Question and answer pair generation method and system based on large language model and data enhancement

Abstract

The invention provides a question-answer pair generation method and system based on a large language model and data enhancement, and the method comprises the steps: carrying out the processing of a text of a target domain, decomposing the text into a semantic unit set, and carrying out the iteration of a multi-intention question-answer generation frame to generate candidate questions; performing semantic focus drift detection, and forming a core problem set according to the context fingerprint bidirectional matching adjustment problem; generating a preliminary answer based on the set, and extracting a semantic divergence index by using cross-modal association enhancement logic; establishing a semantic focus track file, mapping the question and the enhanced answer set, and screening out an optimal question and answer pair combination; and outputting a question and answer sequence containing a context traceability chain, an integrated semantic trajectory, divergence and a cross-modal intensity factor. The efficiency of identifying and generating the high-quality question and answer pairs from the complex text is improved.

Description

Technical Field

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

[0002] In modern knowledge management and intelligent question-answer systems, it is crucial to generate high-efficiency and accurate question-answer pairs for large-scale text data in specific fields. This need is particularly evident in highly specialized and information-intensive fields such as healthcare, legal consulting, and scientific research. These fields not only require the system to accurately understand complex semantics but also need to have cross-modal data processing capabilities to support knowledge expression forms with pictures and texts, and ensure the logic and integrity of the answers.

[0003] Current technical solutions usually adopt large language models based on deep learning combined with data augmentation technology to achieve the automatic generation of question-answer pairs. Such methods first decompose the original text into multi-granularity semantic units to form a preliminary set of semantic units. Then, using a pre-trained language model framework, a candidate question set is generated through multiple iterations of the intention recognition process. In addition, to improve the quality and pertinence of the questions, a probability masking strategy is introduced to dynamically adjust the direction of question generation. For the generated question set, existing solutions also adopt a semantic focus drift detection and adaptive reconstruction mechanism to ensure the bidirectional matching of questions and context fingerprints, and then screen out the core question set. Finally, these questions will be used to generate preliminary answers, and the logical constraints of the answers are enhanced through cross-modal association to form the final answer set.

[0004] Although the above methods have improved the performance of the question-answer system to a certain extent, there are still some deficiencies. First, in the semantic unit decomposition stage, although basic semantic information can be captured, the understanding of deep semantic structures is still insufficient, resulting in possible semantic deviations in the generated questions and answers. Second, although the probability masking strategy in the existing solutions helps to adjust the direction of question generation, its flexibility is limited and it is difficult to cope with complex and changing actual application scenarios. Finally, regarding the application of cross-modal association, existing methods focus more on information integration at the text level, and less exploration of the effective integration of non-text information such as images and audio, which limits the expressiveness and accuracy of the system in a multimedia environment. Summary of the Invention

[0005] The embodiments of the present application provide a method and system for generating question-answer pairs based on large language models and data augmentation, 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, an embodiment of the present application provides a method for generating question-answer pairs based on a large language model and data augmentation, including: Receiving an original text input in a target knowledge domain, decomposing the original text input into multi-granularity semantic units, generating a set of semantic units with context fingerprints, and binding adjacent context association tags to each semantic unit in the set of semantic units; Inputting the set of semantic units into a pre-set multi-intent question generation framework, generating a set of candidate questions through multi-round intent iteration, and dynamically adjusting the question generation direction by triggering a probability masking strategy each time; Performing semantic focus drift detection on the set of candidate questions, performing bidirectional trajectory matching between the questions and the context fingerprints of the set of semantic units through an adaptive reconstruction mechanism, and dynamically triggering question recombination according to a focus displacement threshold to generate a set of core questions; Generating a preliminary answer set based on the set of core questions, enhancing the answer logic constraints through cross-modal association, generating an enhanced answer set and extracting a semantic divergence index; Establishing a semantic focus trajectory file based on the context association tags, performing multi-dimensional topological mapping on the set of core questions and the enhanced answer set, driving aggregation through the trajectory file, and screening target question-answer pair combinations according to the connection strength; Outputting a sequence of question-answer pairs including a context fingerprint traceability chain, where each question-answer pair integrates the semantic focus trajectory file, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier.

[0007] Optionally, the performing semantic focus drift detection on the set of candidate questions, performing bidirectional trajectory matching between the questions and the context fingerprints of the set of semantic units through an adaptive reconstruction mechanism, and dynamically triggering question recombination according to a focus displacement threshold to generate a set of core questions includes: Constructing a context fingerprint trajectory file for each semantic unit in the set of semantic units, and converting the context association tags into multi-dimensional semantic focus vectors through vector space projection; Performing reverse semantic trajectory analysis on each question in the set of candidate questions, extracting the potential semantic focus vector of the question, and performing bidirectional dynamic alignment between the potential semantic focus vector and the multi-dimensional semantic focus vectors to calculate the trajectory offset index between each question semantic unit pair; According to a preset displacement threshold interval, divide the trajectory offset index into a repairable offset interval and an irreparable offset interval, trigger an adaptive recombination rule for the problems in the repairable offset interval, generate alternative candidate problems through semantic focus vector interpolation technology, perform an elimination marking operation on the problems in the irreparable offset interval, and merge the alternative candidate problems with the uneliminated problems in the candidate problem set to generate a recombined candidate problem set; Recalculate the trajectory offset index based on the recombined candidate problem set. When the trajectory offset index converges to a preset stability threshold range, generate a core problem set; otherwise, iteratively execute semantic focus vector interpolation and trajectory offset verification.

[0008] Optionally, the step of dividing the trajectory offset index into a repairable offset interval and an irreparable offset interval according to a preset displacement threshold interval, triggering an adaptive recombination rule for the problems in the repairable offset interval, generating alternative candidate problems through semantic focus vector interpolation technology, performing an elimination marking operation on the problems in the irreparable offset interval, and merging the alternative candidate problems with the uneliminated problems in the candidate problem set to generate a recombined candidate problem set includes: Based on the context fingerprint trajectory archive, construct a historical semantic focus vector library based on the context association markers, and perform a sliding window aggregation analysis on the context association markers to generate a spatio-temporal distribution map of the historical semantic focus vectors; Calculate the normalized weight value of the trajectory offset index for each problem in the candidate problem set, and combine it with the evolution trend parameter of the corresponding semantic unit in the spatio-temporal distribution map to determine whether the problems in the candidate problem set fall into the repairable offset interval; For the problems in the repairable offset interval, extract the adjacent semantic focus vector set from the historical semantic focus vector library, calculate the interpolation weight coefficient according to the evolution trend parameter, perform linear interpolation in the vector space to generate alternative candidate problems, and attach a recombination identifier to the alternative candidate problems; Mix and sort the alternative candidate problems with the uneliminated problems in the candidate problem set, perform stability verification on the mixed problem set based on the recombination identifier. If the decrease amplitude of the trajectory offset index of the alternative candidate problems does not reach the preset repair gain threshold, adjust the interpolation weight coefficient to generate a recombined candidate problem set.

[0009] Optionally, for the problems within the repairable offset interval, extract the adjacent semantic focus vector set from the historical semantic focus vector library, calculate the interpolation weight coefficient according to the evolution trend parameter, perform linear interpolation in the vector space to generate alternative candidate problems, and attach a recombination identifier to the alternative candidate problems, including: Construct an interpolation path topology graph of the vectors in the historical semantic focus vector library based on the spatio-temporal distribution atlas, extract the evolution trend parameter set corresponding to each semantic unit through sliding window aggregation analysis, and generate a multi-level interpolation path candidate pool; Analyze the position coordinates of the semantic focus vector of the problem within the repairable offset interval in the interpolation path topology graph, calculate the semantic continuity index between adjacent vectors according to the evolution trend parameter, and screen the candidate interpolation vector groups that meet the path coherence constraint; Construct a dynamic interpolation weight matrix based on the candidate interpolation vector group, generate an interpolation ratio coefficient through the semantic density distribution and evolution direction parameter between nodes in the interpolation path topology graph, perform constrained linear interpolation operation in the vector space to generate alternative candidate problems, and attach a dynamic recombination identifier to the alternative candidate problems.

