefficient based on the evolution trend parameters, and performs linear interpolation in the vector space to generate alternative candidate questions. The system attaches a reorganization identifier to these alternative candidate questions and sorts them mixedly with the original candidate questions that have not been eliminated. Finally, the stability of the mixed question set is verified based on the reorganization identifier. If the decrease in the trajectory offset index of the alternative candidate question does not reach the preset repair gain threshold, the interpolation weight coefficient is adjusted to regenerate the reorganized candidate question set.

[0103] The following is a specific embodiment:

[0104] In an intelligent customer service scenario, a user inquires about the specific details of the return policy for an electronic product. The system first decomposes the received query text into multi-granular semantic units, identifying key phrases such as "electronic products" and "return policy." The system then constructs a contextual fingerprint trajectory profile for each semantic unit and generates a library of historical semantic focus vectors. Using sliding window aggregation analysis, the system generates a spatiotemporal distribution map of these semantic units, demonstrating their changing trends over time.

[0105] Next, the system generates multiple candidate questions, such as: "What is the return period for this electronic product?", "What supporting documents are required for return?", etc. The system calculates the normalized weight value of the trajectory offset indicator for each question, and combines the evolutionary trend parameters of the corresponding semantic unit in the spatiotemporal distribution map to evaluate whether the question belongs to the repairable offset interval. Suppose the question "How to contact after-sales service?" is judged to be unrepairable and is directly eliminated. As for the question "What supporting documents are required for return?", since its trajectory offset is small and belongs to the repairable interval, the system extracts the adjacent semantic focus vector set from the historical semantic focus vector library, and calculates the interpolation weight coefficient based on the evolutionary trend parameters to generate an alternative candidate question: "What documents need to be prepared for return?"

[0106] The system attaches reorganization identifiers to these alternative candidate questions and sorts them together with the original candidate questions that were not eliminated. The system then verifies the stability of the mixed question set. If it is found that the trajectory deviation indicator of some alternative candidate questions has not decreased by a preset repair gain threshold (for example, "What documents are required for returns?" still has a slight deviation), the interpolation weight coefficient is adjusted to regenerate a more accurate alternative candidate question. Ultimately, the system outputs a set of optimized core questions, which significantly improves the accuracy of the question-answering system and the user experience.

[0107] To further improve the accuracy and flexibility of semantic focus drift detection, in certain embodiments, step 103 extracts a set of adjacent semantic focus vectors from the historical semantic focus vector library for the question within the repairable offset interval, calculates an interpolation weight coefficient based on the evolution trend parameter, performs linear interpolation in the vector space to generate an alternative candidate question, and adds a reorganization identifier to the alternative candidate question, further comprising:

[0108] Based on the spatiotemporal distribution map, an interpolation path topology map of the vectors in the historical semantic focus vector library is constructed, and a set of evolutionary trend parameters corresponding to each semantic unit is extracted through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool; the position coordinates of the semantic focus vectors in the interpolation path topology map are analyzed for the problems in the repairable offset interval, and the semantic continuity index between adjacent vectors is calculated according to the evolutionary trend parameters, and the candidate interpolation vector group that meets the path coherence constraint is screened; based on the candidate interpolation vector group, a dynamic interpolation weight matrix is constructed, and an interpolation scale coefficient is generated through the semantic density distribution and evolutionary direction parameters between nodes in the interpolation path topology map, and a constrained linear interpolation operation is performed in the vector space to generate alternative candidate problems, and a dynamic reorganization identifier is attached to the alternative candidate problems.

[0109] In this embodiment, the interpolation path topology map is constructed based on the spatiotemporal distribution map, which shows the relationship between historical semantic focus vectors and their changing trends over time. The evolution trend parameter set of each semantic unit contains its changes in different time periods, which is used to predict future semantic changes. The multi-level interpolation path candidate pool is a series of possible interpolation paths generated based on these evolution trend parameters. The semantic continuity index is used to measure the similarity or coherence between adjacent semantic focus vectors to ensure that the interpolated vector maintains semantic consistency.

[0110] In an embodiment of the present application, first, the system constructs an interpolation path topology map based on the spatiotemporal distribution map, and extracts the set of evolutionary trend parameters corresponding to each semantic unit through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool. Then, for problems within the repairable offset interval, the system analyzes the position coordinates of its semantic focus vector in the interpolation path topology map, and calculates the semantic continuity index between adjacent vectors based on the evolutionary trend parameters, and screens out candidate interpolation vector groups that meet the path coherence constraints. Based on these candidate interpolation vector groups, the system constructs a dynamic interpolation weight matrix, and generates an interpolation scale coefficient through the semantic density distribution and evolutionary direction parameters between nodes in the interpolation path topology map. Finally, the system performs constrained linear interpolation operations in the vector space to generate alternative candidate problems and attaches dynamic reorganization identifiers to them for subsequent verification and optimization.

