Processing method and system for improving recognition processing efficiency based on gist automatic grouping

Abstract

The application relates to a processing method and system for improving recognition processing efficiency based on automatic grouping of key points, and belongs to the technical field of content recognition processing. The method comprises the following steps: receiving a processing instruction proposed by a user, analyzing all files contained in the processing instruction to obtain an examination key point list and a corresponding grouping thought chain; inputting all examination key points in the examination key point list and the corresponding grouping thought chain into a preset grouping intelligent agent, classifying all examination key points through the grouping intelligent agent to form a plurality of first examination groups; distributing an examination model to each first examination group through an asynchronous calling framework, generating and starting a plurality of examination threads, and the examination threads correspond to the first examination groups one by one, so as to process all examination key points contained in the corresponding first examination group; and obtaining and merging the examination results of all examination threads to output the results for the user to know. The application has the effect of improving the processing efficiency of multiple files.

Description

Technical Field

[0001] This application relates to the field of content recognition and processing technology, and in particular to a processing method and system for improving recognition and processing efficiency based on automatic grouping of key points. Background Technology

[0002] With the development of information technology, the number of documents and materials that enterprises and individuals need to process and review is growing exponentially, especially in the fields of law, finance, scientific research and government affairs, where the demand for multi-document review is becoming increasingly prominent.

[0003] Traditional document processing and review methods mainly rely on manual reading and comparison of document content. When faced with a large number of documents and / or a large number of review points, the above-mentioned manual processing methods are not only inefficient, but also easily affect the accuracy of review and processing, so they need to be improved. Summary of the Invention

[0004] To improve the efficiency and accuracy of review and processing, this application provides a processing method and system that improves the efficiency of identification and processing by automatically grouping key points.

[0005] Firstly, this application provides a processing method for improving recognition efficiency based on automatic grouping of key points, employing the following technical solution:

[0006] Receive processing instructions from users, parse all files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains;

[0007] Input all the review points in the review point list and the corresponding group thinking chain into the preset group intelligence agent, and classify all the review points into several first-level review groups through the group intelligence agent;

[0008] By using an asynchronous call framework, a review model is assigned to each of the first-level review groups, and several review threads are generated and started. Each review thread corresponds one-to-one with a first-level review group to process all the review points contained in the corresponding first-level review group.

[0009] Retrieve and merge the review results from all review threads so that users can see the output.

[0010] By adopting the above technical solution, all the obtained files are first parsed, and all the review points that need to be processed for all the aforementioned files are determined (i.e., the review points included in the review point list). When there are multiple review points, the review task can be dynamically hierarchically grouped by combining the group thinking chain. Finally, the adaptation model is called asynchronously to achieve efficient parallel review, thereby replacing manual review and improving review efficiency.

[0011] Optionally, the review points include at least the entity to be reviewed;

[0012] The process involves inputting all review points from the review point list and their corresponding grouping thought chains into a preset grouping agent. The grouping agent then categorizes all review points into several primary review groups, including:

[0013] All review points in the review point list are aligned with entities. The entity-aligned review points and their corresponding grouped thinking chains are input into a preset grouped intelligent agent. The grouped intelligent agent classifies all review points into several first-level review groups so that different entities correspond to different first-level review groups, generating the first mapping relationship between entities and first-level review groups.

[0014] Based on preset review types, the review points in each primary review group are classified to generate secondary review groups; wherein, the review types include at least plain text review type and visual image review type;

[0015] The asynchronous call framework allocates a review model to each of the first-level review groups, generates and starts several review threads, and each review thread corresponds one-to-one with a first-level review group, including:

[0016] For each of the secondary review groups, a review model is matched, and several review threads are generated and started, with each review thread corresponding to a secondary review group.

[0017] By adopting the above technical solution, the primary grouping of review points is realized based on entities (i.e., forming first-level review groups). Then, according to different review types, the first-level review groups are further classified to match dedicated review models and improve review efficiency.

[0018] Optionally, the step of inputting all review points in the review point list and the corresponding grouping thought chains into a preset grouping agent, and classifying all review points into several primary review groups through the grouping agent, further includes:

[0019] Analyze the review points contained in each of the secondary review groups, determine the file list corresponding to each of the secondary review groups, the file list contains the files that need to be called to execute the review points contained in the corresponding secondary review group; establish a second mapping relationship between the review points in the secondary review group and the files in the corresponding file list;

[0020] Based on the second mapping relationship, all review points corresponding to the same document in each of the secondary review groups are taken as associated review points and merged to form a tertiary review group;

[0021] Establish a relationship between Level 3 review groups that belong to different Level 2 review groups but correspond to the same documents;

[0022] The process involves matching a review model to each of the secondary review groups, generating and starting several review threads, with each review thread corresponding one-to-one with a secondary review group, including:

[0023] Each secondary review group generates a review thread. Based on the tertiary review groups contained in each secondary review group, the review thread corresponding to each secondary review group is split into sub-threads that correspond one-to-one with each tertiary review group. Based on the association between the tertiary review groups, the association between the corresponding sub-threads is established.

