Manuscript review optimization methods, systems, equipment, storage media, and program products

Abstract

This application provides a method, system, device, storage medium, and program product for optimizing manuscript review. The method includes obtaining the manuscript to be reviewed and the preset standards for corresponding target inspection items; retrieving target examples with relevance higher than a preset threshold from a historical calibration data repository; generating incremental prompts based on the target examples and preset inspection standards; inputting the incremental prompts and the manuscript to be reviewed into a pre-trained review model to generate a preliminary review report; obtaining correction data for each inspection item in the preliminary review report; and updating the historical calibration data repository based on the correction data. This method is used to continuously expand the historical calibration data repository with the help of manually calibrated data, optimize the incremental prompt generation logic, and improve the accuracy and consistency of intelligent manuscript review.

Description

Technical Field

[0001] This application relates to the field of text processing, and in particular to a method, system, device, storage medium, and program product for optimizing manuscript review. Background Technology

[0002] As the scale of text content production continues to expand, the demand for efficient and accurate manuscript review is becoming increasingly urgent. Intelligent review technology, with its processing efficiency advantage, has gradually become a core means in the industry to improve the efficiency of review work, playing an important role in reducing manual review costs and shortening the review cycle.

[0003] In existing technologies, intelligent manuscript review systems typically use preset inspection rules to automatically screen various indicators of manuscripts and generate corresponding preliminary review reports. The review process relies on fixed algorithm models and rule bases to complete the initial judgment of the manuscript, and then feeds the results back to the human review stage. Human reviewers verify and correct the results of the intelligent review, but the corrected information is only used as the conclusion of a single review.

[0004] There is a data gap between the existing manuscript review system and the calibration process. The effective data generated by the correction is not fully utilized to optimize the intelligent review prompt strategy. As a result, the accuracy of intelligent review is difficult to continuously improve with the increase of the number of reviews. There is always a problem that the review results deviate significantly from the actual needs, and it cannot meet the requirements of long-term, stable, high-quality review. Summary of the Invention

[0005] This application provides a manuscript review optimization method, system, device, storage medium, and program product to improve the accuracy of manuscript review.

[0006] In a first aspect, embodiments of this application provide a manuscript review optimization method, including:

[0007] Obtain the manuscript to be reviewed and the preset inspection standards for the corresponding target inspection items;

[0008] Retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations.

[0009] Based on the target example and the preset inspection item standards, incremental prompt words are generated;

[0010] The incremental prompts and the manuscript to be reviewed are input into a pre-trained review model for analysis, and a corresponding preliminary review report is generated.

[0011] Obtain the revised data for each inspection item in the preliminary review report;

[0012] The historical calibration data repository is updated based on the corrected data, wherein uncorrected inspection items and their corresponding audit results are stored as positive examples, the original text fragments and original audit results corresponding to corrected inspection items are stored as negative examples, and the corrected content is stored as a positive example.

[0013] In one possible implementation, retrieving target examples from a historical calibration data repository that have a relevance to the manuscript to be reviewed exceeding a preset threshold includes:

[0014] The original text fragments of the manuscript to be reviewed are vectorized using a pre-trained vector model to obtain the manuscript feature vectors.

[0015] Calculate the vector similarity between the manuscript feature vector and the original text vector of each example in the historical calibration data repository;

[0016] Examples whose vector similarity is higher than a preset threshold are selected as candidate examples;

[0017] The candidate examples are matched and filtered with the target check items to obtain the target examples.

[0018] In one possible implementation, based on the target example and the preset check item standards, incremental prompt words are generated, including:

[0019] The target examples are categorized into positive and negative examples to obtain the categorized example data.

[0020] The categorized example data is integrated with preset inspection item standards to construct incremental prompt words.

[0021] In one possible implementation, it also includes:

[0022] Determine whether the length of the incremental prompt word exceeds the context window length of the pre-trained review model;

[0023] If the number of examples exceeds the limit, the classified example data will be clustered by similarity according to positive and negative examples to obtain multiple example groups.

[0024] Information is extracted and content is compressed for each example group to obtain the core example information corresponding to each example group;

[0025] Based on the core example information of all example groups and the preset check item standards, incremental prompt words with the required length are generated.

[0026] In one possible implementation, before obtaining the revised data for each inspection item in the preliminary review report, the method further includes:

[0027] Display the preliminary review report and indicate the original source of each inspection item, and generate a review display page;

[0028] The correction data is obtained by receiving the editing and correction operation through the calibration entry on the audit display page.

[0029] In one possible implementation, after updating the historical calibration data repository based on the corrected data, the method further includes:

[0030] Monitor whether the amount of data in the historical calibration data repository exceeds a preset storage threshold;

[0031] If the amount of data in the historical calibration data repository exceeds a preset storage threshold, the examples in the historical calibration data repository are grouped according to the semantic features of the original text and the type of the check item to obtain multiple semantic groups.

[0032] Calculate the semantic similarity between examples within each semantic group, and remove duplicate examples whose semantic similarity is higher than a preset deduplication threshold;

[0033] The original core fragments, review conclusions and correction basis of the deduplicated semantic groups are merged to generate a condensed example for each semantic group;

[0034] Replace the corresponding semantic group in the historical calibration data repository with the condensed example.

[0035] Secondly, embodiments of this application provide a manuscript review and optimization system, including:

[0036] The acquisition module is used to acquire the manuscript to be reviewed and the preset inspection standards for the corresponding target inspection items;

[0037] The retrieval module is used to retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations.

