Data augmentation generation method and system

Abstract

The application provides a data enhancement generation method and system. The method comprises the following steps: in response to an initial requirement, generating an initial draft, and inversely extracting feature information from the initial draft to generate retrieval basis information; based on the retrieval basis information, performing accurate matching retrieval on a preset memory cache library to obtain a target text slice as a context; generating a target text by comprehensively combining the context and the initial draft; performing requirement verification on the target text to obtain a verification result; when the verification result indicates that the target text has a missing feature vector, adding the missing feature vector to the initial draft, and re-performing the inverse extraction of feature information and the accurate matching retrieval until the verification result indicates that there is no missing feature vector, and outputting a final text. The application realizes accurate recall of structured data and dynamic closed-loop completion of generated content, and solves the problems of context missing and iterative dead loop caused by initial retrieval completion.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to data augmentation generation methods and systems. Background Technology

[0002] Existing Retrieval-Augmented Generation (RAG) schemes typically transform user input into semantic vectors, perform similarity searches in a vector database, and then use the search results as context input to a large model for single-step generation. However, this approach has the following drawbacks: First, vector retrieval suffers from precision loss in structured fields (such as table names and numeric IDs), leading to data illusions in the large model; second, it relies on the initial user input for forward retrieval, making it impossible to predict the underlying data tables to be called during the generation process, resulting in incomplete context; third, it lacks a closed-loop error correction mechanism, meaning that when generated content is missing, it can only be rewritten within the existing context and cannot trigger a re-retrieval by the underlying data engine. Summary of the Invention The technical problem to be solved by this invention is to provide a data augmentation generation method and system that can achieve accurate retrieval of structured data and dynamic completion of generated content.

[0003] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A data augmentation generation method, characterized by comprising: In response to the initial requirements, an initial draft is generated, and feature information is extracted from the initial draft in reverse to generate retrieval basis information; Based on the search criteria information, an exact match search is performed on the preset memory cache library to obtain the target text slice as context; The target text is generated by combining the aforementioned context with the initial draft; The target text is subjected to requirement verification to obtain a verification result. When the verification result indicates that the target text has a missing feature vector, the missing feature vector is added to the initial draft, and the reverse feature information extraction and the exact matching retrieval are re-executed until the verification result indicates that the missing feature vector does not exist, and the final text is output.

[0004] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A data augmentation generation system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the various steps of the aforementioned data augmentation generation method.

[0005] The beneficial effects of this invention are as follows: By extracting feature words in reverse to generate retrieval basis information, the traditional forward retrieval method is changed, making feature words more accurately reflect actual data needs and solving the problem of missing context caused by incomplete initial retrieval; based on the retrieval basis information, a precise retrieval is performed on the memory cache library, avoiding feature ambiguity in the vectorization process of structured data and achieving precise recall of structured fields; after generating target text by combining context and initial draft, requirement verification is performed to obtain verification results. When the verification results show that the target text has missing feature vectors, the missing feature vectors are added to the initial draft and reverse extraction and precise retrieval are re-executed, forming a closed-loop feedback mechanism. This mechanism enables the system to autonomously trigger a re-retrieval of the underlying cache library to obtain missing knowledge when it finds omissions in the generated content, rather than simply rewriting the language in the existing context. This effectively solves the iterative dead loop problem caused by the lack of a closed-loop error correction mechanism, and achieves dynamic completion and quality improvement of the generated content. Attached Figure Description

[0006] Figure 1 This is a flowchart of a data augmentation generation method according to an embodiment of the present invention; Figure 2 This is a flowchart of the dynamic re-retrieval closed loop in a data augmentation generation method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a data augmentation generation system according to an embodiment of the present invention. Detailed Implementation

[0007] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0008] In existing technologies, retrieval-enhanced generation schemes convert user input into semantic vectors, perform similarity retrieval, and then input the vectors into a large model to generate text. However, vector retrieval suffers from feature ambiguity in structured fields, failing to accurately recall data and causing data illusion; forward retrieval, which relies on the initial input, cannot predict the required data table, resulting in incomplete context; and it lacks a closed-loop error correction mechanism, meaning that when content is missing, it can only be rewritten instead of being re-retrieved, leading to an iterative dead loop.