[0010] Optionally, recalculate the trajectory offset index based on the recombined candidate problem set. When the trajectory offset index converges to a preset stability threshold range, generate a core problem set; otherwise, iteratively execute semantic focus vector interpolation and trajectory offset verification, including: Construct a dynamic convergence determination model for the trajectory offset index, statistically analyze the trajectory offset index in the historical recombination process through a sliding window, and generate stability baseline parameters and confidence intervals; Input the recombined candidate problem set into the dynamic convergence determination model, calculate the convergence difference degree between the trajectory offset index and the stability baseline parameter. If the convergence difference degree meets the preset stability threshold range, trigger the iteration termination condition; When the convergence difference degree exceeds the preset stability threshold range, adjust the weight coefficient of semantic focus vector interpolation according to the gradient direction of the convergence difference degree, and generate an incremental candidate problem set based on the unused semantic focus vectors in the context fingerprint trajectory file; Perform cross-round trajectory offset verification on the incremental candidate problem set, extract the trajectory offset fluctuation parameters generated during the verification process, and feedback the trajectory offset fluctuation parameters to the dynamic convergence determination model to update the stability baseline parameters; If the continuous iteration rounds reach the preset maximum recombination depth threshold, forcibly trigger the iteration termination condition, and output a core problem set carrying the convergence difference degree parameter and the incremental recombination identifier.

[0011] Optionally, generating a preliminary answer set based on the set of core questions, enhancing the answer logic constraints through cross-modal association, generating an enhanced answer set, and extracting a semantic divergence metric, including: Constructing a multi-modal semantic space mapping network, converting each question in the set of core questions into a semantic query vector, and retrieving a set of candidate answer fragments in a multi-modal data source in the target knowledge domain through a distributed inference engine; Performing a semantic alignment operation on the set of candidate answer fragments, calculating a semantic matching metric between the semantic query vector and the set of candidate answer fragments through a cross-modal association matrix, and filtering out a set of answer fragments that meet the preliminary logical constraints; Generating a logical enhancement weight coefficient for the set of answer fragments based on the cross-modal association matrix, converting the set of answer fragments into a preliminary answer set through a multi-modal semantic fusion algorithm, and extracting a semantic divergence metric for the preliminary answer set.

[0012] Optionally, establishing a semantic focus trajectory archive based on the context association tags, performing a multi-dimensional topological mapping of the set of core questions and the enhanced answer set, driving aggregation through the trajectory archive, and filtering out a target Q&A pair combination according to the connection strength, including: Constructing a multi-dimensional topological structure of the semantic focus trajectory archive, mapping the set of semantic units, the set of core questions, and the enhanced answer set into a set of nodes in a topological graph through the context association tags, and generating an initial connection strength weight between nodes based on the semantic divergence metric and the logical consistency scoring parameter; Performing semantic focus trajectory parsing on the set of nodes, extracting the context fingerprint trajectory encoding of each node, and generating a semantic association matrix between nodes through the similarity calculation of the context fingerprint trajectory encoding; Constructing a dynamic connection strength calculation model based on the semantic association matrix and the initial connection strength weight, optimizing the connection strength weight between nodes through multiple rounds of iteration, and generating a topological connection graph in a stable state; Performing clustering analysis on the topological connection graph, dividing the set of nodes into multiple semantic focus clusters according to a connection strength threshold, where each cluster contains a strongly associated combination of question nodes and answer nodes; Performing logical consistency verification on the Q&A pair combinations within the semantic focus cluster, extracting the logical conflict parameters generated during the verification process, feeding the logical conflict parameters back to the dynamic connection strength calculation model to update the connection strength weight, and screening out a target Q&A pair combination, where the target Q&A pair combination contains an optimized value of the logical consistency scoring parameter and the semantic divergence metric.

[0013] Second aspect, an embodiment of the present application provides a question-and-answer pair generation system based on a large language model and data augmentation, including: A receiving module, configured to receive an original text input in a target knowledge domain, decompose the original text input into multi-granularity semantic units, generate a set of semantic units with context fingerprints, and bind adjacent context association tags associated with each semantic unit in the set of semantic units; An input module, configured to input the set of semantic units into a preset multi-intention question generation framework, generate a set of candidate questions through multi-round intention iteration, and trigger a probability masking strategy to dynamically adjust the question generation direction each time; A matching module, configured to perform semantic focus drift detection on the set of candidate questions, perform two-way trajectory matching between the questions and the context fingerprints of the set of semantic units through an adaptive reconstruction mechanism, and dynamically trigger question recombination according to a focus displacement threshold to generate a set of core questions; A generation module, configured to generate a preliminary answer set based on the set of core questions, enhance the answer logic constraint through cross-modal association, generate an enhanced answer set and extract a semantic divergence index; A building module, configured to build a semantic focus trajectory file based on the context association tags, perform multi-dimensional topological mapping on the set of core questions and the enhanced answer set, drive aggregation through the trajectory file, and screen out target question-and-answer pair combinations according to the connection strength; An output module, configured to output a sequence of question-and-answer pairs including a context fingerprint traceability chain, where each question-and-answer pair integrates the semantic focus trajectory file, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topological identifier.

[0014] Third aspect, an embodiment of the present application provides a computing device, including a processor and a memory, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute the method for generating a question-and-answer pair based on a large language model and data augmentation according to any one of the first aspects.

[0015] Fourth aspect, an embodiment of the present application provides a computer storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method for generating a question-and-answer pair based on a large language model and data augmentation according to any one of the first aspects is implemented.

[0016] In the embodiments of the present application, the original text input of the 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, and each semantic unit in the set of semantic units is bound with an associated adjacent context association mark; the set of semantic units is input into a preset multi-intent question-and-answer generation framework, and a set of candidate questions is generated through multi-round intent iteration, and the question generation direction is dynamically adjusted by triggering a probability masking strategy in each iteration; semantic focus drift detection is performed on the set of candidate questions, and the questions are bidirectionally trajectory-matched with the context fingerprints of the set of semantic units through an adaptive reconstruction mechanism, and question recombination is dynamically triggered according to the focus displacement threshold to generate a set of core questions; a preliminary answer set is generated based on the set of core questions, the answer logic constraint is enhanced through cross-modal association, an enhanced answer set is generated and the semantic divergence index is extracted; a semantic focus trajectory file is established based on the context association mark, the set of core questions and the enhanced answer set are subjected to multi-dimensional topological mapping, and aggregation is driven by the trajectory file, and the target question-and-answer pair combination is screened according to the connection strength; a question-and-answer pair sequence including a context fingerprint traceability chain is output, and each question-and-answer pair integrates the semantic focus trajectory file, the semantic divergence index and the strength factor of the cross-modal association to generate a multi-level knowledge topological identifier.

[0017] The technical solution of the present application has the following beneficial effects: The present application not only improves the efficiency of identifying and generating high-quality question-and-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 context association marks and semantic focus trajectory files, multi-level topological identification of knowledge is realized, which helps to better understand and apply information, and greatly optimizes the knowledge management and retrieval process.