[0038] The generation module is used to generate incremental prompt words based on the target example and the preset inspection item standards;

[0039] The analysis module is used to input the incremental prompts and the manuscript to be reviewed into a pre-trained review model for analysis and to generate a corresponding preliminary review report.

[0040] The extraction module is used to obtain the correction data for each inspection item in the preliminary review report;

[0041] The update module is used to update the historical calibration data repository according to the corrected data, wherein the uncorrected inspection items and their corresponding audit results are stored as positive examples, the original text fragments and original audit results corresponding to the corrected inspection items are stored as negative examples, and the corrected content is stored as a positive example.

[0042] In one possible implementation, the retrieval module is specifically used for:

[0043] The original text fragments of the manuscript to be reviewed are vectorized using a pre-trained vector model to obtain the manuscript feature vectors.

[0044] Calculate the vector similarity between the manuscript feature vector and the original text vector of each example in the historical calibration data repository;

[0045] Examples whose vector similarity is higher than a preset threshold are selected as candidate examples;

[0046] The candidate examples are matched and filtered with the target check items to obtain the target examples.

[0047] In one possible implementation, the generation module is specifically used for:

[0048] The target examples are categorized into positive and negative examples to obtain the categorized example data.

[0049] The categorized example data is integrated with preset inspection item standards to construct incremental prompt words.

[0050] In one possible implementation, the generation module is further configured to:

[0051] Determine whether the length of the incremental prompt word exceeds the context window length of the pre-trained review model;

[0052] If the number of examples exceeds the limit, the classified example data will be clustered by similarity according to positive and negative examples to obtain multiple example groups.

[0053] Information is extracted and content is compressed for each example group to obtain the core example information corresponding to each example group;

[0054] Based on the core example information of all example groups and the preset check item standards, incremental prompt words with the required length are generated.

[0055] In one possible implementation, the extraction module is further configured to:

[0056] Display the preliminary review report and indicate the original source of each inspection item, and generate a review display page;

[0057] The correction data is obtained by receiving the editing and correction operation through the calibration entry on the audit display page.

[0058] In one possible implementation, the updating module is further configured to:

[0059] Monitor whether the amount of data in the historical calibration data repository exceeds a preset storage threshold;

[0060] If the amount of data in the historical calibration data repository exceeds a preset storage threshold, the examples in the historical calibration data repository are grouped according to the semantic features of the original text and the type of the check item to obtain multiple semantic groups.

[0061] Calculate the semantic similarity between examples within each semantic group, and remove duplicate examples whose semantic similarity is higher than a preset deduplication threshold;

[0062] The original core fragments, review conclusions and correction basis of the deduplicated semantic groups are merged to generate a condensed example for each semantic group;

[0063] Replace the corresponding semantic group in the historical calibration data repository with the condensed example.

[0064] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0065] The memory stores computer-executed instructions;

[0066] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0067] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0068] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0069] The manuscript review optimization method, system, device, storage medium, and program products provided in this application combine target inspection item standards with historical calibration data repository examples to generate incremental prompts adapted to the review scenario. These incremental prompts, along with the manuscript to be reviewed, are simultaneously input into a pre-trained review model to complete the initial review. Correction data corresponding to the initial review results is collected, and the historical calibration data repository is updated and improved in reverse, constructing a closed-loop review optimization process. By transforming manual calibration information into effective review basis, the bias caused by single, fixed review rules is mitigated, enabling long-term retention and reuse of historical calibration information. This continuously optimizes the generation basis of incremental prompts, gradually improving the accuracy and relevance of intelligent review results, reducing the frequency of subsequent manual corrections, and promoting continuous iteration of intelligent review quality. Ultimately, this improves the overall manuscript review efficiency and ensures that review results consistently match actual review needs over the long term. Attached Figure Description

[0070] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0071] Figure 1 A schematic diagram illustrating the application scenario of the manuscript review optimization method provided in this application embodiment;

[0072] Figure 2 A flowchart illustrating the manuscript review optimization method provided in this application embodiment;

[0073] Figure 3 This is a schematic diagram of the manuscript review and optimization system provided in the embodiments of this application;

[0074] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0075] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0076] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with some aspects of this application as detailed in the appended claims.

[0077] First, let me explain the terms used in this application:

[0078] Target inspection items: These refer to the various review dimensions that need to be verified for compliance in the manuscript to be reviewed, providing a clear direction for judgment in the manuscript review process;

[0079] Pre-defined inspection standards: These are standardized criteria that are set in advance to determine whether the target inspection items of a manuscript under review are compliant.

[0080] Historical calibration data repository: This refers to a database that stores correct and incorrect calibration data generated from historical audits. Positive examples are compliance audit data, and negative examples are deviation audit data, which are used to provide a reference for intelligent audits.

[0081] Target examples: These refer to historical review data selected from the historical calibration data repository that meet preset standards in terms of semantic relevance and check item matching with the manuscript to be reviewed.

[0082] Incremental prompts: These are instruction texts generated by combining target examples with preset inspection standards, used to input into the pre-trained audit model to improve audit accuracy;

[0083] Pre-trained review model: refers to a data model that has been trained on a large amount of text data and has the ability to perform text analysis and compliance verification, and is used to perform the initial review of manuscripts to be reviewed;

[0084] Context window length: refers to the maximum length threshold of text instructions that the pre-trained auditing model can normally receive, parse, and process;

[0085] Preliminary review report: refers to the structured review document generated by the pre-trained review model after it has completed the review of the manuscript to be reviewed based on the incremental prompts. It includes the review results of each check item, the explanation of the results, the modification suggestions and the original source.

