Multi-model scheduling and structure semantic fusion document review method and device

Abstract

The invention provides a multi-model scheduling and structural semantic fusion document review method, which comprises the following steps of: rearranging contents in a to-be-reviewed document on the basis of a semantic hierarchical structure of the to-be-reviewed document to obtain a structured recombined document; constructing a review task graph of the to-be-reviewed document; and configuring a task route for each review task in the review task graph, executing the plurality of review tasks based on the review task graph configured with the task routes, and determining a review result of the structured recombined document. In the face of a large number of dense document scenes, the automatic document review process is achieved, the review efficiency is effectively improved, meanwhile, the accuracy and consistency of review results are guaranteed by means of structured processing and accurate model calling, and the requirements of services for high efficiency, high accuracy and high consistency are fully met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a document review method and device based on multi-model scheduling and structure semantic fusion. BACKGROUND

[0002] With the continuous growth of compliance management and document flow in information-intensive industries, the demand for document review is increasing.

[0003] The review process of the document in the existing method is still highly dependent on manual implementation, and the document review method relying on manual completion in the existing method is low in efficiency. In the face of a large number of intensive document scenarios, it is difficult to meet the business needs of high efficiency, high accuracy and high consistency. SUMMARY

[0004] The present application provides a document review method and device based on multi-model scheduling and structure semantic fusion, to solve the defect that the document review method relying on manual completion in the prior art is low in efficiency, and to realize the business needs of high efficiency, high accuracy and high consistency in the face of a large number of intensive document scenarios.

[0005] The present application provides a document review method based on multi-model scheduling and structure semantic fusion, comprising the following steps: Based on the semantic hierarchical structure of the document to be audited, the content in the document to be audited is rearranged to obtain a structured reorganized document; A review task graph of the document to be audited is constructed, the review task graph is constructed based on a plurality of review tasks of the document to be audited as nodes, and an upstream and downstream dependency relationship between each review task as an edge; Each review task in the review task graph is configured with a task routing, and based on the review task graph configured with the task routing, the plurality of review tasks are executed to determine the review result of the structured reorganized document, the task routing is used to call the review model corresponding to the review task from the model library and obtain the content segment required for review by the review task from the structured reorganized document in the process of review task execution.

[0006] According to the document review method based on multi-model scheduling and structure semantic fusion provided by the present application, the determination process of the task routing comprises: The content segment required for review by the review task is determined from the structured reorganized document; The semantic feature analysis of the content segment is performed to determine the semantic analysis result of the content segment, and based on the semantic analysis result, the review model matched with the semantic analysis result is determined from the model library, and the execution path information of the matched review model is determined; Construct the task routing based on the index information of the content segment and the execution path information.

[0007] According to the document review method provided by the application, the review task graph of the to-be-reviewed document is constructed, which includes: Based on the task target of the to-be-reviewed document, a plurality of review tasks of the to-be-reviewed document are determined. A directed acyclic graph of the plurality of review tasks is constructed by taking the plurality of review tasks as nodes and the upstream and downstream dependency relationship between the review tasks as edges, and the directed acyclic graph is taken as the review task graph.

[0008] According to the document review method provided by the application, after determining the review result of the structured reorganized document, the method further includes: Determine the mapping relationship between the to-be-reviewed document and the structured reorganized document. Based on the mapping relationship, determine the marking position of the review result in the to-be-reviewed document. Based on the marking position, mark the review result in the to-be-reviewed document.

[0009] According to the document review method provided by the application, the model library includes a large language model, a fine-tuning review model, a rule-based symbolic reasoning model, and a structure analysis model.

[0010] According to the document review method provided by the application, the review task graph configured with the task routing is used to execute the plurality of review tasks, which further includes: During the execution of the review task, two review models of the same category are called from the model library to review the content segment required by the review task, and a first output result and a second output result are determined. Determine the deviation of the first output result and the second output result, and in the case that the deviation is less than a preset deviation threshold, take the first output result as the review result of the review task. In the case that the deviation is greater than or equal to the preset deviation threshold, an artificial review process is triggered.