[0018] Further, in the process of performing semantic focus drift detection on the set of candidate questions in the embodiments of the present application, a context fingerprint trajectory file of each semantic unit is constructed, and the context association mark is converted into a multi-dimensional semantic focus vector by using vector space projection technology. Then, reverse semantic trajectory analysis is performed on each question in the set of candidate questions, its potential semantic focus vector is extracted, and it is bidirectionally dynamically aligned with the multi-dimensional semantic focus vector to calculate the trajectory offset index between the question semantic unit pairs. According to the preset displacement threshold interval, the repairable and non-repairable offset intervals are distinguished. For the former, an adaptive recombination rule is applied and alternative candidate questions are generated through semantic focus vector interpolation, and the latter is eliminated. Finally, the trajectory offset is re-evaluated based on the recombined set of candidate questions until it converges within the stability threshold range, and a set of core questions is output.

[0019] 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 better conforms to the semantic structure of the original text. By applying the semantic focus vector interpolation technique, it can not only effectively repair some offset problems but also improve its accuracy and pertinence while maintaining the question intention. In addition, the iterative process of performing semantic focus vector interpolation and trajectory offset verification further enhances the robustness and stability of the system, making the finally output set of core questions better reflect the true needs of users and the deep structure of the knowledge domain. This method greatly improves the ability of the question-answering system to process complex texts and enhances the user experience.

[0020] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. Brief Description of the Drawings

[0021] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a flowchart of a method for generating question-answer pairs based on a large language model and data augmentation provided by an embodiment of the present application; Figure 2 It is a schematic structural diagram of a system for generating question-answer pairs based on a large language model and data augmentation provided by an embodiment of the present application; Figure 3 It is a schematic structural diagram of a computing device provided by an embodiment of the present application. Detailed Embodiments

[0023] To enable those skilled in the art to better understand the solution of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application.

[0024] In some of the processes described in the specification, claims, and the above-mentioned drawings of this application, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any order of execution. Additionally, 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 such as "first", "second", etc. in this text are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0025] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0026] Figure 1 The present application provides a flowchart of a method for generating question-and-answer pairs based on a large language model and data augmentation, as Figure 1 shown, the method includes: Step 101: Receive the original text input of the target knowledge domain, perform multi-granularity semantic unit decomposition on the original text input, and generate a set of semantic units with context fingerprints; In this step, multi-granularity semantic unit decomposition is a process of parsing the original text input into semantic units at different levels (such as sentences, phrases, or words), and attaching context fingerprints and adjacent context association tags to each semantic unit. The context fingerprint helps to determine the unique position of each semantic unit in its surrounding environment, while the context association tag identifies the adjacent semantic units. This method helps to more accurately understand the text content and its background information.

[0027] In actual operation, the system first receives the original text input of the target knowledge domain, and then uses natural language processing technology to parse the text at multiple levels to generate a set of semantic units with context information. During this process, by analyzing the relationship between each semantic unit and its context before and after, the corresponding association tags are added to facilitate better understanding and processing of this information in subsequent steps.

[0028] For example, in an intelligent customer service scenario, assume that a user asks for specific details about the return policy of a certain electronic product. After receiving the query, the system first decomposes this text into multi-granularity semantic units. For example, "I want to know the return policy of this electronic product" is decomposed into multiple semantic units: "I want to know", "this electronic product", and "return policy". The system adds context fingerprints and association markers to each semantic unit. For example, "I want to know" may be bound to the relevant context of the user's need expression, while "return policy" is associated with specific terms and conditions. This step ensures the accuracy of subsequent question generation and answer extraction.

[0029] Step 102: Input the set of semantic units into a pre-set multi-intent question generation framework, and generate a set of candidate questions through multi-round intent iteration. Each iteration triggers a probability masking strategy to dynamically adjust the question generation direction. In this step, the multi-intent question generation framework is a deep learning-based system used to generate a series of possible questions based on the set of semantic units. In each iteration, the probability masking strategy dynamically adjusts the question generation direction to increase the diversity and pertinence of the questions, ensuring that the generated questions cover a wide range and meet the actual needs.

[0030] In actual operation, input the set of semantic units generated in Step 101 into the pre-set question generation framework, and generate a set of candidate questions through multiple iterations. During this process, a probability masking strategy is adopted to increase the flexibility and pertinence of question generation, ensuring that the generated questions are not only diverse but also highly practical.

[0031] For example, in the above intelligent customer service scenario, the system uses the multi-intent question generation framework to generate multiple relevant questions for the query "I want to know the return policy of this electronic product", such as: "How long is the return period of this electronic product?", "What supporting documents are required for return?", "Can the product be returned if it is damaged?", etc. These questions cover different aspects that the user may be interested in, thus providing comprehensive information support.

[0032] Step 103: Perform semantic focus drift detection on the set of candidate questions, and through an adaptive reconstruction mechanism, perform two-way trajectory matching between the questions and the context fingerprints of the set of semantic units. Dynamically trigger question recombination according to the focus displacement threshold to generate a set of core questions. In this step, semantic focus drift detection aims to evaluate the matching degree between the candidate questions and the original semantic units, achieve two-way trajectory matching through an adaptive reconstruction mechanism, and decide whether to recombine the questions according to the displacement threshold to generate a more accurate set of core questions. This method can effectively reduce the problem deviation caused by semantic drift.

[0033] In actual operation, semantic focus drift detection is performed on the candidate question set, and the trajectory offset between each question and the semantic unit set is calculated. Based on a preset displacement threshold, the question reorganization rule is triggered to generate a core question set. This process ensures the accuracy and pertinence of the questions.

[0034] For example, continuing with the above case, the system detects that some questions such as "How to contact after-sales service?" are related to returns but deviate from the main concern "return policy". Therefore, the system readjusts these questions to generate a core question set that is more focused on the return process, such as: "What is the validity period of the return policy?", "What documents need to be prepared for returns?", etc. These questions directly respond to the user's initial query and improve the user experience.

[0035] Step 104: Generate a preliminary answer set based on the core question set, enhance the answer logic constraint through cross-modal association, generate an enhanced answer set, and extract the semantic divergence index; In this step, enhancing the answer logic constraint through cross-modal association means combining information in various data forms such as text and images to optimize the quality of the preliminary answer and extract the semantic divergence index that measures the diversity of answers. This method improves the integrity and logic of the answers by integrating multiple information sources.

[0036] In actual operation, a preliminary answer set is generated based on the core question set. By introducing cross-modal association (such as charts, videos, etc.), the logic and integrity of the answers are enhanced, and at the same time, the semantic divergence index is extracted to evaluate the diversity of the answers. This step ensures that the answers are not only detailed but also easy to understand.

[0037] For example, for the core question "What documents need to be prepared for returns?", the system not only provides a detailed document list but also attaches a sample table picture showing the correct filling format. In addition, the system calculates the semantic divergence index to ensure that the information provided is both comprehensive and not redundant, avoiding the problem of information overload. This multi-modal answering method enables users to quickly obtain the required information and operate correctly.

[0038] Step 105: Establish a semantic focus trajectory file based on the context association tag, perform multi-dimensional topological mapping on the core question set and the enhanced answer set, drive aggregation through the trajectory file, and screen out the target Q&A pair combination according to the connection strength; In this step, the semantic focus trajectory file established based on the context association tag performs multi-dimensional topological mapping on the core question and the enhanced answer set, uses the trajectory file to drive aggregation, and screens out the optimal Q&A combination. This method enhances the coherence and consistency of the Q&A pairs through the context association tag and the trajectory file.