[0009] To at least address the aforementioned issues, an initial draft is generated, and feature information is extracted from it using reverse engineering as the retrieval basis. Based on this retrieval basis, relevant text slices are precisely retrieved from the memory cache as context. The target text is then generated by combining the context with the initial draft. The target text undergoes requirement validation; if missing features are found, they are added to the initial draft, and the reverse extraction and precise retrieval are re-executed. This process is iterated until validation passes, at which point the final text is output. This approach enables accurate retrieval of structured data and dynamic closed-loop completion of generated content, avoiding data feature loss caused by vector retrieval and resolving the issues of missing context and iterative loops due to incomplete initial retrieval.

[0010] The following describes a data augmentation generation method of the present invention in detail. Please refer to [link / reference]. Figure 1 The method 100 includes steps 110 to 140.

[0011] Step 110: In response to the initial requirements, generate an initial draft, and extract feature information from the initial draft in reverse to generate retrieval basis information.

[0012] For example, the system receives initial requirements from the user, generates an initial draft, and then extracts proper nouns as feature information from the generated draft to generate search criteria. The term "reverse engineering" refers to a process different from traditional methods that directly extract keywords from the user's initial requirements. This method first allows a large model to generate a preliminary draft, and then the large model, based on the generated draft, considers "what data I need" to extract feature information. This approach ensures that the extracted feature information more closely reflects the actual data requirements during the generation process, rather than relying solely on the limited information input by the user.

[0013] Step 120: Based on the retrieval criteria information, perform an exact match retrieval on the preset memory cache library to obtain the target text slice as context.

[0014] For example, based on the feature words in the retrieval criteria information, the system traverses each text slice in the memory cache, calculates the frequency of the feature words in the slice, and uses the top K slices with the highest frequency as the context. Step 130: Generate the target text by combining the context and the initial draft.

[0015] For example, the system concatenates the target text slices recalled in step 120 with the initial draft generated in step 110, inputs them into the large language model, and generates the target text.

[0016] Step 140: Perform requirement verification on the target text to obtain the verification result. When the verification result indicates that the target text has missing feature vectors, add the missing feature vectors to the initial draft and re-execute the reverse feature information extraction and exact matching retrieval until the verification result indicates that there are no missing feature vectors, and output the final text.

[0017] For example, the system calls the verification model to verify the target text. If missing features are found, a missing feature vector is generated and added to the end of the initial draft. The updated draft is then executed again from step 110 to step 130 until the verification passes and the final text is output.

[0018] As described above, by extracting feature information in reverse to generate retrieval criteria, the traditional forward retrieval method is changed, making feature words more accurately reflect actual data needs and solving the problem of missing context caused by incomplete initial retrieval. Based on the retrieval criteria, a precise retrieval is performed on the memory cache library, avoiding feature ambiguity in the vectorization process of structured data and achieving accurate recall of structured fields. Combined with the requirement verification and closed-loop feedback mechanism, the system can autonomously trigger a re-retrieval to obtain missing knowledge when content is found to be missing, rather than simply rewriting the language in the existing context. This effectively solves the problem of iterative dead loop and achieves dynamic completion and quality improvement of generated content.

[0019] In one alternative implementation, step 121 is included before step 120 performs an exact match retrieval on the preset memory cache library.

[0020] Step 121: Read the preset structured data source, slice the structured data source according to the preset row number threshold to obtain text slices, inject metadata for precise matching and retrieval into each text slice, serialize all text slices into plain text format, and load them into memory to generate a memory cache library.

[0021] For example, the system reads a structured data table in Excel or CSV format, sets each 100 rows as a slice, and injects the source file name, form name, and column name header of the slice as metadata into the header of each slice. All slices after the metadata injection are converted into plain text format and loaded into the system memory to form a memory cache library that can be accurately retrieved later.