[0024] The asynchronous call framework assigns a review model and a review engine to each sub-thread. The review engine controls the execution of the corresponding sub-thread according to a preset review order until the corresponding review thread is completed. When the target sub-thread is executed, sub-threads that are related to the target sub-thread are also triggered to be executed. The target sub-thread can be any sub-thread.

[0025] By adopting the above technical solution, after completing the grouping of secondary review groups, each secondary review group is further classified into tertiary review groups based on the types of files that need to be called when the review points contained in each secondary review group are executed. That is, it is assumed that the review points contained in the tertiary review groups are for the review of the same file. Accordingly, the review thread corresponding to each secondary review group is split into several sub-threads. Each sub-thread corresponds one-to-one with a file, and all sub-threads corresponding to the same file are bound (i.e., an association relationship is established). Whenever any sub-thread is executed and the corresponding file is loaded, all sub-threads corresponding to that file are executed in parallel. In this way, sub-threads across groups are bound and executed synchronously through file association, thereby maximizing the utilization rate of a single file load, achieving the effect of parallel review of a single file load-multiple review model, dynamically balancing file I / O and computing resources, and avoiding resource contention caused by excessive parallelism.

[0026] Optionally, the step of matching a review model to each of the secondary review groups, generating and starting several review threads, and wherein each review thread corresponds one-to-one with a secondary review group, further includes:

[0027] Related sub-threads are categorized into sub-thread lists. The review points contained in each sub-thread within the same sub-thread list are analyzed. Based on the analysis results, the dependencies between all sub-threads in each sub-thread list are determined. The execution mode of all sub-threads in each sub-thread list is determined according to the dependencies, and the execution order of sub-threads with sequential execution mode is determined. The execution mode includes at least parallel execution or sequential execution.

[0028] Based on the execution mode and execution order, a scheduling scheme is generated for each sub-thread list;

[0029] The step of simultaneously triggering the execution of a sub-thread that is related to the target sub-thread when executing the target sub-thread includes:

[0030] When executing the target sub-thread, the execution of sub-threads that are related to the target sub-thread is triggered according to the scheduling scheme corresponding to the target sub-thread.

[0031] By adopting the above technical solution, in the process of collaborative review of sub-threads across groups, the execution order dependency of all sub-threads in the same sub-thread list is dynamically determined, and a scheduling scheme is generated based on the dependency to coordinate the execution order between sub-threads, taking into account both efficiency and correctness, and avoiding conflicts.

[0032] Optionally, the method further includes:

[0033] Analyze the specific review content of all sub-threads within the same sub-thread list, and determine the similarity of the specific review content of the review points contained in different sub-thread lists;

[0034] Establish reuse relationships between review points in different sub-thread lists that meet preset similarity requirements, and identify them as reused review points. Merge reused review points with reuse relationships to form a set of reused review points.

[0035] Analyze the distinguishing and similar review content among the reuse review points in each set of reuse review points, belonging to different sub-thread lists and having reuse relationships. The distinguishing review content includes at least the entity to be reviewed, and the similar content is the same review content in the specific review content corresponding to the reuse review points with reuse relationships.

[0036] A reuse template is generated for each set of reuse review points, and the reuse template includes similar content and differentiating review content, and a reuse model is constructed for executing the reuse template;

[0037] The step of simultaneously triggering the execution of a sub-thread that is associated with the target sub-thread while executing the target sub-thread also includes:

[0038] If the target sub-thread contains reusable review points, the corresponding reusable model is invoked, and the reusable review points are executed through the reusable model according to the differentiated review content corresponding to the target sub-thread; all review points that are not reusable review points contained in the target sub-thread are executed through the review model corresponding to the target sub-thread.

[0039] By adopting the above technical solution, review points with similar review content to those included in other sub-threads (i.e., reused review points) are extracted from sub-threads. Review templates and reuse models are generated separately for reused review points. The common logic (i.e., similar content) in the reuse template is separated from the differentiated review content. The reused review points are executed separately using the reuse model. During this process, only the differentiated review content needs to be switched to review all review points with reuse relationships.

[0040] Optionally, the method further includes:

[0041] Record the review path of the reuse model for the key points of reuse review, and output a natural language interpretation report. The review path includes at least the reuse template content, the corresponding distinguishing review content, and the location information of the distinguishing review content in the corresponding file.

[0042] By adopting the above technical solution, the complete path of reuse review is recorded and an explanation report is output for users to know, thus solving the problem that reuse review results lack interpretability and are difficult to manually review.