[0039] In actual operation, a semantic focus trajectory file is constructed based on context - related markers, the set of core questions is matched with the enhanced answer set, and the best Q&A pair combinations are filtered out through connection strength. This process ensures the best match of Q&A pairs and improves the overall performance of the system.

[0040] For example, in an intelligent customer service system, based on context - related markers, the system matches "What documents are needed when returning goods?" with a detailed document list and example table pictures to form high - quality Q&A pairs. For example, the system shows a text description of the list of documents required for returning goods, accompanied by a clear example table to guide users on how to fill out these documents. This approach not only improves the readability of information but also enhances users' trust.

[0041] Step 106: Output a sequence of Q&A pairs containing the context fingerprint traceability chain, where each Q&A pair integrates the semantic focus trajectory file, the semantic divergence index, and the strength factor of cross - modal association to generate a multi - level knowledge topology identifier.

[0042] In this step, a sequence of Q&A pairs containing the context fingerprint traceability chain is output. Each Q&A pair integrates the semantic focus trajectory file, the semantic divergence index, and the strength factor of cross - modal association to form a multi - level knowledge topology identifier. This method enhances the user experience through the integration of rich background information and data types.

[0043] In actual operation, the final output is a sequence of Q&A pairs supported by the context fingerprint traceability chain. These Q&A pairs not only contain rich background information but also enhance users' understanding and experience through the integration of multiple data types. This process ensures the high availability of Q&A pairs and user satisfaction.

[0044] For example, the intelligent customer service system provides users with a complete Q&A sequence, including all steps from asking a question to obtaining a detailed answer, accompanied by a flowchart and operation guide for returning goods. For example, for the question "What documents are needed when returning goods?", the system not only provides a detailed document list and example table but also attaches a video tutorial on the return process to help users more intuitively understand the whole process. This multi - level knowledge presentation method greatly improves user satisfaction and the efficiency of problem - solving.

[0045] Through the method of the above six steps, the intelligent customer service system can efficiently and accurately extract and generate high - quality Q&A pairs from complex texts. This method not only improves the pertinence of questions and the accuracy of answers but also enhances users' understanding and experience through the integration of cross - modal data. In addition, through the application of the semantic focus trajectory file and the context fingerprint traceability chain, the robustness and reliability of the system are further improved, enabling the intelligent customer service to better serve diverse and complex user needs, greatly improving user satisfaction and system efficiency.

[0046] In order to further improve the accuracy and flexibility of semantic focus drift detection, in some embodiments, for the semantic focus drift detection performed on the candidate question set in step 103, a two-way trajectory matching is performed between the questions and the context fingerprints of the semantic unit set through an adaptive reconstruction mechanism, and question recombination is dynamically triggered according to the focus displacement threshold to generate a core question set, including: Construct a context fingerprint trajectory profile for each semantic unit in the semantic unit set, and convert the context association markers into multi-dimensional semantic focus vectors through vector space projection; perform reverse semantic trajectory analysis on each question in the candidate question set, extract the potential semantic focus vectors of the questions, and perform two-way dynamic alignment between the potential semantic focus vectors and the multi-dimensional semantic focus vectors to calculate the trajectory offset index between each pair of question semantic units; according to the preset displacement threshold interval, divide the trajectory offset index into a repairable offset interval and an irreparable offset interval, trigger the adaptive recombination rule for the questions in the repairable offset interval, generate alternative candidate questions through semantic focus vector interpolation technology, perform an elimination marking operation on the questions in the irreparable offset interval, and merge the alternative candidate questions with the uneliminated questions in the candidate question set to generate a recombined candidate question set; recalculate the trajectory offset index based on the recombined candidate question set, and when the trajectory offset index converges to the preset stability threshold range, generate a core question set, otherwise iteratively perform semantic focus vector interpolation and trajectory offset verification.

[0047] In this embodiment, the context fingerprint trajectory profile refers to constructing a time series profile containing its context information for each semantic unit. Through vector space projection, the system converts these context association markers into multi-dimensional semantic focus vectors for subsequent comparison and analysis. This method enables the system to more accurately capture the position of each semantic unit in its context environment and its relationship with other units. Reverse semantic trajectory analysis is used to start from the question, trace back its potential semantic focus vector, and compare it with the multi-dimensional semantic focus vector of the original semantic unit to evaluate the degree of deviation.

[0048] In the embodiments of the present application, first, the system constructs a context fingerprint trajectory file for each semantic unit and converts it into a multi-dimensional semantic focus vector. Then, reverse semantic trajectory analysis is performed on each question in the candidate question set to extract its potential semantic focus vector, and a two-way dynamic alignment is performed with the multi-dimensional semantic focus vector to calculate the trajectory offset index. Next, according to the preset displacement threshold interval, the repairable and non-repairable offset intervals are distinguished. For the questions in the repairable offset interval, the semantic focus vector interpolation technology is used to generate alternative candidate questions; while for the questions in the non-repairable offset interval, they are directly eliminated. Finally, the trajectory offset index is recalculated based on the recombined candidate question set until it converges within the preset stability threshold range, and the core question set is output.

[0049] The following is a specific example: In an intelligent customer service system, a user asks about the warranty policy of a certain smartphone. First, the system performs multi-granularity semantic unit decomposition on the received query text and identifies key phrases such as "smartphone" and "warranty policy". Then, the system constructs a context fingerprint trajectory file for each semantic unit and converts the context association markers into multi-dimensional semantic focus vectors through vector space projection. For example, "warranty policy" may be marked as associated with "product model", "purchase date", etc.

[0050] Subsequently, the system generates multiple candidate questions, such as: "How long is the warranty period of this smartphone?", "What proof materials need to be provided during the warranty period?", etc. By performing reverse semantic trajectory analysis on these questions, the system extracts the potential semantic focus vector of each question and performs a two-way dynamic alignment with the multi-dimensional semantic focus vector of the original semantic unit to calculate the trajectory offset index. Suppose the system finds that some questions (such as "How to contact after-sales service?") have a large trajectory offset and belong to the non-repairable offset interval, so they are marked for elimination. For other questions in the repairable offset interval, the system uses the semantic focus vector interpolation technology to generate alternative candidate questions, such as: "Can the battery be replaced during the warranty period?" Finally, the system recalculates the trajectory offset index based on the recombined candidate question set. After multiple iterative adjustments, until the trajectory offset indices of all questions converge within the preset stability threshold range, a set of core question sets is output. In this way, the intelligent customer service system can not only more accurately understand the user's query intention, but also provide answers that are more in line with the actual needs, significantly improving the user experience and service quality.

[0051] In order to further improve the accuracy and flexibility of semantic focus drift detection, in some embodiments, in step 103, according to a preset displacement threshold interval, the trajectory offset index is divided into a repairable offset interval and an irreparable offset interval. An adaptive recombination rule is triggered for the 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 irreparable offset interval, and the alternative candidate problems are merged with the uneliminated problems in the candidate problem set to generate a recombined candidate problem set. It further includes: Based on the context fingerprint trajectory archive, construct a historical semantic focus vector library based on the context association markers, and perform a sliding window aggregation analysis on the context association markers to generate a spatio-temporal distribution map of the historical semantic focus vectors; calculate the normalized weight value of the trajectory offset index for each problem in the candidate problem set, and combine the evolution trend parameters of the corresponding semantic units in the spatio-temporal distribution map to determine whether the problems in the candidate problem set fall into the repairable offset interval; for the problems in the repairable offset interval, extract the adjacent semantic focus vector set from the historical semantic focus vector library, calculate the interpolation weight coefficient according to the evolution trend parameters, perform linear interpolation in the vector space to generate alternative candidate problems, and attach a recombination identifier to the alternative candidate problems; perform a mixed sorting on the alternative candidate problems and the uneliminated problems in the candidate problem set, and perform a stability verification on the mixed problem set based on the recombination identifier. If the decrease amplitude of the trajectory offset index of the alternative candidate problems does not reach the preset repair gain threshold, adjust the interpolation weight coefficient to generate a recombined candidate problem set.