[0086] Correction data refers to the rectification guidance data extracted from the preliminary review report for the deviation issues of each inspection item, including the optimization direction of the inspection item and the basis for the qualification judgment;

[0087] Semantic group: refers to a collection of examples that are grouped according to the semantic features of the original text and the types of inspection items, and have similar review attributes and text features;

[0088] Condensed Examples: These are standardized examples generated by extracting and integrating the core information of examples within a semantic group, while retaining their complete audit reference value and reducing the data volume.

[0089] In existing technologies, manuscript review is conducted using fixed review rules, which makes it difficult to optimize the review criteria by incorporating historical calibration information. This results in technical problems such as insufficient accuracy of review results and the inability to reuse calibration data.

[0090] The manuscript review optimization method provided in this application completes the initial review of manuscripts by generating incremental prompts by combining inspection standards with historical examples, and updates the example library based on the initial review correction data. This solves the technical problems of the inability to reuse calibration information and the difficulty in improving the accuracy of review in the existing intelligent manuscript review.

[0091] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0092] Figure 1This is a schematic diagram illustrating the application scenario of the manuscript review optimization method provided in the embodiments of this application, such as... Figure 1 As shown, it includes: terminal 101 and server 102.

[0093] Terminal 101 is used to submit manuscripts to be reviewed and corresponding target inspection item requirements, and to display the review results to the server and receive manual corrections. Server 102 is used to receive manuscripts to be reviewed and target inspection item requirements submitted by the terminal, obtain preset inspection item standards and historical calibration data storage, retrieve target examples and generate incremental prompts, import pre-trained review models to generate preliminary review reports, and send them back to the terminal. At the same time, it receives correction data collected by the terminal and updates the historical calibration data storage.

[0094] Figure 2 This is a flowchart illustrating the manuscript review optimization method provided in this application embodiment. The execution entity of this embodiment can be... Figure 1 The server 102 in the illustrated embodiment can also be other computer-related devices, and this embodiment is not particularly limited.

[0095] like Figure 2 As shown, the manuscript review and optimization method includes the following steps:

[0096] Step S201: Obtain the manuscript to be reviewed and the preset inspection standards for the corresponding target inspection items.

[0097] Specifically, the process involves extracting text documents to be reviewed, simultaneously retrieving standardized review specifications corresponding to the target inspection items matching the documents, clarifying the judgment criteria and review requirements for each inspection item, and completing the collection and integration of basic information for the initial review phase. Based on the standardized text review process, a correspondence between review criteria and documents to be processed is established, clarifying the core principles of review execution, providing a directional execution basis for subsequent review processes, locking in the core scope and judgment standards of the review, ensuring that subsequent review operations align with established review specifications, achieving precise matching between review criteria and documents to be reviewed, and providing directional support for subsequent example retrieval and prompt word generation stages.

[0098] Step S202: Retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that are higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations.

[0099] Specifically, various review examples generated through historical manual calibration are extracted from the historical calibration data repository. Using vector similarity calculations, examples with a matching degree exceeding a preset value to the manuscript to be reviewed are selected and designated as target examples for subsequent prompt word generation. Based on text feature vector matching logic, the relevance between the manuscript to be reviewed and historical calibration examples is quantitatively compared, and highly adaptable historical data is selected. Effective examples that fit the manuscript to be reviewed are then selected from the historical calibration data, providing historical reference data that aligns with the actual review scenario for prompt word generation. This allows for rapid acquisition of historical reference information suitable for the current review task, improving the targeting of subsequent prompt word generation.

[0100] Step S203: Generate incremental prompts based on the target example and preset check item standards.

[0101] Specifically, the selected target examples and preset inspection standards are integrated. After classifying and sorting the examples into positive and negative examples, the two types of information are merged to construct incremental prompts adapted to the current audit task. By combining standardized audit rules with historical practical calibration experience, instruction texts with clear audit guidance are constructed through information fusion. This provides refined audit instructions for the pre-trained audit model, transforming audit standards into refined audit instructions that the model can recognize and execute. This optimizes the execution logic of the model's audit and generates intelligent audit instructions that fit actual audit needs.

[0102] Step S204: Input the incremental prompts and the manuscript to be reviewed into the pre-trained review model for analysis, and generate the corresponding preliminary review report.

[0103] Specifically, the completed incremental prompts are simultaneously imported into the pre-trained review model along with the manuscript to be reviewed. The model analyzes the manuscript based on the incremental prompts and the prompts extracted by the model, outputting a structured preliminary review report that includes the name of each inspection item, the inspection results, a description of the results, suggested revisions, and the original source. By leveraging the text analysis capabilities of the pre-trained review model and increasing the accuracy of the review by adding incremental prompts, the model automates the screening and judgment of various aspects of the manuscript, outputs structured review results, quickly generates a standardized preliminary review report, and completes the manuscript review process.

[0104] Step S205: Obtain the revised data for each inspection item in the preliminary review report.