[0111] The following is a specific embodiment:

[0112] In an intelligent customer service scenario, a user inquires about the warranty policy for a particular smartphone. The system first decomposes the received query text into multi-granular semantic units, identifying key phrases such as "smartphone" and "warranty policy." The system then constructs a contextual fingerprint trajectory profile for each semantic unit and generates a library of historical semantic focus vectors. Using sliding window aggregation analysis, the system generates a spatiotemporal distribution map of these semantic units, demonstrating their changing trends over time and context.

[0113] Next, the system generates multiple candidate questions, such as "What is the warranty period for this smartphone?" and "What documentation is required during the warranty period?" The system calculates the trajectory offset index for each question and, combined with the evolutionary trend parameters of the corresponding semantic unit in the spatiotemporal distribution map, assesses whether the question falls within the repairable offset range. For example, the question "Can the battery be replaced during the warranty period?" is determined to fall within the repairable offset range. The system constructs an interpolation path topology graph based on the spatiotemporal distribution map and, through sliding window aggregation analysis, extracts the set of evolutionary trend parameters corresponding to each semantic unit, generating a multi-level interpolation path candidate pool. Next, the system analyzes the position coordinates of the semantic focus vector for the question "Can the battery be replaced during the warranty period?" within the interpolation path topology graph and calculates the semantic continuity index between adjacent vectors based on the evolutionary trend parameters. After selecting candidate interpolation vector groups that meet path coherence constraints, the system constructs a dynamic interpolation weight matrix based on these candidate interpolation vector groups.

[0114] The system generates interpolation scaling coefficients by interpolating the semantic density distribution and evolution direction parameters between nodes in the path topology graph, and performs constrained linear interpolation operations in the vector space to generate an alternative candidate question: "How to deal with battery failure during the warranty period?" The system attaches a dynamic reorganization identifier to this alternative candidate question to facilitate subsequent verification of its stability.

[0115] The system then mixes and ranks the generated alternative candidate questions with the original candidate questions that survived the elimination phase and performs stability verification. If the trajectory offset indicator for some alternative candidate questions does not decrease by the preset repair gain threshold, the interpolation weight coefficient is adjusted to regenerate a more accurate alternative candidate question. Ultimately, the system outputs a set of optimized core questions, significantly improving the accuracy of the question-answering system and the user experience, allowing users to obtain answers that better suit their needs.

[0116] In order to further improve the stability and convergence efficiency of semantic focus drift detection, in some embodiments, the trajectory offset index is recalculated based on the reorganized candidate question set in step 103. When the trajectory offset index converges to a preset stability threshold range, a core question set is generated. Otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed, further comprising:

[0117] A dynamic convergence determination model for the trajectory offset indicator is constructed, and the trajectory offset indicator in the historical reorganization process is statistically analyzed using a sliding window to generate a stability baseline parameter and a confidence interval. The reorganized candidate question set is input into the dynamic convergence determination model, and the convergence difference between the trajectory offset indicator and the stability baseline parameter is calculated. If the convergence difference satisfies a preset stability threshold range, an iteration termination condition is triggered. When the convergence difference exceeds the preset stability threshold range, the weight coefficient of the semantic focus vector interpolation is adjusted according to the gradient direction of the convergence difference, and an incremental candidate question set is generated based on the unused semantic focus vectors in the context fingerprint trajectory archive. Cross-round trajectory offset verification is performed on the incremental candidate question set, and the trajectory offset fluctuation parameter generated during the verification process is extracted. The trajectory offset fluctuation parameter is fed back to the dynamic convergence determination model to update the stability baseline parameter. If the consecutive iteration rounds reach the preset maximum reorganization depth threshold, the iteration termination condition is forcibly triggered, and a core question set carrying the convergence difference parameter and the incremental reorganization identifier is output.

[0118] In this embodiment, the dynamic convergence determination model is used to assess whether the trajectory offset indicator has reached a stable state. It uses a sliding window technique to statistically analyze the trajectory offset indicator during the historical reconstruction process to generate stability baseline parameters and confidence intervals. These parameters measure the difference between the current trajectory offset indicator and the expected stable state. The convergence difference represents the gap between the current trajectory offset indicator and the stability baseline parameters, while the incremental candidate question set is a new set of questions generated from unused semantic focus vectors in the contextual fingerprint trajectory archive.