[0052] In this embodiment, the historical semantic focus vector library is constructed based on the context fingerprint trajectory archive, which contains the semantic focus vectors of each semantic unit at different time points. These vectors form a spatio-temporal distribution map through sliding window aggregation analysis, showing the trend of semantic units changing over time and context. The normalized weight value is used to measure the deviation degree between each candidate problem and its original semantic unit, while the evolution trend parameter helps predict the possible future change direction.

[0053] In the embodiments of the present application, first, the system constructs a historical semantic focus vector library based on the context fingerprint trajectory archive, and generates a spatio-temporal distribution map by performing sliding window aggregation analysis on the context association tags. Next, the system calculates the normalized weight values of the trajectory offset metrics for each question in the candidate question set, and determines whether the question belongs to the repairable offset interval in combination with the evolution trend parameters of the corresponding semantic units in the spatio-temporal distribution map. For repairable questions, the system extracts the adjacent semantic focus vector set from the historical semantic focus vector library, calculates the interpolation weight coefficient according to the evolution trend parameter, and performs linear interpolation in the vector space to generate alternative candidate questions. The system attaches a recombination identifier to these alternative candidate questions and mixes and sorts them with the original candidate questions that have not been eliminated. Finally, based on the recombination identifier, the system performs stability verification on the mixed question set. If the decline amplitude of the trajectory offset metric of the alternative candidate question does not reach the preset repair gain threshold, the interpolation weight coefficient is adjusted to regenerate the recombined candidate question set.

[0054] The following is a specific embodiment: In an intelligent customer service scenario, the user asks about the specific details of the return policy for a certain electronic product. The system first performs multi-granularity semantic unit decomposition on the received query text and identifies key phrases such as "electronic product" and "return policy". Then, the system constructs a context fingerprint trajectory archive for each semantic unit and generates a historical semantic focus vector library. Through sliding window aggregation analysis, the system generates a spatio-temporal distribution map of these semantic units, showing their changing trends over time.

[0055] Next, the system generates multiple candidate questions, such as: "How long is the return period for this electronic product?", "What supporting documents are required for returns?", etc. The system calculates the normalized weight values of the trajectory offset metrics for each question and evaluates whether the question belongs to the repairable offset interval in combination with the evolution trend parameters of the corresponding semantic units in the spatio-temporal distribution map. Suppose the question "How to contact after-sales service?" is determined to be non-repairable and is directly eliminated. For the question "What supporting documents are required for returns?", since its trajectory offset is small and it 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 according to the evolution trend parameter to generate an alternative candidate question: "What documents need to be prepared for returns?". The system assigns recombination identifiers to these alternative candidate questions and mixes and sorts them with the original candidate questions that have not been eliminated. Subsequently, the system verifies the stability of the mixed set of questions. If it is found that the decline in the trajectory offset index of some alternative candidate questions does not reach the preset repair gain threshold (for example, there is still a slight offset in "What documents need to be prepared when returning a product?"), the interpolation weight coefficient is adjusted to regenerate more accurate alternative candidate questions. Finally, the system outputs an optimized set of core questions, significantly improving the accuracy and user experience of the question-and-answer system.

[0056] To further improve the accuracy and flexibility of semantic focus drift detection, in some embodiments, for the questions within the repairable offset interval in step 103, extracting an adjacent semantic focus vector set 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 alternative candidate questions, and assigning a recombination identifier to the alternative candidate questions further includes: Constructing an interpolation path topology graph of the vectors in the historical semantic focus vector library based on the spatio-temporal distribution atlas, extracting a set of evolution trend parameters corresponding to each semantic unit through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool; analyzing the position coordinates of the semantic focus vectors of the questions within the repairable offset interval in the interpolation path topology graph, calculating the semantic continuity index between adjacent vectors according to the evolution trend parameter, and screening a candidate interpolation vector group that satisfies the path coherence constraint; constructing a dynamic interpolation weight matrix based on the candidate interpolation vector group, generating an interpolation ratio coefficient through the semantic density distribution and evolution direction parameter between nodes in the interpolation path topology graph, performing constrained linear interpolation operations in the vector space to generate alternative candidate questions, and assigning a dynamic recombination identifier to the alternative candidate questions.

[0057] In this embodiment, the interpolation path topology graph is constructed based on the spatio-temporal distribution atlas, which shows the relationships between historical semantic focus vectors and their changing trends over time. The set of evolution trend parameters for each semantic unit includes its changes in different time periods and 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 vectors maintain semantic consistency.

[0058] In the embodiments of the present application, first, the system constructs an interpolation path topology graph based on the spatio-temporal distribution map, and extracts the set of evolution trend parameters corresponding to each semantic unit through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool. Next, for the problems within the repairable offset interval, the system analyzes the position coordinates of the semantic focus vector of the problem in the interpolation path topology graph, and calculates the semantic continuity index between adjacent vectors according to the evolution trend parameters, and filters out the 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 an interpolation ratio coefficient through the semantic density distribution and evolution direction parameters between the nodes in the interpolation path topology graph. Finally, the system performs a constrained linear interpolation operation in the vector space to generate alternative candidate problems and attach dynamic reorganization identifiers to them for subsequent verification and optimization.

[0059] The following is a specific embodiment: In an intelligent customer service scenario, the user asks about the warranty policy of a certain smartphone. First, the system decomposes the received query text into multi-granularity semantic units and identifies key phrases such as "smartphone" and "warranty policy". Then, the system constructs a context fingerprint trajectory archive for each semantic unit and generates a historical semantic focus vector library. Through sliding window aggregation analysis, the system generates the spatio-temporal distribution map of these semantic units, showing their changing trends over time and context.

[0060] Next, the system generates multiple candidate problems, such as: "How long is the warranty period of this smartphone?", "What proof materials need to be provided during the warranty period?", etc. The system calculates the trajectory offset index of each problem and combines the evolution trend parameters of the corresponding semantic unit in the spatio-temporal distribution map to evaluate whether the problem belongs to the repairable offset interval. Suppose the problem "Can the battery be replaced during the warranty period?" is determined to be a problem within the repairable offset interval. The system constructs an interpolation path topology graph based on the spatio-temporal distribution map, and extracts the set of evolution trend parameters corresponding to each semantic unit through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool. Then, the system analyzes the position coordinates of the semantic focus vector of the problem "Can the battery be replaced during the warranty period?" in the interpolation path topology graph, and calculates the semantic continuity index between adjacent vectors according to the evolution trend parameters. After filtering out the candidate interpolation vector groups that meet the path coherence constraint, the system constructs a dynamic interpolation weight matrix based on these candidate interpolation vector groups.

[0061] The system generates an interpolation ratio coefficient through the semantic density distribution and evolution direction parameters between the nodes in the interpolation path topology graph, and performs a constrained linear interpolation operation in the vector space to generate an alternative candidate problem: "How to handle battery failure during the warranty period?" The system attaches a dynamic reorganization identifier to this alternative candidate problem for subsequent verification of its stability.