[0105] Specifically, the inspection results, explanations, and suggested modifications for each inspection item in the preliminary review report are extracted and integrated to form a set of corrected data for each inspection item, completing the structured extraction and aggregation of data. Based on text data structured parsing technology, multi-dimensional audit information in the preliminary review report is extracted, and the inspection item information is transformed into standardized corrected data, realizing the aggregation of audit data. Core data with corrective guidance value is extracted from the preliminary review report, providing data support for the subsequent updating of the historical calibration data repository, forming standardized and unified corrected data.

[0106] Step S206: Update the historical calibration data repository according to the corrected data, wherein the uncorrected inspection items and their corresponding audit results are stored as positive examples, the original text fragments and original audit results corresponding to the corrected inspection items are stored as negative examples, and the corrected content is stored as a positive example.

[0107] Specifically, uncorrected check items and their corresponding review results are stored as positive examples to represent the review pattern where the initial review results meet the editor's expectations; corrected check items and their corresponding original text fragments and original review results are stored as negative examples to represent the deviation pattern between the initial review results and actual needs; when corrected content is stored as a positive example, it must include the basis for the correction (such as the reason for the modification, the rule basis, or the specific adjustment logic) to ensure that the positive example data completely maps the correction logic.

[0108] The data collected from the initial review report is transformed into structured data and categorized into positive and negative examples based on the type of correction (positive / negative). Uncorrected check items and review results are directly used as positive examples, indicating that the initial review results meet expectations. For corrected check items, the original text and review results are used as negative examples, while the corrected content is used as a positive example, forming an "error-correction" pair for subsequent optimization of enhancement prompts. By injecting the corrected and valid data into a historical example library, the library is dynamically supplemented and iterated, continuously increasing the volume of historical calibration data. This ensures the example library always contains the latest review standards and practical experience, providing a reference for subsequent review tasks and forming a closed-loop iteration mechanism for review data.

[0109] The manuscript review optimization method provided in this invention integrates historical calibration data and standardized review rules to construct intelligent review instructions, and optimizes the review reference library in reverse based on the correction data. This breaks the rigid mode of traditional intelligent review, realizes the long-term reuse of review data and the continuous improvement of review quality. While ensuring review efficiency, it steadily improves the accuracy and stability of intelligent manuscript review, simplifies the review process, optimizes the review results, and significantly improves the overall operational efficiency of manuscript review.

[0110] This embodiment provides a detailed description of the process in the above embodiment of retrieving target examples from the historical calibration data repository whose relevance to the manuscript to be reviewed is higher than a preset threshold. The specific implementation of this process includes the following steps:

[0111] Step a1: Vectorize the original text fragments of the manuscript to be reviewed using a pre-trained vector model to obtain the manuscript feature vector.

[0112] Specifically, the core original text fragments of the manuscript to be reviewed are extracted and input into a pre-trained vector model for feature mapping operations. This transforms unstructured text information into high-dimensional numerical vectors, completing the digital extraction of the manuscript's text features. Based on natural language processing technology, a deep learning model is used to quantify the semantic and structural features of the text, achieving a vectorized expression of the text content. This transforms the text manuscript into computable numerical features, providing basic data support for subsequent similarity measurement calculations. This achieves a structured expression of the text features of the manuscript to be reviewed while preserving the core semantic information of the original text.

[0113] Step a2: Calculate the vector similarity between the manuscript feature vector and the original text vector of each example in the historical calibration data repository.

[0114] Specifically, spatial distance calculation algorithms can be used to quantify the spatial distance between the manuscript's feature vector and the original text vector of each example in the historical calibration database, thus calculating the semantic similarity between the two. By utilizing the distance measurement principle between high-dimensional spatial vectors, the semantic similarity between different text vectors can be quantitatively calculated, enabling an objective numerical determination of text content relevance. This standardizes the quantification of the semantic relevance between the manuscript under review and historical examples, providing a quantitative basis for selecting highly relevant examples and transforming abstract text relevance into a quantifiable numerical indicator. Other similarity algorithms can also be selected according to actual needs; no specific restrictions are imposed here.

[0115] Step a3: Select examples whose vector similarity is higher than a preset threshold as candidate examples.

[0116] Specifically, by pre-setting a similarity threshold, the calculated vector similarity values ​​are compared with the preset threshold to filter out examples whose similarity exceeds the threshold range, thus defining them as a candidate example set. Based on a preset relevance judgment standard, the quantitative calculation results are subjected to threshold screening to achieve preliminary extraction of highly relevant historical data. From the historical examples, samples that are highly semantically related to the manuscript to be reviewed are selected, completing the preliminary screening of highly relevant examples and simplifying the scope of data processing.

[0117] Step a4: Match and filter the candidate examples with the target check items to obtain the target examples.

[0118] Specifically, by extracting the target inspection item tags from the manuscripts to be reviewed, matching them with the corresponding inspection item tags of candidate examples, and retaining examples that meet the target inspection item requirements, these are identified as target examples. Combining the specific inspection item dimensions of the review task, the initially screened candidate examples are targeted for matching, achieving precise screening guided by the review process. A second filtering of candidate examples is then performed based on specific review needs to ensure that the selected examples are suitable for the current review task, providing high-quality reference data matching the target inspection items.

[0119] This invention transforms text into high-dimensional semantic vectors through a pre-trained vector model. By measuring the semantic relevance between the manuscript and historical examples through vector similarity, highly relevant candidate examples are selected, achieving accurate extraction of highly relevant and well-fitting historical examples. This effectively improves the efficiency and quality of example retrieval, providing solid and accurate data support for the generation of subsequent incremental prompts, and ensuring the accuracy and efficiency of the overall review process.