[0022] As described above, by slicing the two-dimensional structured data table according to row thresholds, the table data is reduced to one-dimensional text slices, effectively avoiding the token overflow problem during single processing. Metadata containing the source file name, form name, and column header is injected into each slice, providing source traceability capability for subsequent retrieval and location. The entire data is loaded into memory in plain text format, abandoning traditional vector databases and inverted indexes, achieving lossless storage and extremely fast traversal retrieval of structured data, laying the data foundation for subsequent precise matching retrieval.

[0023] In an alternative implementation, step 121 injects metadata for precise matching retrieval location into each text slice, including step 122.

[0024] Step 122: Inject metadata into each text slice, serialize the text slices after metadata injection into a semi-structured data format, and store them in the memory cache library.

[0025] For example, the system generates corresponding metadata for each text slice, which includes the source file name, form name, and column header. The metadata and slice content are serialized together into semi-structured data in JSON format and stored in a memory cache library for locating the source of the matching slice during subsequent precise matching retrieval.

[0026] As described above, by injecting metadata into each text slice, each slice carries independent source identification information. The slices and metadata are co-serialized into a semi-structured data format for storage. This maintains the parsability of the data structure and provides a location basis for subsequent accurate matching and retrieval. When a search matches a slice, the system can quickly trace the source data table and field information to which the slice belongs through its metadata. This allows for accurate reference to the data source when generating target text, effectively avoiding problems such as incorrect data referencing or unclear sources.

[0027] In one alternative implementation, step 110 reverse-engineers feature information from the initial draft to generate retrieval basis information, including steps 111 and 112.

[0028] Step 111: Input the initial draft into the large language model, and constrain the large language model to extract feature information based on the content of the initial draft to obtain the extraction result.

[0029] For example, the system takes the generated initial draft as input, calls the large language model and applies constraint instructions, requiring the large language model to extract only proper nouns appearing in the draft, such as prop names, table names, field names, system names, etc., ignoring descriptive and modifying general words, and outputs preliminary feature information extraction results.

[0030] Step 112: Filter out common pronouns in the extraction results and perform deduplication logic on the extraction results to generate a set of search keys as the basis for retrieval.

[0031] For example, the system performs secondary filtering on the initial extraction results to remove any possible residual generic pronouns, such as "this," "that," "it," and "they." Then, it performs deduplication on the remaining feature information, merging the same feature words that appear multiple times into one, and finally generating a set of unique search keys, which serves as the search basis information needed for subsequent precise matching searches.

[0032] As described above, by first generating an initial draft and then extracting feature information from it in reverse, the traditional forward retrieval method, which relies on direct keyword extraction from initial user input, is changed. This makes the extracted feature words closer to the actual data requirements during the generation process. Simultaneously, by constraining the large language model to extract only proper nouns, filtering general pronouns, and performing deduplication logic, noise information is effectively eliminated, ensuring the conciseness and accuracy of the retrieval criteria. This provides high-quality query conditions for subsequent exact matching retrieval, avoiding low retrieval efficiency or unexpected recall results due to redundant or repetitive retrieval criteria.

[0033] In one alternative implementation, step 120 includes 123.

[0034] Step 123: Calculate the frequency of occurrence of the retrieval information in each text slice in the memory cache library, obtain the score of the text slice based on the frequency of occurrence, and use the preset number of target text slices with the highest scores as context.

[0035] For example, the system uses the generated set of search keys as query conditions, traverses each text slice in the memory cache, counts the frequency of each search key in the slice, calculates the comprehensive score of each slice based on the frequency of each search key and its preset weight, sorts all slices from high to low score, and extracts the top K slices with the highest scores (e.g., K=15) as the target text slices, which are then output as context. As described above, a precise retrieval mechanism based on word frequency statistics replaces traditional semantic vector retrieval. By calculating and quantifying the frequency of retrieval information in text slices, hard matching of structured fields (such as table names and numerical IDs) is achieved, avoiding feature loss due to semantic ambiguity during vectorization. Simultaneously, a strategy of ranking by score and extracting a predetermined number of highest-scoring slices ensures that the recalled context contains the most relevant data while controlling the context length of the input large model, thus preventing token overflow. Compared to existing solutions, this approach can improve the recall accuracy of structured data to nearly 100%, fundamentally suppressing data illusions caused by inaccurate retrieval in large models.