[0062] Subsequently, the system mixes and sorts the generated alternative candidate questions with the original candidate questions that have not been eliminated, and performs stability verification. If it is found that the decline in the trajectory offset index of some alternative candidate questions does not reach the preset repair gain threshold, the interpolation weight coefficient is adjusted to regenerate more accurate alternative candidate questions. Finally, the system outputs a set of optimized core question sets, significantly improving the accuracy and user experience of the question-and-answer system, enabling users to obtain answers that better meet their needs.

[0063] To further improve the stability and convergence efficiency of semantic focus drift detection, in some embodiments, in step 103, the trajectory offset index is recalculated based on the recombined candidate question set. When the trajectory offset index converges within a preset stability threshold range, a core question set is generated; otherwise, semantic focus vector interpolation and trajectory offset verification are iteratively executed, and it further includes: Construct a dynamic convergence determination model for the trajectory offset index. By using a sliding window to statistically analyze the trajectory offset index in the historical recombination process, stability baseline parameters and confidence intervals are generated; the recombined candidate question set is input into the dynamic convergence determination model to calculate the convergence difference degree between the trajectory offset index and the stability baseline parameters. If the convergence difference degree meets the preset stability threshold range, the iterative termination condition is triggered; when the convergence difference degree exceeds the preset stability threshold range, the weight coefficient of semantic focus vector interpolation is adjusted according to the gradient direction of the convergence difference degree, and an incremental candidate question set is generated based on the semantic focus vectors in the context fingerprint trajectory file that have not been utilized; perform cross-round trajectory offset verification on the incremental candidate question set, extract the trajectory offset fluctuation parameters generated during the verification process, and feedback the trajectory offset fluctuation parameters to the dynamic convergence determination model to update the stability baseline parameters; if the continuous iteration rounds reach the preset maximum recombination depth threshold, the iterative termination condition is forcibly triggered, and a core question set carrying the convergence difference degree parameter and the incremental recombination identifier is output.

[0064] In this embodiment, the dynamic convergence determination model is a model used to evaluate whether the trajectory offset index reaches a stable state. It statistically analyzes the trajectory offset index in the historical recombination process through a sliding window technique to generate stability baseline parameters and confidence intervals. These parameters are used to measure the difference between the current trajectory offset index and the expected stable state. The convergence difference degree represents the gap between the current trajectory offset index and the stability baseline parameters, and the incremental candidate question set is a new question set generated from the semantic focus vectors in the context fingerprint trajectory file that have not been utilized.

[0065] In the embodiments of the present application, first, the system constructs a dynamic convergence determination model for the trajectory offset index, and generates stability baseline parameters and confidence intervals by statistically analyzing the trajectory offset index in the historical recombination process through a sliding window. Next, the system inputs the set of candidate problems after recombination into the dynamic convergence determination model, and calculates the convergence difference degree between the trajectory offset index and the stability baseline parameters. If the convergence difference degree meets the preset stability threshold range, the iteration termination condition is triggered and the set of core problems is output. If it exceeds the threshold range, the weight coefficient of semantic focus vector interpolation is adjusted according to the gradient direction of the convergence difference degree, and an incremental candidate problem set is generated based on the unused semantic focus vectors in the context fingerprint trajectory archive. Then, cross-round trajectory offset verification is performed on the incremental candidate problem set, the trajectory offset fluctuation parameters generated during the verification process are extracted, and fed back to the dynamic convergence determination model to update the stability baseline parameters. If the continuous iteration rounds reach the maximum recombination depth threshold, the iteration is forced to terminate and the final set of core problems is output.

[0066] The following is a specific embodiment: In an intelligent customer service scenario, a user asks about the return policy of a certain electronic product. First, the system performs multi-granularity semantic unit decomposition on the received query text, and identifies key phrases such as "electronic product" and "return policy". Then, the system constructs a context fingerprint trajectory archive for each semantic unit and generates a historical semantic focus vector library. Through sliding window aggregation analysis, the system generates a spatio-temporal distribution map of these semantic units, showing their changing trends over time and context.

[0067] Next, the system generates multiple candidate problems, such as: "What is the return period of this electronic product?", "What supporting documents are required for returns?", etc. The system calculates the trajectory offset index for each problem, and combines the evolution trend parameters of the corresponding semantic unit in the spatio-temporal distribution map to evaluate whether the problem belongs to the repairable offset interval. Suppose the problem "What documents are required for returns?" is determined to be a problem within the repairable offset interval. The system constructs an interpolation path topology map based on the spatio-temporal distribution map, and extracts the set of evolution trend parameters corresponding to each semantic unit through sliding window aggregation analysis to generate a multi-level interpolation path candidate pool. Then, the system analyzes the position coordinates of the semantic focus vector of the problem "What documents are required for returns?" in the interpolation path topology map, and calculates the semantic continuity index between adjacent vectors according to the evolution trend parameters, and filters out the 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 an interpolation ratio coefficient through the semantic density distribution and evolution direction parameters between the nodes in the interpolation path topology map, and generates an alternative candidate problem: "What documents do I need to prepare for returns?" The system constructs a dynamic convergence determination model for the trajectory offset metric, and generates stability baseline parameters and confidence intervals by statistically calculating the trajectory offset metric during the historical recombination process through a sliding window. Next, the system inputs the set of candidate problems after recombination into the dynamic convergence determination model to calculate the convergence difference degree between the trajectory offset metric and the stability baseline parameters. Assuming that the convergence difference degree meets the preset stability threshold range, the iteration termination condition is triggered and the set of core problems is output.

[0068] However, if the convergence difference degree exceeds the preset stability threshold range, the system adjusts the weight coefficient of the semantic focus vector interpolation according to the gradient direction of the convergence difference degree, and generates an incremental set of candidate problems based on the unused semantic focus vectors in the context fingerprint trajectory archive, such as: "What forms need to be filled out when returning a product?" The system performs cross-round trajectory offset verification on the incremental set of candidate problems, extracts the trajectory offset fluctuation parameters generated during the verification process, and feeds them back to the dynamic convergence determination model to update the stability baseline parameters. If the maximum recombination depth threshold is reached for consecutive iteration rounds, the system forcibly terminates the iteration and outputs the final set of core problems, which includes the convergence difference degree parameter and the incremental recombination identifier, ensuring that the finally generated problem set is more accurate and meets the user's needs. In this way, the intelligent customer service system can handle complex queries more effectively, improving the user experience and service quality.

[0069] In order to further improve the accuracy and logical integrity of the preliminary answer set, in some embodiments, generating the preliminary answer set based on the set of core problems in step 104 enhances the answer logic constraints through cross-modal association, generates an enhanced answer set, and extracts the semantic divergence metric, including: Construct a multi-modal semantic space mapping network, convert each problem in the set of core problems into a semantic query vector, and retrieve a set of candidate answer fragments in the multi-modal data source of the target knowledge domain through a distributed inference engine; perform a semantic alignment operation on the set of candidate answer fragments, calculate the semantic matching metric between the semantic query vector and the set of candidate answer fragments through a cross-modal association matrix, and filter out the set of answer fragments that meet the preliminary logical constraints; generate the logical enhancement weight coefficient of the set of answer fragments based on the cross-modal association matrix, convert the set of answer fragments into a preliminary answer set through a multi-modal semantic fusion algorithm, and extract the semantic divergence metric of the preliminary answer set.

[0070] In this embodiment, the multimodal semantic space mapping network is a technology for mapping various types of data such as text, images, and audio into a unified semantic space. The semantic query vector is the representation form of the core problem in this space after transformation, and is used for subsequent retrieval and matching. The distributed inference 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 different modal data, so as to ensure 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 preliminary answer set.