[0120] This embodiment provides a detailed description of the process of generating incremental prompt words based on target examples and preset inspection standards in the above embodiments. The specific implementation of this process includes the following steps:

[0121] Step b1: Classify and organize the target examples into positive and negative examples to obtain the classified example data.

[0122] Specifically, the target example dataset to be processed is obtained, and it is classified and labeled according to positive and negative examples to form structured example data. Unstructured text is transformed into high-dimensional numerical features through vector mapping, realizing the digital and structured expression of example content. This provides standardized and quantifiable basic data support for subsequent prompt word construction, enabling orderly management of example data and facilitating accurate matching, filtering, and integration of data in subsequent processes.

[0123] Step b2 involves integrating the categorized example data with preset check item standards to construct incremental prompt words.

[0124] Specifically, the categorized positive and negative example data are deeply integrated with preset inspection standards. Based on preset inspection labeling and audit rules, incremental prompts with clear guidance and execution are constructed. Through information fusion, standardized audit rules are quantitatively matched with business scenario requirements to generate targeted instruction texts. This provides clear and explicit execution basis for the automated audit process, and the generated incremental prompts map audit requirements, providing a clear execution path for the audit.

[0125] This invention, through target example classification and incremental prompt word construction, achieves fully automated processing from raw data collection and structured classification to standardized prompt word construction, effectively improving the accuracy and efficiency of prompt word generation. It can accurately match the review needs under different business scenarios and provide stable and efficient technical support for automated review work.

[0126] In one specific embodiment, the above embodiment further includes the following steps:

[0127] Step c1: Determine whether the length of the incremental prompt word exceeds the context window length of the pre-trained review model.

[0128] Specifically, the raw data of the incremental prompt words to be generated is obtained and compared with the context window length threshold of the pre-trained review model to determine whether the length of the raw prompt words exceeds the processing range of the model. Based on natural language processing technology, the matching degree between the prompt word length and the model's context window is quantified, redundant information is pre-screened, providing clear guidance for subsequent content simplification, ensuring that the prompt word length is adapted to the model's processing capabilities, guaranteeing the basic feasibility of the review process, and making the prompt word length match the context window requirements of the pre-trained review model, eliminating the impact of information redundancy on the model's computational efficiency.

[0129] In step c2, if the number of cases exceeds the limit, the classified example data will be clustered by similarity according to positive and negative examples to obtain multiple example groups.

[0130] Specifically, for the filtered categorized example data, similarity is quantified based on positive and negative example attributes, and example data with similar semantic features are integrated into several independent example groups. During clustering, check item labels must be retained as constraints to ensure that cases within each example group correspond to the same check item type; the clustering results must be verified through check item labels, eliminating example groups irrelevant to the current review task to ensure the relevance of subsequent prompt word generation. Based on textual semantic similarity calculation, example data is transformed into quantifiable feature representations through vector mapping, achieving accurate aggregation of similar examples, standardizing the classification logic of example data, simplifying the complexity of subsequent information processing, ensuring that each group of example data has a unified semantic orientation, providing standardized data support for core information extraction, and forming example groups with clear structure and consistent semantics.

[0131] Step c3 involves extracting information and compressing content for each example group to obtain the core example information corresponding to each example group.

[0132] Specifically, each example group undergoes in-depth information analysis to extract core semantic features and compress the content, eliminating irrelevant and redundant expressions while retaining key review guidance information. Through feature selection and quantitative integration, the information in the example groups is transformed into concise and standardized core content, extracting the core value of the example groups and ensuring a high degree of alignment between the core information and the preset inspection standards.

[0133] Step c4: Based on the core example information and preset check item standards of all example groups, generate incremental prompt words with the required length.

[0134] Specifically, based on the core information of all example groups, and combined with preset inspection standards, incremental prompts are constructed with appropriate length and precise semantics. This ensures that the prompts fully comply with the context window requirements and review specifications of the pre-trained review model. Through information fusion and vector optimization, the core example information is precisely matched with the review standards to generate standardized and compliant prompt instructions. This unifies the prompt generation standards, ensures that the prompt length is compatible with the model's processing capabilities, and guarantees the semantic precision of the prompts, providing a clear and effective execution basis for the overall review process.

[0135] This invention clusters positive and negative example data based on semantic similarity, merging similar cases into example groups to reduce redundant descriptions. Information is extracted from each example group, retaining core semantic features (such as error type and correction logic) while removing irrelevant details to generate concise core example information. Based on the core example information and preset inspection standards, length-appropriate enhanced prompts are constructed. Similarity clustering avoids redundant duplicate cases, improving the compactness of example information. Content compression ensures that the enhanced prompts remain within the model's processing capabilities while retaining key review logic to prevent information loss. The optimized prompts are easier for the model to parse, improving the efficiency and accuracy of the initial review.

[0136] This embodiment provides a detailed description of the process prior to obtaining the corrected data for each inspection item in the preliminary review report as described in the above embodiments. The specific implementation of this process also includes the following steps:

[0137] Step d1: Display the preliminary review report and indicate the original source of each inspection item, and generate the review display page.