[0043] Optionally, the method further includes:

[0044] Receive correction instructions from users and verify user identity; wherein, the correction instructions include at least the reused template to be corrected and the correction content for the corresponding differentiation review content;

[0045] When user authentication is successful, the reuse template contained in the correction instruction is corrected based on the correction instruction.

[0046] By adopting the above technical solutions, a manual review channel is provided to help correct the differential review content in the reuse template, optimize the reuse template, and ensure the accuracy of the reuse model's execution.

[0047] Secondly, this application provides a processing system for improving recognition efficiency based on automatic grouping of key points, including:

[0048] The file recognition and parsing module is used to receive processing instructions from users, parse all the files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains;

[0049] The review point classification module is used to input all review points in the review point list and the corresponding group thinking chain into a preset group intelligence agent, and the group intelligence agent classifies all review points into several primary review groups.

[0050] The multi-threaded synchronization enabling module is used to asynchronously call the framework to assign a review model to each of the first-level review groups, generate and start several review threads, and the review threads correspond one-to-one with the first-level review groups to process all the review points contained in the corresponding first-level review group.

[0051] The processing result output module is used to obtain and merge the review results of all review threads so that users can know the output results.

[0052] Thirdly, this application provides a processing apparatus for improving recognition processing efficiency based on key point automatic grouping, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any of the first aspects.

[0053] Fourthly, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in any of the first aspects.

[0054] In summary, this application includes at least one of the following beneficial technical effects:

[0055] 1. This application aims to solve the problem of low review efficiency in traditional technologies. Specifically, it unifies the entity representation of review points through entity alignment technology, realizes dynamic hierarchical grouping of review tasks by combining grouping mind chain, and finally achieves efficient parallel review by asynchronous multi-threaded call adaptation model.

[0056] 2. Furthermore, this application uses entities as the classification basis to achieve primary grouping of examination points (i.e., forming first-level examination groups), and then further classifies the first-level examination groups according to different examination types to match dedicated examination models and improve examination efficiency. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating a method for improving recognition and processing efficiency based on automatic grouping of key points, as disclosed in an embodiment of this application.

[0059] Figure 2 This is a structural block diagram of a processing system that improves recognition and processing efficiency based on automatic grouping of key points, as disclosed in an embodiment of this application.

[0060] Figure labeling: 201, File recognition and parsing module; 202, Review point classification module; 203, Multi-threaded synchronization enabling module; 204, Processing result output module. Detailed Implementation

[0061] The following is in conjunction with the appendix Figure 1-2 This application will be described in further detail.

[0062] This application discloses a processing method for improving recognition and processing efficiency based on automatic point grouping (hereinafter referred to as the processing method). The processing method is applicable to the scenario of automated review and processing of multiple documents. The executing entity of the processing method is a processing system for improving recognition and processing efficiency based on automatic point grouping (hereinafter referred to as the processing system). The following will be described in conjunction with the appendix. Figure 1 The specific process steps for the system to execute processing methods are described in detail.

[0063] S101 receives the processing instructions from the user, parses all the files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains.

[0064] In implementation, users can access the processing system via a webpage and upload files using the upload button pre-installed on the system's interface. Files can be multiple and of various types, such as plain text and images, thus triggering processing instructions. After receiving all the files included in the processing instructions, the system parses them. This parsing process includes: extracting text and tables from PDF and Word documents using open-source libraries (such as Apache PDFBox and Python's pdfplumber / docx libraries); recognizing documents in images and scanned documents using OCR technology (such as Tesseract and Alibaba Cloud OCR API); and extracting entities (such as names, addresses, and dates) from the extracted content using pre-trained models (such as SpaCy and Alibaba Cloud NLP).

[0065] Next, the processing system determines the specific processing requirements triggered by the user based on the parsing results. Accordingly, the processing requirements can be directly entered by the user when triggering the processing command, or the system can automatically parse the processing requirements based on the pre-stored correspondence between relevant processing requirements and files. For example, if the processing requirement is to open a store, the corresponding required documents include "business license", "lease contract", "hygiene permit", etc.

[0066] In addition, the processing system pre-stores several processing requirements and a list of review points corresponding to each requirement. The list of review points includes several review points, which can be considered as specific review content for a specific entity in the corresponding document, such as checking whether entity B in document A meets the requirements. Therefore, review points correspond to entities, and thus, review points at least include the entity to be reviewed. The grouping mind chain is a pre-defined grouping criterion used to classify all review points in the review point list into several first review groups; such as the COT mind chain.

[0067] Specifically, the processing method also includes the following steps:

[0068] S102, input all review points in the review point list and the corresponding group thinking chain into the preset group intelligence agent, and classify all review points into several first-level review groups through the group intelligence agent;

[0069] S103 uses an asynchronous call framework to assign a review model to each level review group, generate and start several review threads, and the review threads correspond one-to-one with the level review group to process all the review points contained in the corresponding level review group.