[0119] In an embodiment of the present application, first, the system constructs a dynamic convergence judgment model for the trajectory offset indicator, and uses a sliding window to count the trajectory offset indicators in the historical reorganization process to generate stability baseline parameters and confidence intervals. Then, the system inputs the reorganized candidate question set into the dynamic convergence judgment model to calculate the convergence difference between the trajectory offset indicator and the stability baseline parameter. If the convergence difference meets the preset stability threshold range, the iteration termination condition is triggered and the core question set is output. If it exceeds the threshold range, the weight coefficient of the semantic focus vector interpolation is adjusted according to the gradient direction of the convergence difference, and an incremental candidate question set is generated based on the unused semantic focus vectors in the context fingerprint trajectory archive. Then, a cross-round trajectory offset verification is performed on the incremental candidate question set, and the trajectory offset fluctuation parameters generated during the verification process are extracted and fed back to the dynamic convergence judgment model to update the stability baseline parameters. If the continuous iteration rounds reach the maximum reorganization depth threshold, the iteration is forcibly terminated and the final core question set is output.

[0120] The following is a specific embodiment:

[0121] In an intelligent customer service scenario, a user inquires about the return policy for an electronic product. The system first decomposes the query text into multi-granular semantic units, identifying key phrases such as "electronic products" and "return policy." The system then constructs a contextual fingerprint trajectory profile for each semantic unit and generates a library of historical semantic focus vectors. Using sliding window aggregation analysis, the system generates a spatiotemporal distribution map of these semantic units, demonstrating their changing trends over time and context.

[0122] Next, the system generates multiple candidate questions, such as "What is the return period for this electronic product?" and "What documentation is required for a return?" The system calculates the trajectory offset metric for each question and, combined with the evolutionary trend parameters of the corresponding semantic unit in the spatiotemporal distribution map, assesses whether the question falls within the fixable offset range. For example, the question "What documents are required for a return?" is determined to fall within the fixable offset range. The system constructs an interpolation path topology graph based on the spatiotemporal distribution map and extracts the set of evolutionary trend parameters corresponding to each semantic unit through sliding window aggregation analysis, generating a multi-level interpolation path candidate pool. Next, the system analyzes the position coordinates of the semantic focus vector for the question "What documents are required for a return?" within the interpolation path topology graph. Based on the evolutionary trend parameters, it calculates the semantic continuity metric between adjacent vectors, selecting candidate interpolation vector groups that meet the path coherence constraint. Based on these candidate interpolation vector groups, the system constructs a dynamic interpolation weight matrix and generates interpolation scaling coefficients based on the semantic density distribution and evolutionary direction parameters between nodes in the interpolation path topology graph, generating an alternative candidate question: "What documents are required for a return?"

[0123] The system constructs a dynamic convergence determination model for the trajectory deviation metric and uses a sliding window to statistically analyze the trajectory deviation metric during the historical reorganization process to generate stability baseline parameters and confidence intervals. The system then inputs the reorganized candidate problem set into the dynamic convergence determination model and calculates the convergence difference between the trajectory deviation metric and the stability baseline parameters. If the convergence difference meets the preset stability threshold, the iteration termination condition is triggered and the core problem set is output.

[0124] However, if the convergence difference exceeds the preset stability threshold, the system adjusts the weight coefficient of the semantic focus vector interpolation according to the gradient direction of the convergence difference and generates an incremental candidate question set based on the unused semantic focus vectors in the context fingerprint trajectory archive. For example, "What forms do I need to fill out when returning a product?" The system performs cross-round trajectory offset verification on the incremental candidate question set, extracts the trajectory offset fluctuation parameters generated during the verification process, and feeds them back to the dynamic convergence judgment model to update the stability baseline parameters. If the maximum reorganization depth threshold is reached in consecutive iterations, the system forcibly terminates the iteration and outputs the final core question set, which includes the convergence difference parameter and incremental reorganization identifier, to ensure that the final generated question set is more accurate and meets user needs. In this way, the intelligent customer service system can handle complex queries more effectively and improve user experience and service quality.