[0036] In one alternative implementation, step 123 obtains a score for the text slice based on its frequency of occurrence, including step 1231.

[0037] Steps 1, 2, 3, 1

[0038] in, The score is given to the text slice, where n is the total number of feature information in the retrieval criteria. Let be the frequency of the i-th feature information in text slice j. The weights are the feature information.

[0039] For example, suppose the search criteria contain three keywords: "rusty gears," "abandoned giant ship," and "teleportation point," each with a weight of 1. If "rusty gears" appears twice, "abandoned giant ship" appears once, and "teleportation point" appears zero times in a certain text slice, then the score for that slice is 2×1 + 1×1 + 0×1 = 3 points. After the system calculates the scores for all slices sequentially, it sorts them in descending order of score and selects the slices with the highest scores as the search results.

[0040] As described above, the degree of matching between search criteria information and text segments is quantified into a comparable numerical score using a word frequency statistical formula. The formula contains... This reflects the hit frequency of a single feature word in the slice. The system allows for weighted adjustments based on the importance of different feature words (e.g., core table names can be assigned higher weights), and the final summation yields a comprehensive score for each slice. This scoring mechanism is simple and efficient, requiring no complex vector calculations, and is particularly suitable for retrieval tasks in structured data scenarios where precise matching is crucial. By setting feature weights, the system can flexibly adjust retrieval preferences according to actual business needs, further improving the accuracy and adaptability of the retrieval.

[0041] In one optional implementation, step 140 performs requirement verification on the target text to obtain a verification result, including step 141.

[0042] Step 141: Decompose the initial requirements into multiple requirement sub-items, call the verification model to check whether the target text covers all requirement sub-items one by one, and output the quantitative score and missing feature vector based on the verification results of the verification model.

[0043] For example, the system breaks down the initial user-input requirements into several indivisible atomic requirements, such as "needs to include item exchange logic," "needs to include map teleportation point configuration," and "needs to include character experience point calculation." It then calls an independent large-scale validation language model to verify whether the generated target text covers each of these requirement sub-items. After validation, the system outputs a quantitative score of 0-100 (e.g., 75 points) and a missing feature vector recording the missing requirement sub-items, such as "no map teleportation point configuration found."

[0044] Step 140 adds the missing feature vectors to the initial draft, including step 142.

[0045] Step 142: Concatenate the missing feature vectors into the initial draft using the specified identifier format.

[0046] For example, the system adds a specific identifier prefix "[Lead Planner Review Feedback:]" to the output missing feature vector, such as "No map teleportation point configuration found", to form a formatted feedback message. Then, this feedback message is appended to the end of the original initial draft to generate a new draft carrying the missing information, which is used to trigger the reverse extraction and precise retrieval in the next iteration.

[0047] As described above, by breaking down the initial requirements into multiple sub-requirements, fine-grained verification of the generated content is achieved, rather than simply evaluating overall quality. Calling independent verification models for one-by-one checks ensures the objectivity and fairness of the verification process, avoiding self-verification bias that might arise from using the same model for both generation and verification. The output quantitative score intuitively reflects the completeness of the generated content, while missing feature vectors accurately pinpoint key information that needs to be supplemented. Concatenating the missing feature vectors to the initial draft using a specified identifier format forms structured feedback information, enabling the system to seamlessly pass on identified issues to the next iteration. This provides clear supplementary targets for closed-loop re-retrieval, effectively guiding subsequent reverse extraction and data retrieval.

[0048] In an optional implementation, step 140 re-executes the reverse feature information extraction and precise matching retrieval, and further includes step 1401.

[0049] Step 1401: Merge the newly extracted feature information with all previously extracted feature information to remove duplicates, and use the merged and deduplicated feature information as the updated retrieval basis information.

[0050] For example, in the first iteration, the system extracted "rusty gears" and "abandoned giant ships" as search criteria. After verification, it was found that "teleportation point configuration" was missing. In the second iteration, "teleportation points" were newly extracted as feature information. The system merged the newly extracted "teleportation points" with the previously extracted "rusty gears" and "abandoned giant ships," removed duplicates, and generated updated search criteria information of "rusty gears," "abandoned giant ships," and "teleportation points," which served as the query conditions for the next round of exact match retrieval.