[0070] Specifically, S102 and S103 include the following sub-steps:

[0071] All review points in the review point list are aligned with entities. The entity-aligned review points and their corresponding grouped thinking chains are then input into a preset grouped intelligent agent. The grouped intelligent agent categorizes all review points into several first-level review groups, so that different entities correspond to different first-level review groups, thus generating the first mapping relationship between entities and first-level review groups.

[0072] Based on the preset review types, the review points in each first-level review group are classified to generate second-level review groups; among them, the review types include at least plain text review type and visual image review type;

[0073] Analyze the review points contained in each secondary review group, determine the file list corresponding to each secondary review group, and the file list contains the files that need to be called to execute the review points contained in the corresponding secondary review group; establish a second mapping relationship between the review points in the secondary review group and the files in the corresponding file list;

[0074] Based on the second mapping relationship, all review points corresponding to the same document in each secondary review group are taken as associated review points and merged to form a tertiary review group;

[0075] Establish a relationship between Level 3 review groups that belong to different Level 2 review groups but correspond to the same documents;

[0076] Each secondary review group generates a review thread. Based on the tertiary review groups contained in each secondary review group, the review thread corresponding to each secondary review group is split into sub-threads that correspond one-to-one with each tertiary review group. Based on the association between the tertiary review groups, the association between the corresponding sub-threads is established.

[0077] Related child threads are categorized into child thread lists. The review criteria for each child thread within the same list are analyzed. Based on the analysis, the dependencies between all child threads in each child thread list are determined. The execution mode of all child threads in each list is then determined based on these dependencies, and the execution order of child threads with sequential execution mode is determined. The execution mode includes at least parallel execution or sequential execution.

[0078] Based on the execution mode and execution order, a scheduling scheme is generated for each sub-thread list;

[0079] The asynchronous call framework assigns an inspection model and inspection engine to each sub-thread. The inspection engine controls the corresponding inspection model to execute the corresponding sub-thread according to the preset inspection order until the corresponding inspection thread is completed. When the target sub-thread is executed, the scheduling scheme corresponding to the target sub-thread triggers the execution of sub-threads that are related to the target sub-thread.

[0080] In implementation, based on existing entity alignment technology, all review points in the review point list are processed for entity alignment. Then, a grouping agent is used to classify all review points in the review point list into several first-level review groups based on the grouping thought chain. In this embodiment, the grouping logic of the grouping thought chain is as follows: using entities as the grouping basis, review points containing the same entity to be reviewed are classified into the same first-level review group to achieve a one-to-one correspondence between entities and first-level review groups. The corresponding first mapping relationship and the ID of the entity name and its corresponding first-level review group are used to store the entity name and its corresponding first-level review group.

[0081] Furthermore, the processing system determines the review type of each review point based on its specific review content, the corresponding file type, and the entity name. Here, the review type is either plain text or visual image review. For example, if the file type corresponding to a review point is plain text, then its review type is considered plain text; if the file type is image, or the corresponding entity is an image within a file, then its review type is visual image review. The processing system further categorizes all review points within each primary review group into secondary review groups according to the classification logic of "review points of the same review type belong to the same group."

[0082] Furthermore, the processing system analyzes all the documents required for all review points within each secondary review group and generates a document list.

[0083] Based on the number of files in the file list corresponding to each secondary review group, all review points in each secondary review group are further classified into several third review groups according to the classification logic of "review points for the same file are grouped together". The corresponding second mapping relationship is used to store the ID of each third review group and its corresponding file name. Next, the processing system uses the second mapping relationship to establish an association relationship between all third review groups with the same file name. The review points contained in the third review group are the associated review points mentioned above.

[0084] Then, the processing system generates a review thread for each secondary review group. This review thread refers to the execution flow that carries out all the review points contained within the secondary review group. The processing system further divides the review thread into sub-threads corresponding one-to-one with the tertiary review groups contained within each secondary review group. These sub-threads can be considered as executing the execution flow of all the review points within their corresponding tertiary review groups. Furthermore, for tertiary review groups with related relationships, corresponding sub-threads are also associated.

[0085] Next, a pre-built review model is invoked for each sub-thread through a preset asynchronous call framework (such as Celery, Ray). Depending on the review type corresponding to the review points, the review model is also divided into text review model NLP model (such as BERT, FastText) and visual review model CV model (such as YOLO, ResNet) to execute all the review points contained in the corresponding sub-thread.

[0086] Furthermore, since each secondary review group's corresponding review thread includes several sub-threads, in order to control the execution order of all sub-threads, the processing system also configures a review engine for each secondary review group's corresponding review thread. The review engine is used to control the review model to execute all sub-threads contained in the corresponding review thread sequentially according to a preset review order, until all sub-threads have been executed. The preset review order here is a random order predefined by the system, but it can also be customized by the user.