[0125] To further improve the accuracy and logical completeness of the preliminary answer set, in certain embodiments, generating a preliminary answer set based on the core question set in step 104, enhancing answer logic constraints through cross-modal association, generating an enhanced answer set, and extracting a semantic divergence indicator include:

[0126] A multimodal semantic space mapping network is constructed, each question in the core question set is converted into a semantic query vector, and a set of candidate answer fragments is retrieved from the multimodal data source of the target knowledge domain through a distributed inference engine; a semantic alignment operation is performed on the candidate answer fragment set, and the semantic matching index between the semantic query vector and the candidate answer fragment set is calculated through a cross-modal association matrix, and a set of answer fragments that meets the preliminary logical constraints is screened; a logical enhancement weight coefficient of the answer fragment set is generated based on the cross-modal association matrix, and the answer fragment set is converted into a preliminary answer set through a multimodal semantic fusion algorithm, and the semantic divergence index of the preliminary answer set is extracted.

[0127] In this embodiment, the multimodal semantic space mapping network is a technology used to map various types of data such as text, images, and audio into a unified semantic space. The semantic query vector is the representation of the core question in this space after conversion, which is used for subsequent retrieval and matching. The distributed reasoning engine can efficiently retrieve relevant information from multimodal data sources in the target knowledge domain. The cross-modal association matrix is used to measure the semantic similarity between data of different modalities, thereby ensuring that the answer fragments not only meet the logical requirements, but also supplement information from multiple perspectives. The logical enhancement weight coefficient is used to adjust the importance of different answer fragments to generate a more complete and accurate set of preliminary answers.

[0128] In an embodiment of the present application, first, the system constructs a multimodal semantic space mapping network and converts each question in the core question set into a semantic query vector. Next, a distributed reasoning engine is used to retrieve a set of candidate answer fragments in the multimodal data source of the target knowledge domain. Then, a semantic alignment operation is performed on the candidate answer fragment set, and the semantic matching index between the semantic query vector and the candidate answer fragment set is calculated using the cross-modal association matrix to screen out the answer fragment set that meets the preliminary logical constraints. Next, the logical enhancement weight coefficient of the answer fragment set is generated based on the cross-modal association matrix, and these answer fragments are combined into a preliminary answer set through a multimodal semantic fusion algorithm. Finally, the system extracts the semantic divergence index of the preliminary answer set to evaluate its diversity and coverage.

[0129] The following is a specific embodiment:

[0130] In an intelligent customer service scenario, a user asks about the warranty policy of a certain smartphone. The system first processes the core question set received and identifies key phrases such as "smartphone" and "warranty policy." The system constructs a multimodal semantic space mapping network to convert each core question (for example, "How long is the warranty period for this smartphone?") into a semantic query vector. This step ensures that the question can be accurately represented in a unified semantic space. Next, the system uses a distributed inference engine to retrieve a set of relevant candidate answer fragments from multimodal data sources in the target knowledge domain (including product manuals, user reviews, FAQ pages, etc.). For example, the system may find the following fragments: "The warranty period is one year," "Proof of purchase is required," and a screenshot showing the warranty terms.

[0131] The system performs a semantic alignment operation on the set of candidate answer fragments, and calculates the semantic matching index between the semantic query vector and the set of candidate answer fragments through the cross-modal association matrix. Suppose the system finds that the fragment "the warranty period is one year" is highly matched with the original question, and other fragments also provide useful but not completely directly related information. The logical enhancement weight coefficient of the answer fragment set is generated based on the cross-modal association matrix. For example, "the warranty period is one year" has a higher weight because it is the key information that directly answers the user's question; while "proof of purchase is required", although important, is logically secondary information. The system combines these answer fragments into a preliminary answer set through a multimodal semantic fusion algorithm, such as: "The warranty period of this smartphone is one year, and you need to provide proof of purchase to enjoy the warranty service."

[0132] Finally, the system extracts a semantic divergence metric from the preliminary answer set to assess its diversity and coverage. Suppose the system finds that the current answer set already covers the main information points, but to further enhance the user experience, the system decides to add some additional details, such as the scope of services during the warranty period or how to contact after-sales service.

[0133] In order to further improve the logical consistency and semantic coherence of the question-answer pair combination, in certain embodiments, the step 105 includes establishing a semantic focus trajectory archive based on the contextual association tag, performing a multi-dimensional topological mapping between the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength, including:

[0134] Construct a multi-dimensional topological structure of the semantic focus trajectory archive, map the semantic unit set, the core question set and the enhanced answer set into a node set in the topological graph through the context association mark, and generate the initial connectivity strength weights between nodes based on the semantic divergence index and the logical consistency scoring parameter; perform semantic focus trajectory analysis on the node set, extract the context fingerprint trajectory code of each node, and generate a semantic correlation matrix between nodes through the similarity calculation of the context fingerprint trajectory code; construct a dynamic connectivity strength calculation model based on the semantic correlation matrix and the initial connectivity strength weights, and optimize the connectivity strength between nodes through multiple rounds of iterations. degree weight, and generate a topological connectivity map in a stable state; perform cluster analysis on the topological connectivity map, and divide the node set into multiple semantic focus clusters according to the connectivity strength threshold, wherein each cluster contains a strong correlation combination of question nodes and answer nodes; perform logical consistency verification on the question-answer pair combination in the semantic focus cluster, extract the logical conflict parameters generated in the verification process, and feed the logical conflict parameters back to the dynamic connectivity strength calculation model to update the connectivity strength weight, and screen out the target question-answer pair combination from the updated connectivity strength weight, wherein the target question-answer pair combination contains the optimized value of the logical consistency scoring parameter and the semantic divergence index.