[0071] In the embodiment of the present application, first, the system constructs a multimodal semantic space mapping network and converts each problem in the core problem set into a semantic query vector. Then, the distributed inference engine is used to retrieve a set of candidate answer fragments from multimodal data sources in the target knowledge domain. Next, a semantic alignment operation is performed on the set of candidate answer fragments, and the cross-modal association matrix is used to calculate the semantic matching degree index between the semantic query vector and the set of candidate answer fragments, and the set of answer fragments that meet the preliminary logical constraints is filtered out. Then, based on the cross-modal association matrix, the logical enhancement weight coefficient of the set of answer fragments is generated, 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.

[0072] The following is a specific example: In an intelligent customer service scenario, the user asks about the warranty policy of a certain smartphone. The system first processes the received core problem set and identifies key phrases such as "smartphone" and "warranty policy". The system constructs a multimodal semantic space mapping network and converts each core problem (for example, "How long is the warranty period of this smartphone?") into a semantic query vector. This step ensures that the problem can be accurately represented in the unified semantic space. Then, the system uses the distributed inference engine to retrieve a relevant set of 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", "Purchase vouchers need to be provided", and a screenshot showing the warranty terms.

[0073] The system performs semantic alignment operations on the set of candidate answer segments, and calculates the semantic matching degree index between the semantic query vector and the set of candidate answer segments through a cross-modal correlation matrix. Suppose the system discovers that the segment "The warranty period is one year" highly matches the original question, while other segments also provide useful but not completely directly relevant information. Based on the cross-modal correlation matrix, logical enhancement weight coefficients for the set of answer segments are generated. For example, the weight of "The warranty period is one year" is higher because it is the key information directly answering the user's question; while "Purchase vouchers need to be provided" is important but logically secondary information. The system combines these answer segments into a preliminary answer set through a multi-modal semantic fusion algorithm, such as: "The warranty period of this smartphone is one year. You need to provide purchase vouchers to enjoy the warranty service." Finally, the system extracts the semantic divergence index of the preliminary answer set to evaluate its diversity and coverage. Suppose the system discovers that the current answer set already covers the main information points, but to further improve 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, etc.

[0074] To further improve the logical consistency and semantic coherence of the Q&A pair combinations, in some embodiments, in step 105, a semantic focus trajectory archive is established based on the context association tags, a multi-dimensional topological mapping is performed between the set of core questions and the enhanced answer set, and driven by the trajectory archive for aggregation, and target Q&A pair combinations are screened according to the connection strength, including: Construct a multi-dimensional topological structure of the semantic focus trajectory archive, map the set of semantic units, the set of core questions, and the enhanced answer set into a set of nodes in the topological graph through the context association tags, and generate an initial connection strength weight between the nodes based on the semantic divergence index and the logical consistency scoring parameter; perform semantic focus trajectory parsing on the set of nodes, extract the context fingerprint trajectory encoding of each node, and generate a semantic association degree matrix between the nodes through the similarity calculation of the context fingerprint trajectory encoding; construct a dynamic connection strength calculation model based on the semantic association degree matrix and the initial connection strength weight, and optimize the connection strength weight between the nodes through multiple rounds of iteration to generate a topological connection graph in a stable state; perform clustering analysis on the topological connection graph, divide the set of nodes into multiple semantic focus clusters according to the connection strength threshold, where each cluster contains strongly associated combinations of question nodes and answer nodes; perform logical consistency verification on the Q&A pair combinations within the semantic focus clusters, extract the logical conflict parameters generated during the verification process, feedback the logical conflict parameters to the dynamic connection strength calculation model to update the connection strength weight, and screen out the target Q&A pair combinations from the updated connection strength weight, where the target Q&A pair combinations contain optimized values of the logical consistency scoring parameter and the semantic divergence index.

[0075] In this embodiment, the semantic focus trajectory file is a time-series file that records semantic units and their context information, and is mapped to nodes in a multi-dimensional topological structure through context association tags. The initial connection strength weight between nodes is generated based on the semantic divergence index and the logical consistency scoring parameter, and is used to measure the relationship strength between nodes. The semantic association degree matrix quantifies the semantic relevance between nodes by calculating the similarity of the context fingerprint trajectory codes between nodes. The dynamic connection strength calculation model is a model for optimizing the connection strength weight between nodes, ensuring that the finally generated Q&A pair combinations have high logical consistency and semantic coherence.

[0076] In the embodiment of the present application, first, the system constructs a multi-dimensional topological structure of the semantic focus trajectory file, maps the semantic unit set, the core question set, and the enhanced answer set to a node set in the topological graph, and generates the initial connection strength weight between 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 code of each node, and generates a semantic association degree matrix between nodes through similarity calculation. Based on the semantic association degree matrix and the initial connection strength weight, the system constructs a dynamic connection strength calculation model, and optimizes the connection strength weight between nodes through multiple rounds of iteration to generate a topological connection graph in a stable state. Then, the system performs clustering analysis on the topological connection graph, and divides the node set into multiple semantic focus clusters according to the connection strength threshold. Each cluster contains a strongly associated combination of question nodes and answer nodes. Finally, the system performs logical consistency verification on the Q&A pair combinations within the semantic focus cluster, extracts the logical conflict parameters generated during the verification process, and feeds them back to the dynamic connection strength calculation model to update the connection strength weight, and finally outputs the target Q&A pair combination.

[0077] The following is a specific embodiment: In an intelligent customer service scenario, the user asks about the warranty policy of a certain smartphone. The system first processes the received core question set (such as "How long is the warranty period of this smartphone?") and the enhanced answer set (such as "The warranty period is one year", "Purchase vouchers need to be provided", etc.).

[0078] The system constructs a multi-dimensional topological structure of a semantic focus trajectory archive, mapping a set of semantic units (e.g., "smartphone", "warranty policy"), a set of core questions (such as "How long is the warranty period of this smartphone?"), and a set of enhanced answers (such as "The warranty period is one year", "Purchase vouchers need to be provided") to a set of nodes in the topological graph. Based on the semantic divergence index and the logical consistency scoring parameter, the initial connection strength weights between nodes are generated. For example, the connection strength between the node of "The warranty period is one year" and other nodes is relatively high because it directly answers the user's question. The system performs semantic focus trajectory parsing on the set of nodes, extracts the context fingerprint trajectory encoding of each node, and generates a semantic association matrix between nodes through similarity calculation. Suppose the system discovers a strong semantic association between the two nodes of "The warranty period is one year" and "Purchase vouchers need to be provided" because they both involve specific details of the warranty policy.

[0079] Based on the semantic association matrix and the initial connection strength weights, the system constructs a dynamic connection strength calculation model, and optimizes the connection strength weights between nodes through multiple rounds of iteration to generate a topological connection graph in a stable state. For example, after multiple iterations, the system discovers that the connection strength between "The warranty period is one year" and "Purchase vouchers need to be provided" has been significantly improved, indicating that they are closely related logically. The system performs clustering analysis on the topological connection graph, and divides the set of nodes into multiple semantic focus clusters according to the connection strength threshold. Each cluster contains a strong association combination of question nodes and answer nodes. For example, the system groups "How long is the warranty period of this smartphone?" with "The warranty period is one year" and "Purchase vouchers need to be provided" into one cluster because they jointly form a complete answer.

[0080] The system performs logical consistency verification on the Q&A pair combinations within the semantic focus clusters, extracts the logical conflict parameters generated during the verification process, and feeds them back to the dynamic connection strength calculation model to update the connection strength weights. For example, if the system discovers that a certain answer segment is not completely consistent with the question logic (such as missing key information), it will adjust its connection strength weight to optimize the overall answer quality.