[0087] It should be noted that the processing system will control the synchronous startup of the review engines corresponding to all secondary review groups, so that the review threads of all secondary review groups are executed in parallel. Since each sub-thread of each secondary review group corresponds to an independent review model, all review models will independently execute their corresponding sub-threads under the triggering control of the review engine. Furthermore, if any executed sub-thread (hereinafter referred to as the target sub-thread) is associated with other sub-threads, the processing system will synchronously trigger the review models of all sub-threads associated with the target sub-thread (hereinafter referred to as associated sub-threads) to begin executing the associated sub-threads. That is, currently, within the review thread to which the associated sub-thread belongs, there are multiple sub-threads being executed (i.e., the associated sub-thread and the sub-threads currently being executed by the review engine corresponding to the review thread to which the associated sub-thread belongs). The processing system is used to update the completion progress of each review thread in real time, so that the corresponding review engine can sequentially trigger the execution of the remaining unprocessed sub-threads according to the preset review order.

[0088] Furthermore, the processing system will merge related sub-threads into the same list (i.e., the sub-thread list). As mentioned above, sub-threads in the same sub-thread list will be triggered and executed simultaneously. At this point, considering the possible dependencies between sub-threads in terms of processing logic, such as sub-thread M needing the output of sub-thread N (e.g., sub-thread N extracts text, and sub-thread M performs semantic analysis on the text extracted by sub-thread N), it is necessary to limit the execution mode of dependent sub-threads to sequential execution. The execution order of sub-threads with sequential execution mode will be further determined, while the execution mode of sub-threads without dependencies will be defined as parallel mode. This will ultimately determine the scheduling scheme for all sub-threads in each sub-thread list. The scheduling scheme includes all sub-threads arranged in the order of execution. Finally, when a sub-thread to be executed belongs to any sub-thread list, all sub-threads in the sub-thread list to which the sub-thread belongs will be triggered and executed according to the corresponding scheduling scheme.

[0089] Optionally, the processing methods also include:

[0090] Analyze the specific review content of all sub-threads within the same sub-thread list, and determine the similarity of the specific review content of the review points contained in different sub-thread lists;

[0091] Establish reuse relationships between review points in different sub-thread lists that meet preset similarity requirements, and identify them as reused review points. Merge reused review points with reuse relationships to form a set of reused review points.

[0092] Analyze the distinct review content and similar content among the reuse review points in each set of reuse review points, belonging to different sub-thread lists and having reuse relationships. The distinct review content includes at least the entity to be reviewed, and the similar content is the same review content in the specific review content corresponding to the reuse review points with reuse relationships.

[0093] Generate a reuse template for each set of reuse review points, and the reuse template includes similar content and differentiated review content, and build a reuse model for executing the reuse template;

[0094] When executing a target sub-thread, if the target sub-thread contains reusable review points, the corresponding reusable model is called, and the reusable review points are executed through the reusable model according to the different review content corresponding to the target sub-thread; all review points that are not reusable review points contained in the target sub-thread are executed through the review model corresponding to the target sub-thread.

[0095] In implementation, for example, if the sub-thread list X1 corresponding to file X includes multiple review points (a1, a2, a3); and the sub-thread list Y1 corresponding to file Y includes multiple review points (b1, b2, b3), and the specific review content of (a1, a2, a3) for file X is similar to the specific review content of (b1, b2, b3) for file Y, such as both a1 and b1 extracting text, both a2 and b2 recognizing text content, and both a3 and b3 verifying content clarity; The difference lies not only in the entities being reviewed, but also in the fact that a1 and b1, a2 and b2, and a3 and b3 can be considered as reusable review points with a reuse relationship. The processing system generates a set of reusable review points for each reusable review point with a reuse relationship, such as (a1, b1), (a2, b2), and (a3, b3). In other embodiments, reusable review sets with the same corresponding sub-thread lists can be further merged, such as those corresponding to (a1, b1), (a2, b2), and (a3, b3). The sub-thread lists are all X1 and Y1, so the union of the aforementioned reuse review point sets can be obtained as (a1, b1, a2, b2, a3, b3). Then, a reuse template is generated based on each reuse review point set. The reuse template includes general operation logic (i.e., the specific review content that is common and the same to reuse review points with reuse relationship, i.e., the similar content mentioned above, such as the operation of extracting, identifying, and verifying clarity) and distinguishing review content. The distinguishing review content can specifically be the entity to be reviewed (such as the specific text content extracted from the text described in a1 and the specific text content extracted from the text described in b1). The processing system is used to determine the similar content and distinguishing review content between reuse review points with reuse relationship, generate reuse templates, establish the correspondence between distinguishing review content and reuse review points, and construct a reuse model according to the general operation logic contained in the reuse template. Then, the distinguishing review content is used as input parameters to input the reuse model, and the reuse model is used to execute the specific review content corresponding to the reuse review point.