[0135] In this embodiment, the semantic focus trajectory archive is a time series archive that records semantic units and their contextual information, which is mapped to nodes in a multidimensional topological structure through context-related tags. The initial connectivity strength weights between nodes are generated based on the semantic divergence index and the logical consistency scoring parameter, and are used to measure the strength of the relationship between nodes. The semantic association matrix quantifies the semantic relevance between nodes by calculating the similarity of the context fingerprint trajectory encoding between nodes. The dynamic connectivity strength calculation model is a model for optimizing the connectivity strength weights between nodes to ensure that the final generated question-answer pair combination has high logical consistency and semantic coherence.

[0136] In an embodiment of the present application, first, the system constructs a multidimensional topological structure of the semantic focus trajectory archive, maps the semantic unit set, the core question set and the enhanced answer set into a node set in the topological graph, and generates the initial connectivity strength weights between the nodes based on the semantic divergence index and the logical consistency scoring parameter. Then, the system performs semantic focus trajectory parsing on the node set, extracts the context fingerprint trajectory encoding of each node, and generates a semantic correlation matrix between the nodes through similarity calculation. Based on the semantic correlation matrix and the initial connectivity strength weights, the system constructs a dynamic connectivity strength calculation model, optimizes the connectivity strength weights between the nodes through multiple rounds of iterations, and generates a topological connectivity map in a stable state. Then, the system performs cluster analysis on the topological connectivity map and divides the node set into multiple semantic focus clusters according to the connectivity strength threshold. Each cluster contains a strong correlation combination of question nodes and answer nodes. Finally, the system performs logical consistency verification on the question and answer pair combination within the semantic focus cluster, extracts the logical conflict parameters generated during the verification process, and feeds them back to the dynamic connectivity strength calculation model to update the connectivity strength weights, and finally outputs the target question and answer pair combination.

[0137] The following is a specific embodiment:

[0138] In an intelligent customer service scenario, a user inquires about the warranty policy for a particular smartphone. The system first processes the received core question set (e.g., "What is the warranty period for this smartphone?") and the enhanced answer set (e.g., "The warranty period is one year," "Proof of purchase is required," etc.).

[0139] The system constructs a multidimensional topological structure of a semantic focus trajectory archive, mapping a set of semantic units (e.g., "smartphone," "warranty policy"), a set of core questions (e.g., "How long is the warranty period for this smartphone?"), and a set of enhanced answers (e.g., "The warranty period is one year," "Proof of purchase required") into a set of nodes in a topological graph. Initial connectivity strength weights between nodes are generated based on semantic divergence metrics and logical consistency scoring parameters. For example, the node "Warranty period is one year" has high connectivity strength with other nodes because it directly answers the user's question. The system performs semantic focus trajectory parsing on the node set, extracting the contextual fingerprint trajectory encoding of each node and generating a semantic association matrix between the nodes through similarity calculation. Suppose the system discovers that the nodes "Warranty period is one year" and "Proof of purchase required" have a strong semantic association because they both involve specific details of the warranty policy.

[0140] Based on the semantic association matrix and the initial connectivity strength weights, the system constructs a dynamic connectivity strength calculation model, optimizes the connectivity strength weights between nodes through multiple rounds of iterations, and generates a topological connectivity graph in a stable state. For example, after multiple iterations, the system found that the connectivity strength between "the warranty period is one year" and "proof of purchase is required" has been significantly improved, indicating that the two are logically closely related. The system performs cluster analysis on the topological connectivity graph and divides the node set into multiple semantic focus clusters based on the connectivity strength threshold. Each cluster contains a strongly associated combination of question nodes and answer nodes. For example, the system groups "What is the warranty period for this smartphone?" with "The warranty period is one year" and "proof of purchase is required" into one cluster because they together constitute a complete answer.