[0011] According to the document review method provided by the application, after determining the review result of the structured reorganized document, the method further includes: Obtain a feedback result of the review result. Based on the feedback result, adjust the node weight in the review task graph.

[0012] According to the multi-model scheduling and structure semantic fusion document review method provided by the application, after the feedback result of the review result is obtained, the following steps are further included: Based on the feedback result, a user review preference vector is constructed; The user review preference vector is embedded into a prompt word template to which a review model in the model library belongs.

[0013] The application further provides a multi-model scheduling and structure semantic fusion document review device, which includes the following modules: a reorganization module, configured to rearrange the content in a document to be audited based on the semantic hierarchical structure of the document to be audited, to obtain a structured reorganized document; A review task graph construction module is configured to construct a review task graph of the document to be audited, the review task graph being constructed based on a plurality of review tasks of the document to be audited as nodes and an upstream and downstream dependency relationship between the review tasks as edges; A review module is configured to configure a task route for each review task in the review task graph, and based on the review task graph with the configured task route, execute the plurality of review tasks, determine a review result of the structured reorganized document, and the task route is used to call a review model corresponding to a review task from a model library and obtain a content segment required to be reviewed by the review task in the process of review task execution.

[0014] The application further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the multi-model scheduling and structure semantic fusion document review method as described above when executing the program.

[0015] The multi-model scheduling and structure semantic fusion document review method and device provided by the application rearrange the content based on the semantic hierarchical structure of the document to be audited to obtain a structured reorganized document, construct a review task graph with review tasks as nodes and a dependency relationship between tasks as edges, and configure a task route for each review task, so that the corresponding review model can be accurately called and the required content segment can be obtained when the review task is executed. In the face of a large number of intensive document scenarios, an automated document review process is realized, the review efficiency is effectively improved, and the accuracy and consistency of the review result are guaranteed by means of structured processing and accurate model calling, fully meeting the needs of high efficiency, high accuracy and high consistency of business. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0017] Figure 1 Figure 1 is a flowchart of a document review method provided by the present application.

[0018] Figure 2 Figure 2 is a schematic diagram of a review process provided by the present application.

[0019] Figure 3 Figure 3 is a structural diagram of a document review device provided by the present application.

[0020] Figure 4 Figure 4 is a structural diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0021] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0022] With the continuous growth of compliance management and document flow in information-intensive industries, the document review method relying on manual work in related methods has been difficult to meet the business needs of high efficiency, high accuracy and high consistency. There are the following problems in various review tasks in related methods: On the one hand, the complexity of document structure and the degree of content heterogeneity continue to improve. Documents such as official documents, contracts, technical manuals, and compliance reports usually have highly structured chapter levels, cross-references, numbered clauses, and term definitions. At the same time, there are nonlinear information such as charts, footnotes, and annotations. Traditional NLP models or rule-based text analysis tools are prone to structure recognition errors and semantic decoupling problems when processing such "structured-weak semantic-high density" documents, which further leads to review blind spots.

[0023] On the other hand, although the current mainstream large language model has general text understanding and generation capabilities, it is limited in long document processing, multi-task review concurrency, domain term consistency judgment and other tasks, and its reasoning path is black-boxed, making it difficult to meet the use requirements of high audit transparency scenarios. At the same time, the single model architecture often has uneven generalization ability and high misjudgment rate when facing the diversity of review tasks, and lacks controllability and explainability.

[0024] In addition, most existing document review systems are tool-based or single-point capability modules, such as spelling checking, semantic polishing, format checking, etc., which are difficult to implement complete review logic chain management and do not have the ability to model task dependencies, review strategy orchestration and task scheduling. The system output is usually based on natural language generation results, and lacks structured, traceable and integrated review suggestion data, which hinders the connection with automatic flow processes or audit management platforms.