[0051] As described above, by merging newly extracted feature information with all previously extracted feature information to remove duplicates each time reverse extraction is performed, the system ensures that the retrieval basis information accumulates and expands with each iteration, rather than being overwritten or replaced by newly extracted feature information. This cumulative update mechanism allows the system to retain data requirements that have been validated in previous iterations, while supplementing newly discovered missing features, preventing the recurrence of solved problems due to the loss of historical information. The merging and deduplication operation ensures the conciseness of the retrieval basis information, eliminating duplicate feature words and reducing the redundancy of retrieval calculations. In this way, the system's retrieval scope gradually expands and improves with the increase of iterations until it covers all required features, providing a progressively approximating data guarantee for the final output of complete and accurate target text.

[0052] In an optional implementation, step 140 re-executes the reverse feature information extraction and precise matching retrieval, and also includes step 1402.

[0053] Step 1402: Determine whether the number of iterations already performed has reached the preset maximum number of iterations. If it has, stop the iteration and output the target text of the highest historical score.

[0054] For example, the system presets a maximum of 3 iterations. The target text score generated in the first iteration is 75 points, the score improves to 88 points in the second iteration, and the score remains at 90 points after the third iteration, failing to reach the preset passing threshold. When the system is about to enter the fourth iteration, it determines that it has already executed 3 iterations, reaching the maximum number of iterations, and then stops the loop, selecting the target text with the highest score of 90 points from the three rounds as the final output.

[0055] As described above, setting a maximum iteration limit provides a termination protection mechanism for the closed-loop re-retrieval loop. In practical applications, there may be extreme cases where no matter how many iterations are performed, all required features cannot be fully covered (e.g., the structured data source itself lacks certain information). In such cases, without an upper limit, the system will fall into an infinite loop. By determining whether the number of iterations has reached the preset maximum value, the system can promptly exit the loop and output the highest-scoring text obtained during the iteration process. This ensures both the finiteness and stability of the method and provides a relatively optimal generation result when perfect coverage cannot be achieved. This mechanism effectively avoids the problem of system resource exhaustion or infinite waiting due to abnormal situations, improving the robustness and practicality of the method.

[0056] Please refer to Figure 2 In one embodiment of the present invention, the present invention can be applied to the above-mentioned scheme in the scenario of structured data augmentation generation, especially in the scenario of automated document generation that requires accurate retrieval of data from relational databases and generation of structurally complete text; including S301 to S306.

[0057] S301. The system reads a two-dimensional relational data table in Excel or CSV format and slices the data according to a configurable row count threshold (e.g., N=100 rows). Metadata is injected into the header of each slice, strictly including the source file name, form name, and column header names. Subsequently, all slices are serialized into JSON format and loaded into system memory, forming a "JSON slice cache library." This abandons the concept of inverted indexes and uses full memory loading to support extremely fast traversal. This is equivalent to step 121 above.

[0058] S302. The system receives the initial requirement command input by the user and calls the generative large language model to output a preliminary text draft containing the core logic. Subsequently, the system issues a feature extraction command with strict constraints to the large model, requiring the large model to independently extract the necessary proper nouns (such as table names, field names, specific prop / product names) based on the draft content, and filter out general pronouns. After receiving the output, the system executes deduplication logic to generate a set of search keys. This is equivalent to steps 110, 111, and 112 above.

[0059] S303: The system iterates through the JSON slice cache generated in S301, using the search key set generated in S302 to perform case-insensitive exact substring matching. The score for each slice is calculated using a term frequency formula. Where n is the total number of search keys. Let be the frequency of the i-th search key in slice j. This is the weight of the search key. The system sorts the scores in descending order and extracts the top-K slices with the highest scores as the precise context, where K is a configurable hyperparameter, such as K=15. This is equivalent to steps 120, 123, and 1231 above. S304. The system concatenates the Top-K slices recalled in S303 with the initial draft generated in S302, using this as input to the generation module, and calls the large language model to output the first complete version of the target text. This is equivalent to step 130 above.