[0138] Specifically, the review results of each inspection item in the preliminary review report are visualized and rendered, synchronously linked and matched with the original text of the manuscript to be reviewed, and the text fragments corresponding to each inspection item are located and labeled, integrating and generating a visual interactive page. By establishing a targeted correspondence between the review results and the original text of the manuscript to be reviewed, a structured and visual presentation of review information is achieved, an interactive carrier for the results of manual and intelligent review is built, the correspondence between the review conclusions and the original text content is clearly presented, and the integrated display of the preliminary review report and the original text information is realized, making it easy to locate the text content corresponding to the review items.

[0139] Step d2 involves receiving the editing and correction operations through the calibration entry on the review and display page to obtain the correction data.

[0140] Specifically, through the calibration interaction portal on the review display page, modification and correction instructions for each review result in the initial review report are collected. The modification content within these instructions is extracted and organized into standardized correction data. This transforms manual correction operations into structured data that can be recognized and stored by the system, establishing an interactive channel between manual correction and system data collection. This enables the collection and digital transformation of manual review opinions, retains manual correction information, and forms standardized, reusable correction data, providing data support for subsequent updates to the historical calibration data repository.

[0141] This invention improves operability by visually presenting the relationship between the review results and the original text and establishing a dedicated interactive entry point to receive correction operations. It also improves the efficiency and accuracy of correction data collection, provides a stable and reliable data source for subsequent updates to the historical calibration data repository, and improves the human-machine collaboration of the overall review process.

[0142] This embodiment provides a detailed description of the process after updating the historical calibration data repository based on the correction data in the above embodiments. The specific implementation of this process also includes the following steps:

[0143] Step e1: Monitor whether the amount of data in the historical calibration data repository exceeds the preset storage threshold.

[0144] Specifically, the total number of examples stored in the historical calibration data repository is obtained in real time. This total number of examples is compared with a pre-set storage threshold to determine whether the data volume in the repository exceeds the limit. By routinely and quantitatively determining the data storage scale of the repository, a pre-emptive warning mechanism for data overload is established. This allows for timely control of the data volume in the historical calibration data repository, identifying situations where excessive data leads to decreased retrieval efficiency and excessive storage consumption, and providing a trigger for subsequent repository optimization processes.

[0145] Step e2: If the amount of data in the historical calibration data repository exceeds the preset storage threshold, the examples in the historical calibration data repository are grouped according to the semantic features of the original text and the type of inspection item to obtain multiple semantic groups.

[0146] Specifically, the textual semantic features and corresponding inspection item type information of each example in the database are extracted. Examples with similar features and types are grouped and integrated into multiple independent semantic groups. By combining semantic attributes and business review dimensions, the scattered examples in the database are structurally grouped to achieve data classification and regularization, narrow the scope of subsequent data processing, improve the targeting of subsequent similarity calculations and data simplification, and form a clear hierarchical example grouping.

[0147] Step e3: Calculate the semantic similarity between examples within each semantic group, and remove duplicate examples whose semantic similarity is higher than the preset deduplication threshold.

[0148] Specifically, a one-to-one semantic similarity quantification is performed on all examples within a single semantic group. The calculation results are compared with a preset deduplication threshold, and duplicate examples exceeding the threshold are filtered out. The semantic overlap of examples within a group is quantified through text vector similarity calculation. Based on the threshold, duplicate data is filtered and removed, eliminating highly repetitive example data within the group, reducing redundant information in the database, lowering the computational cost of subsequent data processing, retaining examples with differentiated value, and improving the uniqueness and effectiveness of the data in the database.

[0149] Step e4: Merge the original core fragments, review conclusions and correction basis of the deduplicated semantic groups to generate a condensed example for each semantic group.

[0150] Specifically, after deduplication, three types of key information are extracted from the original core fragments, review conclusions, and correction basis of the examples within the semantic group. This similar information is then integrated to generate a single, concise, condensed example. Based on information fusion and core feature extraction, the core effective information of the examples within the group is retained, redundant expressions are eliminated, and the core value information of multiple examples within the same semantic group is integrated. This process preserves the core value of the data while compressing its volume, generating condensed examples that are both complete in information and concise in size.

[0151] Step e5: Replace the corresponding semantic group in the historical calibration data repository with the condensed example.

[0152] Specifically, the generated condensed examples will replace all existing examples within the original semantic groups, completing the update and replacement of the database data and reconstructing the streamlined database data structure. Through data replacement and database reconstruction, the streamlined condensed data replaces the original scattered data, achieving iterative optimization of the database data. This completes the overall streamlining and optimization of the historical calibration data repository data, reducing the total amount of database data while retaining core audit reference value, optimizing the overall data structure of the historical calibration data repository, reducing database storage footprint, and improving the efficiency of subsequent example retrieval.

[0153] This invention identifies semantically repetitive examples through semantic similarity calculation, eliminates redundant data, retains differentiated cases, merges core information of deduplicated semantic groups, extracts core fragments of the original text, review conclusions, and correction basis, and generates simplified condensed examples. While retaining the core reference value of the examples, it reduces the size of the database, optimizes the data storage structure, improves the efficiency of subsequent example retrieval, and ensures the stable and efficient operation of manuscript review.

[0154] In one specific embodiment, the process also includes the construction of a historical calibration data repository, which is further implemented by the following steps:

[0155] Step f1: Use the pre-trained review model to review the manuscript samples and generate a structured preliminary review report.