[0025] In the specific review process, the following defects exist: Insufficient document structure analysis capability: the document analysis method in the related method cannot accurately restore the logical structure and hierarchical semantics of the document, making it difficult to support fragment understanding and reasoning boundary control of long document content; Lack of strategic arrangement of review tasks: the related method cannot schedule multi-dimensional review goals in a flowchart, and the review tasks lack dependency modeling and upstream and downstream input-output constraints, making it difficult to form a complex logic flow of "term extraction first, then compliance verification"; Single model reasoning capability is limited: facing semantic alignment, term consistency, logical consistency, compliance, expression style and other multi-review tasks, it is difficult for a single general model to cover all task requirements, and it cannot dynamically adapt to content complexity and task priority; Results are not explainable and difficult to structure and integrate: the model output in the related method is mostly natural language suggestions, lacking location anchors, trigger logic, confidence evaluation and other structural meta-information, making it difficult to trace, review and standardize the output results.

[0026] To overcome the defects in the related method, the present application provides a document review method based on multi-model scheduling and structural semantic fusion, Figure 1 The flowchart of the document review method based on multi-model scheduling and structural semantic fusion provided by the present application is shown in Figure 1 As shown in the figure, the method comprises the following steps: Step 110: rearranging the content in the document to be reviewed based on the semantic hierarchical structure of the document to be reviewed to obtain a structured and reorganized document; In step 120, a review task graph of the to-be-reviewed document is constructed, the review task graph is constructed based on a plurality of review tasks of the to-be-reviewed document as nodes and an upstream and downstream dependency relationship between each review task as an edge; In step 130, a task route is configured for each review task in the review task graph, and based on the review task graph with the configured task route, the plurality of review tasks are executed to determine a review result of the structured reorganized document, the task route is used to call a review model corresponding to the review task from a model library and obtain a content segment required to be reviewed by the review task from the structured reorganized document in the process of review task execution.

[0027] The execution subject of the document review method provided by the multi-model scheduling and structure semantic fusion can be an electronic device, a component in the electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Illustratively, the mobile electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), and the non-mobile electronic device can be a server, a network attached storage (NAS), or a personal computer (PC), and the present application is not limited in this regard.

[0028] The technical solutions of the present application will be described in detail below with the computer executing the document review method provided by the multi-model scheduling and structure semantic fusion as an example.

[0029] In step 110, the content in the to-be-reviewed document is rearranged based on the semantic hierarchical structure of the to-be-reviewed document to obtain a structured reorganized document.

[0030] The to-be-reviewed document can be obtained by structurally analyzing the input original document (supporting PDF, Word, scanned copy, etc.) to obtain a text form to-be-reviewed document.

[0031] After obtaining the text form to-be-reviewed document, the to-be-reviewed document is structurally analyzed based on the semantic hierarchical structure to reconstruct the document hierarchical structure corresponding to the semantics and format, and a structured reorganized document convenient for model understanding is output.

[0032] The specific rearrangement process can be implemented based on a layout analysis submodule and a semantic unit generation submodule. The layout analysis submodule: through paragraph extraction, indentation analysis, font size hierarchy, numbering rules, etc., the chapter level, clause structure, table range and appendix reference of the document are identified. A visual language model (such as LayoutLM, Donut, etc.) is used to jointly model the text and spatial layout, improving the positioning and semantic binding accuracy of multi-modal elements such as tables and diagrams.

[0033] The semantic unit generation submodule: further divides semantic units (such as term definition segments, responsibility description segments, conclusion statement segments, etc.) within the structural framework, and generates unit-level vector embedding representations in combination with the context, for subsequent model reasoning and task assignment. Each unit is bound to a unique identifier, with properties such as location anchor, content summary, source structure path, etc.

[0034] It should be noted that based on the semantic hierarchy structure, the content in the document to be reviewed is rearranged, which realizes the parsing of multi-format documents into logical structures with hierarchical and semantic boundaries. The structured reorganized document is the basis for subsequent review task scheduling and model distribution, and its advantage is that it can accurately crop the content boundaries and context windows in long documents.