[0096] When any sub-thread to be executed (i.e., the target sub-thread mentioned above) contains reusable review points, all review points contained in the target sub-thread are further divided into reusable review points and non-reusable review points. The review model is used to execute the specific review content corresponding to the non-reusable review points. At the same time, the reusable model is used to review the specific review content corresponding to the reusable review points. In the process of executing the reusable review points, it is only necessary to input the differentiated review content corresponding to the reusable review points into the reusable model so that the reusable model can review the corresponding differentiated review content according to the general operation logic.

[0097] Optionally, the processing method may also include the following steps:

[0098] Record the review path of the reuse model for the key points of reuse review, and output a natural language interpretation report. The review path includes at least the reuse template content, the corresponding distinguishing review content, and the location information of the distinguishing review content in the corresponding file.

[0099] Receive correction instructions from users and verify user identity; wherein, the correction instructions include at least the reused template to be corrected and the correction content for the corresponding differentiation review content;

[0100] When user authentication is successful, the reuse template contained in the correction instruction is corrected based on the correction instruction.

[0101] In implementation, the processing system is used to record the review path of the reuse model for the key points of reuse review, generate a visual evidence chain (such as highlighting the position of the review content in the corresponding file), and then use LLM (such as GPT-4) to convert the reuse logic into natural language report output. An example output can be "The review result of this clause reuses sub-file A. Since the two match 92%, only the company name field has been adjusted".

[0102] In addition, users can submit a correction instruction and enter their identity information and access key in the correction instruction. The processing system uses the user's identity information and access key to determine whether the user's permissions include correction permissions. If they do, the user's identity is verified and the correction content replaces the difference review content corresponding to the reuse review points.

[0103] S104: Obtain and merge the review results of all review threads for the user to see.

[0104] In implementation, the final review results of all review threads are merged to form the output results, which are then displayed on a preset display interface for users to see. The review results here include at least the determination of whether the review passed or failed, and the determination of whether the review operation was successfully executed.

[0105] Reference Figure 2 This application discloses a processing system for improving recognition efficiency based on automatic grouping of key points, comprising:

[0106] The file recognition and parsing module 201 is used to receive processing instructions from users, parse all the files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains;

[0107] The review point classification module 202 is used to input all review points in the review point list and the corresponding group thinking chain into a preset group intelligence agent, and classify all review points into several first-level review groups through the group intelligence agent;

[0108] The multi-threaded synchronization enabling module 203 is used to asynchronously call the framework to assign a review model to each of the first-level review groups, generate and start several review threads, and the review threads correspond one-to-one with the first-level review groups to process all the review points contained in the corresponding first-level review group.

[0109] The processing result output module 204 is used to obtain and merge the review results of all review threads so that users can know the output results.

[0110] Optionally, the review point classification module 202 is further configured to perform entity alignment processing on all review points in the review point list, input the entity-aligned review points and their corresponding grouped thought chains into a preset grouped intelligent agent, and classify all review points into several first-level review groups through the grouped intelligent agent so that different entities correspond to different first-level review groups, generating a first mapping relationship between entities and first-level review groups; it is also configured to classify the review points in each first-level review group into second-level review groups based on preset review types; wherein, the review types include at least plain text review type and visual image review type;

[0111] The multi-threaded synchronization enabling module 203 is also used to match a review model for each of the secondary review groups, generate and start a number of review threads, and the review threads correspond one-to-one with the secondary review groups.

[0112] Optionally, the review point classification module 202 is further configured to analyze the review points contained in each of the secondary review groups, determine the file list corresponding to each of the secondary review groups, wherein the file list contains the files required to execute the review points contained in the corresponding secondary review group; establish a second mapping relationship between the review points in the secondary review group and the files in the corresponding file list; according to the second mapping relationship, take all review points in each of the secondary review groups that correspond to the same file as associated review points, and merge them to form a tertiary review group; establish an association relationship between tertiary review groups that belong to different secondary review groups and correspond to the same file;

[0113] The multi-threaded synchronization enabling module 203 is also used to generate review threads for each secondary review group. Based on the tertiary review groups included in each secondary review group, the review thread corresponding to each secondary review group is split into sub-threads corresponding to each tertiary review group. Based on the association between the tertiary review groups, the association between the corresponding sub-threads is established. Through an asynchronous call framework, a review model and a review engine are allocated to each sub-thread. The review engine controls the execution of the corresponding sub-thread according to a preset review order until the corresponding review thread is completed. When the target sub-thread is executed, the execution of sub-threads that are associated with the target sub-thread is triggered simultaneously. The target sub-thread can be any sub-thread.