[0141] The system performs logical consistency verification on question-answer pairs within the semantic focus cluster, extracts logical conflict parameters generated during the verification process, and feeds these into the dynamic connectivity strength calculation model to update the connectivity strength weights. For example, if the system finds that a certain answer segment is not fully consistent with the question logic (for example, missing key information), its connectivity strength weight will be adjusted to optimize the overall answer quality.

[0142] Finally, the system selects target question-answer pairs based on the updated connectivity strength weights. These pairs incorporate not only logical consistency scoring parameters but also optimized values for semantic divergence metrics, ensuring comprehensive and accurate answers. For example, a complete answer might be: "This smartphone has a one-year warranty. You will need proof of purchase to qualify for warranty service."

[0143] Figure 2 The present invention provides a structural diagram of a question-answer pair generation system based on a large language model and data enhancement, as shown in FIG. Figure 2 As shown, the system includes:

[0144] A receiving module 21 is configured to receive an original text input in a target knowledge domain, perform multi-granularity semantic unit decomposition on the original text input, and generate a set of semantic units with context fingerprints, wherein each semantic unit in the set of semantic units is bound to an associated adjacent context association tag;

[0145] Input module 22, used to input the semantic unit set into a preset multi-intent question-answering generation framework, generate a candidate question set through multiple rounds of intent iteration, and trigger a probability mask strategy to dynamically adjust the question generation direction at each iteration;

[0146] A matching module 23 is configured to perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization based on a focus shift threshold, and generate a core question set;

[0147] A generation module 24 is configured to generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set, and extract a semantic divergence index;

[0148] An establishment module 25 is configured to establish a semantic focus trajectory archive based on the contextual association mark, perform multi-dimensional topological mapping between the core question set and the enhanced answer set, drive aggregation through the trajectory archive, and screen target question-answer pair combinations based on connectivity strength;

[0149] The output module 26 is used to output a sequence of question-answer pairs containing a context fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier.

[0150] Figure 2 The question-answer pair generation system based on a large language model and data enhancement can be performed Figure 1 The implementation principles and technical effects of the method for generating question-answer pairs based on a large language model and data augmentation described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the system for generating question-answer pairs based on a large language model and data augmentation in the above embodiment has been described in detail in the embodiments of the method and will not be elaborated on here.

[0151] In one possible design, Figure 2 A xx system of the embodiment shown may be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0152] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0153] The processing component 32 is as follows Figure 1 The embodiment provides a question-answer pair generation method based on a large language model and data enhancement.

[0154] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0155] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0156] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0157] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0158] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0159] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0160] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment provides a method for generating question-answer pairs based on a large language model and data enhancement.

[0161] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0162] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0163] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A question-answer pair generation method based on a large language model and data enhancement, characterized in that: include: Receiving original text input of a target knowledge domain, performing multi-granularity semantic unit decomposition on the original text input, generating a set of semantic units with context fingerprints, wherein each semantic unit in the set of semantic units is bound to an associated adjacent context association tag; Input the semantic unit set into a preset multi-intent question-answering generation framework, generate a candidate question set through multiple rounds of intent iteration, and trigger a probability mask strategy to dynamically adjust the question generation direction in each iteration; Perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization according to a focus shift threshold, and generate a core question set; Generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set and extract a semantic divergence indicator; Establishing a semantic focus trajectory archive based on the contextual association tags, performing multi-dimensional topological mapping between the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength; Output a sequence of question-answer pairs containing a contextual fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier; The method performs semantic focus drift detection on the candidate question set, performs bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, and dynamically triggers question reorganization according to a focus displacement threshold to generate a core question set, including: Constructing a context fingerprint trajectory file for each semantic unit in the semantic unit set, and converting the context association mark into a multi-dimensional semantic focus vector through vector space projection; Performing reverse semantic trajectory analysis on each question in the candidate question set, extracting the latent semantic focus vector of the question, and performing bidirectional dynamic alignment between the latent semantic focus vector and the multidimensional semantic focus vector, and calculating a trajectory offset index between each question semantic unit pair; Based on a preset displacement threshold interval, the trajectory offset indicator is divided into a repairable offset interval and an unrepairable offset interval. An adaptive reorganization rule is triggered for problems in the repairable offset interval, and alternative candidate problems are generated through semantic focus vector interpolation technology. An elimination marking operation is performed on the problems in the unrepairable offset interval, and the alternative candidate problems are merged with the non-eliminated problems in the candidate problem set to generate a reorganized candidate problem set. The trajectory offset index is recalculated based on the reorganized candidate question set. When the trajectory offset index converges to a preset stability threshold range, a core question set is generated. Otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed.