[0035] In step 120, a review task graph of the document to be reviewed is constructed, wherein the review task graph is constructed based on a plurality of review tasks of the document to be reviewed as nodes, and an upstream and downstream dependency relationship between each review task as an edge.

[0036] Step 120 is to convert the review target into an executable task dependency graph, realizing the mapping and conversion of "multi-dimensional review target → logical flow node → multi-model scheduling strategy".

[0037] Specifically, according to the pre-defined rule library, industry template or user specified parameters, the required review types (such as term consistency check, compliance element check, logical clause analysis, style specification prompt, etc.) are automatically identified, and a plurality of review tasks are output.

[0038] The review task items can be constructed as a directed acyclic graph (DAG), each node representing a review subtask, and the edge representing the upstream and downstream dependency relationship between the tasks, such as "term extraction → term alignment check", "responsibility subject extraction → right and responsibility parity verification", etc. The task graph supports strategy label (priority, sensitivity, resource level) labeling for subsequent scheduling.

[0039] The task graph is parsed at runtime, the input segment is automatically positioned according to the document semantic structure graph, a "subtask-input subset" mapping table is formed, and a scheduling strategy is dynamically generated in combination with the system resource state, model availability, execution history.

[0040] In step 130, task routing is configured for each review task in the review task graph, and based on the review task graph with the configured task routing, the plurality of review tasks are executed to determine the review result of the structured reorganized document, the task routing being used to call the review model corresponding to the review task from the model library and obtain the content segment to be reviewed by the review task from the structured reorganized document in the process of review task execution.

[0041] Each review task in the review task graph is analyzed, and task routing is configured for each review task in the review task graph. The task routing is used to guide the task execution in the process of task execution.

[0042] The specific guidance includes guiding the selection of the corresponding review model from the model library and determining the content segment to be reviewed from the structured reorganized document. It can be understood that after the task routing is determined, in the process of executing the task based on the review task graph, the corresponding review model can be automatically selected from the model library based on the guidance of the task routing, and the content segment to be reviewed can be determined from the structured reorganized document to achieve the acquisition of necessary tools and information in the process of review task execution.

[0043] The model library is pre-constructed and contains models of a plurality of review analysis processes. The specific models included can be a general language large model, a fine-tuned review model processing, a rule-based symbolic reasoning model, and a model execution scheduler.

[0044] The fine-tuned review model processing can be a contract model and a law model. The rule-based symbolic reasoning model can be used to process review processes such as responsibility attribution rules and key term missing rules.

[0045] In the process of auditing the to-be-audited document, the review task graph with the configured task routing is called to implement. According to the nodes and relationships in the review task graph, the plurality of review tasks in the review task graph are executed in sequence. In the process of executing each review task in the review task graph, the review model corresponding to the review task is called from the model library and the content segment to be reviewed by the review task is obtained from the structured reorganized document.

[0046] After all the nodes in the review task graph are reviewed, the review result of the structured reorganized document is obtained. The review result can be used as the final review result of the to-be-audited document.

[0047] Optionally, the overall implementation flowchart can be as follows Figure 2The review process provided by the application is shown in the implementation schematic diagram. Starting from document parsing and structural semantic modeling, the document is parsed and its structural semantic model is constructed; then entering the review task identification and task graph construction link, the specific review task is determined and the review task graph is constructed; then multi-model scheduling and concurrent review execution are performed, multiple models are scheduled and concurrent review is executed; then review result aggregation and structured output are performed, the review results are integrated and output in a structured form; then the user feedback record and review strategy learning step is entered, the user feedback is recorded and the review strategy is learned accordingly; finally, the system deployment and security control process is completed, and the system implementing the method is deployed and security control is implemented.