[0156] Specifically, manuscript samples are input into a pre-trained review model. Text analysis is performed based on preset check items to generate a structured preliminary review report containing check item names, check standards, check results, result explanations, revision suggestions, and original source information. By performing multi-dimensional matching and judgment on the content of the manuscripts to be reviewed, standardized review results are output according to a fixed format, completing the initial screening of manuscripts and outputting complete and standardized preliminary review conclusions. This provides basic review data for subsequent data collection and example classification, and serves as a preliminary data source for building a historical calibration data repository.

[0157] Step f2: Obtain calibration information for the preliminary review report and complete the correction of the review results.

[0158] Specifically, by conducting targeted corrections to the inspection results, explanations, and modification suggestions, the intelligent preliminary review results are optimized and adjusted. Through external calibration data integration and data correction logic, the preliminary review deviations are corrected, the limitations of the preliminary review are addressed, and accurate review conclusions that conform to the review standards are formed. This ensures the accuracy of the data entering the database, resulting in standardized and highly accurate review result data. It also eliminates data deviations caused by erroneous judgments in the preliminary review, providing high-quality raw data for the construction of a historical calibration data repository.

[0159] Step f3: Collect and organize the full audit data before and after calibration, and generate standardized data entries.

[0160] Specifically, the process involves collecting the initial review results and the final calibration conclusions, integrating relevant information such as inspection items, inspection standards, original descriptions, post-calibration results, and initial review results to form structured data entries and retain them. By uniformly collecting and organizing data from the entire review process, scattered data is transformed into standardized data formats that can be directly categorized and stored. Complete comparison data before and after calibration is retained, and standardized data entries are established. This provides a well-organized data carrier for the categorized storage of historical calibration data, forming complete and directly reusable review data entries. This ensures the integrity and standardization of data during the construction of the historical calibration data repository, facilitating subsequent categorization, storage, and retrieval.

[0161] Step f4: Classify and collect data according to standardized data entries, and build a historical calibration data repository.

[0162] Specifically, correct results after calibration and qualified preliminary review results without correction are uniformly classified as positive example data, while incorrect preliminary review results before correction are classified as negative example data. These two types of data are stored in their respective databases, completing the construction and updating of the historical calibration data repository. The data is categorized and stored according to the correctness of the data assessment, distinguishing between correct and incorrect review examples. This achieves the categorization, accumulation, and long-term storage of experience-based review data, building a clearly structured and accurate historical calibration data repository. This ensures the long-term accumulation of review experience, continuously improves the repository content, and provides stable data support for subsequent manuscript review optimization.

[0163] This invention establishes a complete historical calibration data repository through preliminary review, calibration data acquisition, data organization, and classification and database construction. This enables the long-term accumulation and standardized storage of effective audit data, providing a solid data foundation for subsequent audit process optimization.

[0164] Figure 3 This is a schematic diagram of the manuscript review and optimization system provided in an embodiment of this application. Figure 3 As shown, the manuscript review and optimization system 30 includes:

[0165] The acquisition module 301 is used to acquire the manuscript to be reviewed and the preset inspection standards of the corresponding target inspection items;

[0166] The retrieval module 302 is used to retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations.

[0167] The generation module 303 is used to generate incremental prompts based on the target example and preset inspection item standards;

[0168] Analysis module 304 is used to input incremental prompts and manuscripts to be reviewed into a pre-trained review model for analysis and to generate a corresponding preliminary review report;

[0169] Extraction module 305 is used to obtain the correction data for each inspection item in the preliminary review report;

[0170] The update module 306 is used to update the historical calibration data repository based on the corrected data. Uncorrected inspection items and their corresponding audit results are stored as positive examples, while the original text fragments and original audit results corresponding to corrected inspection items are stored as negative examples, and the corrected content is stored as positive examples.

[0171] In one possible implementation, the retrieval module 302 is specifically used for:

[0172] The original text fragments of the manuscript to be reviewed are vectorized by a pre-trained vector model to obtain the manuscript feature vectors;

[0173] Calculate the vector similarity between the manuscript feature vector and the original text vector of each example in the historical calibration data repository;

[0174] Examples whose vector similarity exceeds a preset threshold are selected as candidate examples;

[0175] The candidate examples are matched and filtered with the target check items to obtain the target examples.

[0176] In one possible implementation, the generation module 303 is specifically used for:

[0177] The target examples are categorized into positive and negative examples to obtain the categorized example data.

[0178] The categorized example data is integrated with the preset inspection standards to construct incremental prompt words.

[0179] In one possible implementation, the generation module 303 is further configured to:

[0180] Determine whether the length of the incremental prompt word exceeds the context window length of the pre-trained review model;

[0181] If the number of cases exceeds the limit, the classified example data will be clustered by similarity according to positive and negative examples to obtain multiple example groups.

[0182] Information is extracted and content is compressed for each example group to obtain the core example information corresponding to each example group;

[0183] Based on the core example information and preset check item standards of all example groups, generate incremental prompt words with the required length.

[0184] In one possible implementation, the extraction module 305 is further configured to:

[0185] Display the preliminary review report and indicate the original source of each inspection item, and generate a review display page;

[0186] The system receives edit and correction data by accessing the calibration portal on the review and display page.