[0081] Finally, from the updated connection strength weights, the system filters out the target Q&A pair combinations. These Q&A pairs not only contain the logical consistency scoring parameter but also the optimized value of the semantic divergence index, ensuring that the provided answers are both comprehensive and accurate. For example, the complete answer output by the system may be: "The warranty period of this smartphone is one year, and you need to provide purchase vouchers to enjoy the warranty service." Figure 2 The following is a schematic structural diagram of a Q&A pair generation system based on a large language model and data augmentation provided by an embodiment of the present application, as Figure 2 shown, the system includes: A receiving module 21, configured to receive an original text input in a target knowledge domain, decompose the original text input into multi-granularity semantic units, generate a set of semantic units with context fingerprints, and bind adjacent context association tags associated with each semantic unit in the set of semantic units; An input module 22, configured to input the set of semantic units into a preset multi-intent question-answer generation framework, generate a set of candidate questions through multi-round intent iteration, and dynamically adjust the question generation direction by triggering a probability masking strategy each time; A matching module 23, configured to perform semantic focus drift detection on the set of candidate questions, perform two-way trajectory matching between the questions and the context fingerprints of the set of semantic units through an adaptive reconstruction mechanism, and dynamically trigger question recombination according to a focus displacement threshold to generate a set of core questions; A generation module 24, configured to generate a preliminary answer set based on the set of core questions, enhance the answer logic constraint through cross-modal association, generate an enhanced answer set and extract a semantic divergence index; A building module 25, configured to build a semantic focus trajectory file based on the context association tags, perform multi-dimensional topological mapping on the set of core questions and the enhanced answer set, drive aggregation through the trajectory file, and screen out target question-answer pair combinations according to the connection strength; An output module 26, configured to output a sequence of question-answer pairs including a context fingerprint traceability chain, where each question-answer pair integrates the semantic focus trajectory file, the semantic divergence index, and the strength factor of the cross-modal association to generate a multi-level knowledge topology identifier.

[0082] Figure 2 The described question-answer pair generation system based on a large language model and data augmentation can execute Figure 1 The described question-answer pair generation method based on a large language model and data augmentation in the illustrated embodiment, and its implementation principle and technical effects will not be elaborated further. For the question-answer pair generation system based on a large language model and data augmentation in the above embodiment, the specific ways in which each module and unit perform operations have been described in detail in the embodiment related to the method, and will not be elaborated here.

[0083] In a possible design, Figure 2 The xx system in the illustrated embodiment can 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; The storage component 31 stores one or more computer instructions, where the one or more computer instructions are called and executed by the processing component 32.

[0084] The processing component 32 above Figure 1A method for generating question-answer pairs based on large language models and data augmentation in the described embodiment.

[0085] Among them, 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 can also be implemented by 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 for executing the above method.

[0086] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage 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 disks or optical discs.

[0087] Of course, the computing device may also necessarily include other components, such as input / output interfaces, display components, communication components, etc.

[0088] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above peripheral interface module can be an output device, an input device, etc.

[0089] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0090] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device can refer to a cloud server, and the above processing component, storage component, etc. can be basic server resources leased or purchased from a cloud computing platform.

[0091] The embodiment of the present application also provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the above Figure 1 A method for generating question-answer pairs based on large language models and data augmentation in the shown embodiment.

[0092] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative work.

[0094] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part 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, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment 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; The semantic unit set is input into a preset multi-intent question-answering generation framework, and a candidate question set is generated through multiple rounds of intent iterations. Each iteration triggers a probability mask strategy to dynamically adjust the question generation direction; Performing semantic focus drift detection on the candidate question set, bidirectionally matching the question with the context fingerprint of the semantic unit set through an adaptive reconstruction mechanism, dynamically triggering question reorganization according to a focus displacement threshold, and generating a core question set; Generate a preliminary answer set based on the core question set, enhance answer logic constraints through cross-modal association, generate an enhanced answer set and extract semantic divergence indicators; 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 according to connectivity strength; The 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 identifier.

2. The method according to claim 1, characterized in that The method 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, dynamically triggering question reorganization according to a focus displacement threshold, and generating a core question set includes: Constructing a context fingerprint trajectory file of 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 a latent semantic focus vector of the question, and dynamically aligning the latent semantic focus vector with the multidimensional semantic focus vector in a bidirectional manner, and calculating a trajectory offset index between each pair of semantic units of the question; According to a preset displacement threshold interval, the trajectory offset index is divided into a repairable offset interval and an unrepairable offset interval, an adaptive reorganization rule is triggered for the problem in the repairable offset interval, an alternative candidate problem is generated through a semantic focus vector interpolation technology, an elimination marking operation is performed on the problem in the unrepairable offset interval, and the alternative candidate problem is 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.

3. The method according to claim 2, 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 in the repairable offset interval, generates alternative candidate problems through semantic focus vector interpolation technology, performs an elimination marking operation on the problems in 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, construct a historical semantic focus vector library based on the context association mark, and perform sliding window aggregation analysis 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 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, 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 in the vector space is performed to generate an alternative candidate problem, and a reorganization identifier is attached 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.

4. The method according to claim 3, characterized in that For the problem in 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 attaching 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; Analyze the position coordinates of the semantic focus vector in the interpolation path topology map for the problem in the repairable offset interval, calculate the semantic continuity index between adjacent vectors according to the evolution trend parameter, and select 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.

5. The method according to claim 4, characterized in that The recalculating the trajectory offset index based on the reorganized candidate question set, generating a core question set when the trajectory offset index converges to a preset stability threshold range, and iteratively performing semantic focus vector interpolation and trajectory offset verification otherwise, includes: A dynamic convergence judgment model of the trajectory deviation index is constructed, and the trajectory deviation index in the historical reorganization process is statistically analyzed through a sliding window to generate a stability baseline parameter and a confidence interval; 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 back the trajectory offset fluctuation parameters to the dynamic convergence determination model to update the stability baseline parameters; If the consecutive iteration rounds reach the preset maximum reorganization depth threshold, the iteration termination condition is triggered forcibly, and a set of core questions carrying convergence difference parameters and incremental reorganization identifiers are output.

6. The method according to claim 1, characterized in that 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, converting each question in the core question set into a semantic query vector, and retrieving a set of candidate answer fragments in a multimodal data source in the target knowledge domain through a distributed reasoning engine; Performing a semantic alignment operation on the candidate answer fragment set, calculating a semantic matching index between the semantic query vector and the candidate answer fragment set through a cross-modal association matrix, and screening an answer fragment set that meets 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.

7. The method according to claim 1, characterized in that The step of establishing a semantic focus trajectory archive based on the context 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 according to connectivity strength includes: Constructing a multi-dimensional topological structure of the 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 context fingerprint trajectory coding of each node, and generating a semantic correlation matrix between nodes by calculating the similarity of the context fingerprint trajectory coding; A dynamic connectivity strength calculation model is constructed based on the semantic relevance matrix and the initial connectivity strength weight, 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, dividing 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 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 optimized values ​​of the logical consistency scoring parameter and the semantic divergence index.

8. A question-answer pair generation system based on a large language model and data enhancement, characterized in that: include: A receiving module is used 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 candidate question set through multiple rounds of intent iterations, and trigger a probability mask strategy to dynamically adjust the question generation direction in each iteration; A matching module is used 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 according to a focus displacement threshold, and generate a core question set; A generation module, used 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; An establishment module is used to establish a semantic focus trajectory archive based on the context 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 the target question-answer pair combination according to the connectivity strength; 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 identifier.

9. 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 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a question-answer pair generation method based on a large language model and data enhancement is implemented as described in any one of claims 1 to 7.

Images