2. The method according to claim 1, characterized in that The method divides the trajectory offset index into a repairable offset interval and an unrepairable offset interval according to a preset displacement threshold interval, triggers an adaptive reorganization rule for problems within the repairable offset interval, generates alternative candidate problems through semantic focus vector interpolation technology, performs an elimination marking operation on problems within the unrepairable offset interval, and merges the alternative candidate problems with the non-eliminated problems in the candidate problem set to generate a reorganized candidate problem set, including: Based on the context fingerprint trajectory archive, a historical semantic focus vector library based on the context association mark is constructed, and a sliding window aggregation analysis is performed on the context association mark to generate a spatiotemporal distribution map of the historical semantic focus vector; Calculating the normalized weight value of the trajectory offset indicator for each problem in the candidate problem set, and combining it with the evolution trend parameter of the corresponding semantic unit in the spatiotemporal distribution map to determine whether the problem in the candidate problem set falls into the repairable offset interval; For the problem within the repairable offset interval, extract a set of adjacent semantic focus vectors from the historical semantic focus vector library, calculate an interpolation weight coefficient based on the evolution trend parameter, perform linear interpolation in the vector space to generate an alternative candidate problem, and attach a reorganization identifier to the alternative candidate problem; The alternative candidate questions are mixed and sorted with the questions that have not been eliminated in the candidate question set, and stability verification is performed on the mixed question set based on the reorganization identifier. If the decrease in the trajectory offset indicator of the alternative candidate question does not reach a preset repair gain threshold, the interpolation weight coefficient is adjusted to generate a reorganized candidate question set.

3. The method according to claim 2, characterized in that For the problem within the repairable offset interval, extracting a set of adjacent semantic focus vectors from the historical semantic focus vector library, calculating an interpolation weight coefficient according to the evolution trend parameter, performing linear interpolation in the vector space to generate an alternative candidate problem, and adding a reorganization identifier to the alternative candidate problem, including: Based on the spatiotemporal distribution map, an interpolation path topology map of the vectors in the historical semantic focus vector library is constructed, and a set of evolutionary trend parameters corresponding to each semantic unit is extracted through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool; Analyzing the position coordinates of the semantic focus vector of the problem within the repairable offset interval in the interpolation path topology graph, calculating the semantic continuity index between adjacent vectors based on the evolution trend parameter, and screening the candidate interpolation vector group that meets the path coherence constraint; A dynamic interpolation weight matrix is constructed based on the candidate interpolation vector group, an interpolation scale coefficient is generated through the semantic density distribution and evolution direction parameters between nodes in the interpolation path topology graph, a constrained linear interpolation operation is performed in the vector space to generate alternative candidate questions, and a dynamic reorganization identifier is attached to the alternative candidate questions.

4. The method according to claim 3, characterized in that The trajectory offset index is recalculated based on the reorganized candidate question set, and when the trajectory offset index converges to a preset stability threshold range, a core question set is generated; otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed, including: Constructing a dynamic convergence judgment model for the trajectory deviation index, and generating stability baseline parameters and confidence intervals by statistically analyzing the trajectory deviation index during the historical reorganization process through a sliding window; Inputting the reorganized candidate problem set into the dynamic convergence determination model, calculating the convergence difference between the trajectory offset index and the stability baseline parameter, and triggering an iteration termination condition if the convergence difference satisfies a preset stability threshold range; When the convergence difference exceeds a preset stability threshold range, adjusting the weight coefficient of the semantic focus vector interpolation according to the gradient direction of the convergence difference, and generating an incremental candidate question set based on the unused semantic focus vectors in the context fingerprint trajectory archive; performing cross-round trajectory offset verification on the incremental candidate problem set, extracting trajectory offset fluctuation parameters generated during the verification process, and feeding the trajectory offset fluctuation parameters back to the dynamic convergence determination model to update the stability baseline parameters; If the consecutive iterations reach the preset maximum reorganization depth threshold, the iteration termination condition is forcibly triggered, and a core problem set carrying a convergence difference parameter and an incremental reorganization identifier is output.

5. The method according to claim 1, wherein The generating of a preliminary answer set based on the core question set, enhancing answer logic constraints through cross-modal association, generating an enhanced answer set and extracting a semantic divergence index includes: Constructing a multimodal semantic space mapping network to convert each question in the core question set into a semantic query vector, and retrieving a set of candidate answer fragments from a multimodal data source in the target knowledge domain through a distributed reasoning engine; Performing a semantic alignment operation on the candidate answer segment set, calculating a semantic matching index between the semantic query vector and the candidate answer segment set through a cross-modal association matrix, and screening a set of answer segments that meet preliminary logical constraints; The logical enhancement weight coefficient of the answer fragment set is generated based on the cross-modal association matrix, the answer fragment set is converted into a preliminary answer set through a multimodal semantic fusion algorithm, and the semantic divergence index of the preliminary answer set is extracted.