[0114] Optionally, the multi-threaded synchronization enabling module 203 is further configured to classify related sub-threads into a sub-thread list, analyze the review points contained in each sub-thread within the same sub-thread list, determine the dependency relationships between all sub-threads in each sub-thread list based on the analysis results, determine the execution mode of all sub-threads in each sub-thread list according to the dependency relationships, and determine the execution order for sub-threads whose execution mode is sequential execution; wherein, the execution mode includes at least parallel execution or sequential execution; generate a scheduling scheme for each sub-thread list according to the execution mode and execution order; and further configured to, when executing a target sub-thread, trigger the execution of sub-threads related to the target sub-thread according to the scheduling scheme corresponding to the target sub-thread.

[0115] Optionally, it also includes a reuse review module, used to analyze the specific review content of all review points contained in all sub-threads belonging to the same sub-thread list, and determine the similarity of the specific review content of review points contained in different sub-thread lists; establish reuse relationships between review points in different sub-thread lists that meet preset similarity requirements, and determine them as reuse review points; merge reuse review points with reuse relationships to form a set of reuse review points; and analyze the distinguished review content and similar content among reuse review points belonging to different sub-thread lists and having reuse relationships in each set of reuse review points, wherein the distinguished review content at least includes the entity to be reviewed, and the similar content is the same review content in the specific review content corresponding to reuse review points with reuse relationships; generate a reuse template for each set of reuse review points, wherein the reuse template includes similar content and distinguished review content, and construct a reuse model for executing the reuse template;

[0116] The multi-threaded synchronization enabling module 203 is also used to, if the target sub-thread contains reusable review points, call the corresponding reuse model, execute the reusable review points through the reuse model according to the different review content corresponding to the target sub-thread; and execute all review points of non-reusable review points contained in the target sub-thread through the review model corresponding to the target sub-thread.

[0117] Optionally, it also includes a review path recording module, used to record the review path of the reuse model for the reuse review points and output a natural language interpretation report, wherein the review path includes at least the reuse template content, the corresponding distinguishing review content, and the location information of the distinguishing review content in the corresponding file.

[0118] Optionally, a reuse template correction module is also included, which is used to receive correction instructions from users and verify user identity; wherein, the correction instructions include at least the reuse template to be corrected and the correction content for correcting the corresponding differentiation review content; when user identity verification is successful, the reuse template included in the correction instructions is corrected based on the correction instructions.

[0119] This application also discloses a processing device for improving recognition and processing efficiency based on automatic grouping of key points. The processing device for improving recognition and processing efficiency based on automatic grouping of key points includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed as described above for the efficient processing method for batch files.

[0120] This application also discloses a computer-readable storage medium that stores a computer program that can be loaded by a processor and executed as described above for improving recognition efficiency through automatic grouping based on key points. The computer-readable storage medium includes, for example, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0121] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0122] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit the scope of protection of the application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

Claims

1. A processing method for improving recognition and processing efficiency based on automatic grouping of key points, characterized in that, include: Receive processing instructions from users, parse all files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains; Input all the review points in the review point list and the corresponding group thinking chain into the preset group intelligence agent, and classify all the review points into several first-level review groups through the group intelligence agent; By using an asynchronous call framework, a review model is assigned to each of the first-level review groups, and several review threads are generated and started. Each review thread corresponds one-to-one with a first-level review group to process all the review points contained in the corresponding first-level review group. Retrieve and merge the review results from all review threads so that users can see the results. The review criteria include at least the entity to be reviewed; The process involves inputting all review points from the review point list and their corresponding grouping thought chains into a preset grouping agent. The grouping agent then categorizes all review points into several primary review groups, including: All review points in the review point list are aligned with entities. The entity-aligned review points and their corresponding grouped thinking chains are input into a preset grouped intelligent agent. The grouped intelligent agent classifies all review points into several first-level review groups so that different entities correspond to different first-level review groups, generating the first mapping relationship between entities and first-level review groups. Based on preset review types, the review points in each primary review group are classified to generate secondary review groups; wherein, the review types include at least plain text review type and visual image review type; The asynchronous call framework allocates a review model to each of the first-level review groups, generates and starts several review threads, and each review thread corresponds one-to-one with a first-level review group, including: For each of the secondary review groups, a review model is matched, and several review threads are generated and started, with each review thread corresponding to a secondary review group. The step of inputting all review points in the review point list and their corresponding grouping thought chains into a preset grouping agent, and then using the grouping agent to classify all review points into several primary review groups, further includes: Analyze the review points contained in each of the secondary review groups, determine the file list corresponding to each of the secondary review groups, the file list contains the files that need to be called to execute the review points contained in the corresponding secondary review group; establish a second mapping relationship between the review points in the secondary review group and the files in the corresponding file list; Based on the second mapping relationship, all review points corresponding to the same document in each of the secondary review groups are taken as associated review points and merged to form a tertiary review group; Establish a relationship between Level 3 review groups that belong to different Level 2 review groups but correspond to the same documents; The process involves matching a review model to each of the secondary review groups, generating and starting several review threads, with each review thread corresponding one-to-one with a secondary review group, including: Each secondary review group generates a review thread. Based on the tertiary review groups contained in each secondary review group, the review thread corresponding to each secondary review group is split into sub-threads that correspond one-to-one with each tertiary review group. Based on the association between the tertiary review groups, the association between the corresponding sub-threads is established. The asynchronous call framework assigns a review model and a review engine to each sub-thread. The review engine controls the execution of the corresponding sub-thread according to a preset review order until the corresponding review thread is completed. When the target sub-thread is executed, sub-threads that are related to the target sub-thread are also triggered to be executed. The target sub-thread can be any sub-thread.