[0060] S305, The system triggers a verification module consisting of another independent large language model instance. The verification module first decomposes the initial requirement instruction into a set of atomic requirement points. The system verifies that the target text generated by S304 covers all the aforementioned atomic requirements. The verification module outputs a standardized JSON format result, including a quantitative score (an integer from 0 to 100) and a missing feature vector (e.g., ["No store redemption logic found", "Missing teleportation point coordinate configuration"]). The system determines whether the output quantitative score reaches a preset threshold (e.g., 95 points). If it does, S306 is executed to output the final enhanced document and the process ends; if it does not, the system enters a dynamic re-retrieval closed loop. This is equivalent to steps 140 and 141 above.

[0061] S306. The system appends the missing feature vector output in S305 to the end of the original draft with a specific identifier (e.g., "[Lead Strategist Review Feedback:]"), triggering a state rollback. The system then re-executes S302 and S303 with the augmented draft, extracting new search keys based on the supplementary feedback information and merging them with the old search keys to remove duplicates, and re-retrieves the previously missed data table slices. This loop has a maximum number of iterations M (e.g., M=3). Before each loop, it checks whether the number of iterations has reached M. If it has, it forcibly outputs the highest-scoring version and ends; if it has not, it continues with a new round of iterations. This is equivalent to steps 140, 1401, and 1402 above.

[0062] Through the above steps, this invention first reduces the dimensionality of relational data sources and constructs them into a plain text in-memory cache library, providing a lossless storage foundation for accurate retrieval; secondly, it receives initial requirements and extracts search keys in reverse, changing the traditional forward retrieval method; then, it achieves accurate recall through word frequency scoring, avoiding the feature ambiguity problem of vector retrieval; finally, through a verification module and a dynamic re-retrieval closed loop, it achieves autonomous completion and iterative optimization of generated content. This process effectively solves the problems of low accuracy, missing context, and iterative dead loops in structured data retrieval, achieving high-precision data augmentation generation.

[0063] Please refer to Figure 3The present invention also provides a data augmentation generation system 300, including a memory 301, a processor 302, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of a data augmentation generation method as described above.

[0064] In summary, this invention provides a data augmentation generation method and system that achieves accurate retrieval of structured data and autonomous completion of generated content through reverse feature information extraction, precise retrieval based on word frequency statistics, and a dynamic re-retrieval closed loop driven by verification feedback.

[0065] The system first reads a preset structured data source, slices the data according to a preset row count threshold, injects metadata for retrieval and location into each slice, serializes all slices into plain text format and loads them into memory, generating an in-memory cache library. This mechanism reduces the dimensionality of a two-dimensional relational data table to one-dimensional text slices, and uses full plain text loading to replace a vector database, achieving lossless storage and extremely fast traversal and retrieval of structured data.

[0066] The system generates an initial draft in response to initial requirements and then extracts feature information from it in reverse as the retrieval basis. Reverse extraction refers to first having the large language model generate a preliminary draft, and then constraining it to extract proper nouns based on the draft content, unlike traditional forward retrieval methods that directly extract keywords from the user's initial requirements. Subsequently, the system performs exact matching retrieval on the in-memory cache based on the retrieval basis, calculates the score of each text slice using a word frequency statistical formula, and uses a predetermined number of slices with the highest scores as context. This mechanism replaces semantic vector matching with word frequency statistics, avoiding the feature ambiguity problem in the vectorization process of structured data, and achieving accurate recall of structured fields such as table names and numerical IDs, fundamentally suppressing the data illusion of large models.

[0067] During the generation and verification phase, the system generates target text by combining the context and the initial draft, and then verifies the target text against requirements. The verification module breaks down the initial requirements into multiple sub-items, checks the coverage of the target text one by one, and outputs a quantitative score and missing feature vectors. When the verification result indicates the existence of missing feature vectors, the system adds the missing feature vectors to the initial draft and re-executes reverse extraction and precise retrieval. At the same time, it merges the newly extracted feature information with all previously extracted feature information to remove duplicates, forming a cumulative retrieval basis update mechanism. This closed-loop feedback mechanism enables the system to autonomously trigger a re-retrieval of the underlying memory cache library to obtain missing knowledge when content omissions are detected, rather than simply rewriting the language in the existing context. The system also has an iteration termination protection mechanism, which judges whether the number of iterations executed has reached the preset maximum number of iterations. If it has, it stops iterating and outputs the target text with the highest historical score, avoiding the infinite loop problem.