6. The method according to claim 1, characterized in that The process of establishing a semantic focus trajectory archive based on the contextual association tag, performing multi-dimensional topological mapping between the core question set and the enhanced answer set, driving aggregation through the trajectory archive, and screening target question-answer pair combinations based on connectivity strength includes: Constructing a multidimensional topological structure of a semantic focus trajectory archive, mapping the semantic unit set, the core question set, and the enhanced answer set into a node set in a topological graph through the context association mark, and generating initial connectivity strength weights between nodes based on the semantic divergence index and the logical consistency scoring parameter; Performing semantic focus trajectory analysis on the node set, extracting the context fingerprint trajectory code of each node, and generating a semantic correlation matrix between nodes by calculating the similarity of the context fingerprint trajectory codes; A dynamic connectivity strength calculation model is constructed based on the semantic relevance matrix and the initial connectivity strength weights, and the connectivity strength weights between nodes are optimized through multiple rounds of iterations to generate a topological connectivity graph in a stable state; Performing cluster analysis on the topological connectivity graph to divide the node set into a plurality of semantic focus clusters according to a connectivity strength threshold, wherein each cluster contains a strongly associated combination of question nodes and answer nodes; Perform logical consistency verification on the question-answer pair combination within the semantic focus cluster, extract the logical conflict parameters generated during the verification process, feed the logical conflict parameters back to the dynamic connectivity strength calculation model to update the connectivity strength weight, and screen out the target question-answer pair combination from the updated connectivity strength weight, wherein the target question-answer pair combination includes the optimized values of the logical consistency scoring parameter and the semantic divergence index.

7. A question-answer pair generation system based on a large language model and data enhancement, characterized in that: include: A receiving module is configured to receive an original text input of a target knowledge domain, perform multi-granularity semantic unit decomposition on the original text input, and generate a set of semantic units with context fingerprints, wherein each semantic unit in the set of semantic units is bound to an associated adjacent context association tag; An input module is used to input the semantic unit set into a preset multi-intent question-answering generation framework, generate a set of candidate questions through multiple rounds of intent iteration, and trigger a probabilistic masking strategy to dynamically adjust the question generation direction at each iteration; a matching module configured to perform semantic focus drift detection on the candidate question set, perform bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically trigger question reorganization based on a focus shift threshold, and generate a core question set; A generation module, configured to generate a preliminary answer set based on the core question set, enhance the answer logic constraints through cross-modal association, generate an enhanced answer set, and extract a semantic divergence indicator; An establishment module is used to establish a semantic focus trajectory archive based on the contextual association mark, perform multi-dimensional topological mapping between the core question set and the enhanced answer set, drive aggregation through the trajectory archive, and screen target question-answer pair combinations based on connectivity strength; An output module is configured to output a sequence of question-answer pairs containing a contextual fingerprint traceability chain, wherein each question-answer pair integrates the semantic focus trajectory archive, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier; The method performs semantic focus drift detection on the candidate question set, performs bidirectional trajectory matching between the question and the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, and dynamically triggers question reorganization according to a focus displacement threshold to generate a core question set, including: Constructing a context fingerprint trajectory file for each semantic unit in the semantic unit set, and converting the context association mark into a multi-dimensional semantic focus vector through vector space projection; Performing reverse semantic trajectory analysis on each question in the candidate question set, extracting the latent semantic focus vector of the question, and performing bidirectional dynamic alignment between the latent semantic focus vector and the multidimensional semantic focus vector, and calculating a trajectory offset index between each question semantic unit pair; Based on a preset displacement threshold interval, the trajectory offset indicator is divided into a repairable offset interval and an unrepairable offset interval. An adaptive reorganization rule is triggered for problems in the repairable offset interval, and alternative candidate problems are generated through semantic focus vector interpolation technology. An elimination marking operation is performed on the problems in the unrepairable offset interval, and the alternative candidate problems are merged with the non-eliminated problems in the candidate problem set to generate a reorganized candidate problem set. The trajectory offset index is recalculated based on the reorganized candidate question set. When the trajectory offset index converges to a preset stability threshold range, a core question set is generated. Otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively performed.

8. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a question-answer pair generation method based on a large language model and data enhancement as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for generating question-answer pairs based on a large language model and data enhancement is implemented as described in any one of claims 1 to 6.

Images