2. The processing method for improving recognition efficiency based on automatic grouping of key points according to claim 1, characterized in that, The process of matching a review model to each of the secondary review groups, generating and starting several review threads, with each review thread corresponding one-to-one with a secondary review group, further includes: Related sub-threads are categorized into sub-thread lists. The review points contained in each sub-thread within the same sub-thread list are analyzed. Based on the analysis results, the dependencies between all sub-threads in each sub-thread list are determined. The execution mode of all sub-threads in each sub-thread list is determined according to the dependencies, and the execution order of sub-threads with sequential execution mode is determined. The execution mode includes at least parallel execution or sequential execution. Based on the execution mode and execution order, a scheduling scheme is generated for each sub-thread list; The step of simultaneously triggering the execution of a sub-thread that is related to the target sub-thread when executing the target sub-thread includes: When executing the target sub-thread, the execution of sub-threads that are related to the target sub-thread is triggered according to the scheduling scheme corresponding to the target sub-thread.

3. The processing method for improving recognition efficiency based on automatic grouping of key points according to claim 2, characterized in that, The method further includes: Analyze the specific review content of all sub-threads within the same sub-thread list, and determine the similarity of the specific review content of the review points contained in different sub-thread lists; Establish reuse relationships between review points in different sub-thread lists that meet preset similarity requirements, and identify them as reused review points. Merge reused review points with reuse relationships to form a set of reused review points. Analyze the distinguishing and similar review content among the reuse review points in each set of reuse review points, belonging to different sub-thread lists and having reuse relationships. The distinguishing review content includes at least the entity to be reviewed, and the similar content is the same review content in the specific review content corresponding to the reuse review points with reuse relationships. A reuse template is generated for each set of reuse review points, and the reuse template includes similar content and differentiating review content, and a reuse model is constructed for executing the reuse template; The step of simultaneously triggering the execution of a sub-thread that is associated with the target sub-thread while executing the target sub-thread also includes: If the target sub-thread contains reusable review points, the corresponding reusable model is invoked, and the reusable review points are executed through the reusable model according to the differentiated review content corresponding to the target sub-thread; all review points that are not reusable review points contained in the target sub-thread are executed through the review model corresponding to the target sub-thread.

4. The processing method for improving recognition efficiency based on automatic grouping of key points according to claim 1, characterized in that, The method further includes: Record the review path of the reuse model for the key points of reuse review, and output a natural language interpretation report. The review path includes at least the reuse template content, the corresponding distinguishing review content, and the location information of the distinguishing review content in the corresponding file.

5. The processing method for improving recognition efficiency based on automatic grouping of key points according to claim 1, characterized in that, The method further includes: Receive correction instructions from users and verify user identity; wherein, the correction instructions include at least the reused template to be corrected and the correction content for the corresponding differentiation review content; When user authentication is successful, the reuse template contained in the correction instruction is corrected based on the correction instruction.

6. A processing system for improving recognition efficiency based on automatic grouping of key points, applied to the processing method for improving recognition efficiency based on automatic grouping of key points as described in claim 1, characterized in that, The system includes: The file recognition and parsing module (201) is used to receive processing instructions from users, parse all the files contained in the processing instructions to obtain a list of review points and corresponding grouped thought chains; The review point classification module (202) is used to input all review points in the review point list and the corresponding group thinking chain into a preset group intelligence agent, and classify all review points into several first-level review groups through the group intelligence agent; The multi-threaded synchronization enabling module (203) is used to allocate a review model to each of the first-level review groups through asynchronous call framework, generate and start a number of review threads, and the review threads correspond one-to-one with the first-level review groups to process all the review points contained in the corresponding first-level review group; The processing result output module (204) is used to obtain and merge the review results of all review threads so that users can know the output results.

7. A processing device for improving recognition and processing efficiency based on automatic grouping of key points, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 5.

Images