[0068] This invention solves the technical problems in related technologies, such as the loss of precision in structured data due to vector retrieval, the lack of context due to forward retrieval, and the iterative dead loop caused by the lack of a closed-loop error correction mechanism. It realizes accurate retrieval of structured data and dynamic closed-loop completion of generated content, enhances the reliability and completeness of text generation in structured data scenarios, and provides a complete technical solution for automated text generation scenarios that rely on underlying relational data tables, such as game development and financial research report generation.

[0069] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A data augmentation generation method, characterized in that, include: In response to the initial requirements, an initial draft is generated, and feature information is extracted from the initial draft in reverse to generate retrieval basis information; Based on the search criteria information, an exact match search is performed on the preset memory cache library to obtain the target text slice as context; The target text is generated by combining the aforementioned context with the initial draft; The target text is subjected to requirement verification to obtain a verification result. When the verification result indicates that the target text has a missing feature vector, the missing feature vector is added to the initial draft, and the reverse feature information extraction and the exact matching retrieval are re-executed until the verification result indicates that the missing feature vector does not exist, and the final text is output.

2. The data augmentation generation method according to claim 1, characterized in that, Before performing an exact match retrieval on the preset memory cache library, the method further includes: A preset structured data source is read, and the structured data source is sliced ​​according to a preset line number threshold to obtain text slices. Metadata for precise matching retrieval and positioning is injected into each text slice. All text slices are serialized into plain text format and loaded into memory to generate the memory cache library.

3. The data augmentation generation method according to claim 2, characterized in that, The injection of metadata for precise matching retrieval into each of the text slices includes: The metadata is injected into each text slice, and the text slice after the metadata injection is serialized into a semi-structured data format and stored in the memory cache library.

4. The data augmentation generation method according to claim 1, characterized in that, The step of reversely extracting feature information from the initial draft to generate retrieval basis information includes: The initial draft is input into a large language model, and the large language model is constrained to extract feature information in reverse based on the content of the initial draft to obtain the extraction result; Filter out common pronouns in the extracted results and perform deduplication logic on the extracted results to generate a set of search keys as the search basis information.

5. The data augmentation generation method according to claim 1, characterized in that, The step of performing an exact match search on a preset memory cache library based on the search criteria information to obtain the target text slice as context includes: Calculate the frequency of occurrence of the retrieval criteria information in each text slice in the memory cache library, obtain the score of the text slice based on the frequency of occurrence, and use the preset number of target text slices with the highest scores as the context.

6. The data augmentation generation method according to claim 5, characterized in that, The step of obtaining the score for the text slice based on the frequency of occurrence includes: in, The score for the text slice, where n is the total number of feature information in the retrieval criteria information. The frequency of the i-th feature information appearing in the text slice j. The weights are those of the feature information.

7. The data augmentation generation method according to claim 1, characterized in that, The requirement validation of the target text to obtain the validation result includes: The initial requirement is broken down into multiple requirement sub-items, and the verification model is called to check whether the target text covers all the requirement sub-items one by one. Based on the verification result of the verification model, a quantitative score and the missing feature vector are output. Adding the missing feature vector to the initial draft includes: The missing feature vectors are concatenated into the initial draft using a specified identifier format.

8. The data augmentation generation method according to claim 1, characterized in that, The re-execution of the reverse feature extraction and the precise matching retrieval also includes: The newly extracted feature information is merged with all previously extracted feature information to remove duplicates, and the merged and deduplicated feature information is used as the updated retrieval basis information.

9. The data augmentation generation method according to claim 1, characterized in that, The re-execution of the reverse feature extraction and the precise matching retrieval also includes: Determine whether the number of iterations already performed has reached the preset maximum number of iterations. If it has, stop the iteration and output the target text with the highest historical score.

10. A data augmentation generation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the data augmentation generation method according to any one of claims 1 to 9.

Images