[0187] In one possible implementation, the update module 306 is further configured to:

[0188] Monitor whether the amount of data in the historical calibration data repository exceeds the preset storage threshold;

[0189] If the amount of data in the historical calibration data repository exceeds the preset storage threshold, the examples in the historical calibration data repository are grouped according to the semantic features of the original text and the type of inspection item to obtain multiple semantic groups;

[0190] Calculate the semantic similarity between examples within each semantic group, and remove duplicate examples whose semantic similarity is higher than the preset deduplication threshold;

[0191] The original core fragments, review conclusions and correction basis of the deduplicated semantic groups are merged to generate a condensed example for each semantic group;

[0192] Replace the corresponding semantic group in the historical calibration data repository with the condensed example.

[0193] The manuscript review and optimization system provided in this embodiment can be used to execute the manuscript review and optimization method described above. Its implementation principle and technical effect are similar, and will not be described again in this embodiment.

[0194] The manuscript review system can be deployed in a SaaS model, supporting interaction with external editing terminals via a network interface. Deployed in the cloud, the system provides API interfaces for external editing terminals to call, enabling a seamless workflow from manuscript submission and initial review report retrieval to correction data feedback. Data storage and computing resources are allocated on demand, supporting elastic scaling to adapt to review needs of varying scales. The SaaS model reduces user deployment costs and improves system availability. The network interface supports multi-terminal collaboration, enabling efficient collaboration between editors and intelligent reviewers. Elastic scaling ensures stable operation of the system under high-concurrency scenarios.

[0195] Figure 4 A schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application, such as... Figure 4 As shown, the electronic device 40 includes at least one processor 401 and a memory 402. Optionally, the electronic device 40 also includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0196] In the specific implementation process, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above method.

[0197] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0198] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0199] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0200] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0201] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0202] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0203] The aforementioned readable storage medium 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 storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0204] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0205] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interface, or indirect coupling or communication connection between systems or units, and may be electrical, mechanical, or other forms.

[0206] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0207] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0208] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0209] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0210] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for optimizing manuscript review, characterized in that, include: Obtain the manuscript to be reviewed and the preset inspection standards for the corresponding target inspection items; Retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations. Based on the target example and the preset inspection item standards, incremental prompt words are generated; The incremental prompts and the manuscript to be reviewed are input into a pre-trained review model for analysis, and a corresponding preliminary review report is generated. Obtain the revised data for each inspection item in the preliminary review report; The historical calibration data repository is updated based on the corrected data, wherein uncorrected inspection items and their corresponding audit results are stored as positive examples, the original text fragments and original audit results corresponding to corrected inspection items are stored as negative examples, and the corrected content is stored as a positive example.

2. The method according to claim 1, characterized in that, Retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold, including: The original text fragments of the manuscript to be reviewed are vectorized using a pre-trained vector model to obtain the manuscript feature vectors. Calculate the vector similarity between the manuscript feature vector and the original text vector of each example in the historical calibration data repository; Examples whose vector similarity is higher than a preset threshold are selected as candidate examples; The candidate examples are matched and filtered with the target check items to obtain the target examples.

3. The method according to claim 2, characterized in that, Based on the target example and the preset inspection item standards, incremental prompt words are generated, including: The target examples are categorized into positive and negative examples to obtain the categorized example data. The categorized example data is integrated with preset inspection item standards to construct incremental prompt words.

4. The method according to claim 3, characterized in that, Also includes: Determine whether the length of the incremental prompt word exceeds the context window length of the pre-trained review model; If the number of examples exceeds the limit, the classified example data will be clustered by similarity according to positive and negative examples to obtain multiple example groups. Information is extracted and content is compressed for each example group to obtain the core example information corresponding to each example group; Based on the core example information of all example groups and the preset check item standards, incremental prompt words with the required length are generated.

5. The method according to claim 1, characterized in that, Before obtaining the revised data for each inspection item in the preliminary review report, the process also includes: Display the preliminary review report and indicate the original source of each inspection item, and generate a review display page; The correction data is obtained by receiving the editing and correction operation through the calibration entry on the audit display page.

6. The method according to claim 1, characterized in that, After updating the historical calibration data repository based on the corrected data, the method further includes: Monitor whether the amount of data in the historical calibration data repository exceeds a preset storage threshold; If the amount of data in the historical calibration data repository exceeds a preset storage threshold, the examples in the historical calibration data repository are grouped according to the semantic features of the original text and the type of the check item to obtain multiple semantic groups. Calculate the semantic similarity between examples within each semantic group, and remove duplicate examples whose semantic similarity is higher than a preset deduplication threshold; The original core fragments, review conclusions and correction basis of the deduplicated semantic groups are merged to generate a condensed example for each semantic group; Replace the corresponding semantic group in the historical calibration data repository with the condensed example.

7. A manuscript review and optimization system, characterized in that, include: The acquisition module is used to acquire the manuscript to be reviewed and the preset inspection standards for the corresponding target inspection items; The retrieval module is used to retrieve target examples from the historical calibration data repository that have a relevance to the manuscript to be reviewed that is higher than a preset threshold. The historical calibration data repository stores positive and negative examples generated based on historical calibration operations. The generation module is used to generate incremental prompt words based on the target example and the preset inspection item standards; The analysis module is used to input the incremental prompts and the manuscript to be reviewed into a pre-trained review model for analysis and to generate a corresponding preliminary review report. The extraction module is used to obtain the correction data for each inspection item in the preliminary review report; The update module is used to update the historical calibration data repository according to the corrected data, wherein the uncorrected inspection items and their corresponding audit results are stored as positive examples, the original text fragments and original audit results corresponding to the corrected inspection items are stored as negative examples, and the corrected content is stored as a positive example.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.

Images