[0048] The multi-model scheduling and structural semantic fusion document review method provided by the application rearranges the content of the document to be reviewed according to the semantic hierarchical structure to obtain a structured reorganized document, constructs a review task graph with review tasks as nodes and dependency relationships between tasks as edges, and configures a task route for each review task. When executing the review task, the corresponding review model can be accurately called and the required content segment can be obtained. In the face of a large number of intensive document scenarios, an automated document review process is realized, which effectively improves the review efficiency, and at the same time, the accuracy and consistency of the review results are guaranteed by means of structured processing and accurate model calling, fully meeting the needs of business for high efficiency, high accuracy and high consistency.

[0049] In one embodiment, the determination process of the task route includes: determining the content segment that needs to be reviewed by the review task from the structured reorganized document; performing semantic feature analysis on the content segment to determine the semantic analysis result of the content segment, and determining the review model matched with the semantic analysis result from the model library based on the semantic analysis result, and determining the execution path information of the matched review model; based on the index information of the content segment and the execution path information, the task route is constructed.

[0050] From the structured reorganized document obtained by rearranging according to the semantic hierarchical structure, the specific content segment required by each review task is determined. According to the specific requirements of the review task, the text range closely related to the task can be identified in combination with the structured information of the document, so as to ensure that the extracted content segment can completely cover the key information required for review.

[0051] After obtaining the content segment, comprehensive semantic feature analysis is performed thereon. Based on natural language processing technology, the content segment can be analyzed from multiple dimensions such as vocabulary, grammar and semantic relationship, and then the semantic analysis result of the content segment is determined.

[0052] Based on the above semantic analysis result, a review model matching the semantic analysis result is selected from a pre-constructed model library. The model library stores a plurality of review models for different types of semantic features and review requirements, each model having specific functions and application scope. By comparing the semantic analysis result of the content segment with the input requirements and functional characteristics of each review model, the most suitable review model for processing the content segment can be accurately found for accurate and efficient review of the content segment. Meanwhile, the execution path information of the matching review model is determined.

[0053] In combination with the index information of the content segment in the structured reorganized document and the determined execution path information, a task routing is constructed. The index information can quickly locate the position of the content segment in the document, facilitating quick access and reference to the document in the subsequent review process; the execution path information provides a clear route for the processing flow of the review model. The task routing is constructed by organically combining the two, so that during the execution of the review task, the corresponding review model can be quickly and accurately called from the model library according to the task routing, and the content segment is processed according to the established execution path, realizing efficient and automated execution of the review task, and greatly improving the efficiency and quality of document review.

[0054] In one embodiment, a review task graph of the document to be audited is constructed, including: determining a plurality of review tasks of the document to be audited based on the task target of the document to be audited; constructing a directed acyclic graph of the plurality of review tasks with the plurality of review tasks as nodes and the upstream and downstream dependency relationships between the review tasks as edges, and taking the directed acyclic graph as the review task graph.

[0055] The review task items are constructed into a directed acyclic graph (DAG), each node representing a review subtask, and the edges representing the upstream and downstream dependency relationships between the tasks, such as "term extraction → term alignment check", "responsibility subject extraction → responsibility parity verification", etc. The task graph supports policy label (priority, sensitivity, resource level) labeling for subsequent scheduling.

[0056] In one embodiment, after determining the review result of the structured reorganized document, the method further includes: determining a mapping relationship between the document to be audited and the structured reorganized document; determining a marked position of the review result in the document to be audited based on the mapping relationship; and marking the review result in the document to be audited based on the marked position.

[0057] To determine the mapping relationship between the to-be-audited document and the structured reorganized document. Since the structured reorganized document is obtained by rearranging the content based on the semantic hierarchy of the to-be-audited document, the text order, paragraph structure, etc. of the document may change during the rearrangement process, but there is still an inherent logical correspondence between the original document and each content element in the reorganized document. By deeply analyzing the conversion rules, semantic associations, and unique identifiers of content elements in the document processing process, etc. information, the mapping relationship can be accurately established.

[0058] Based on the determined mapping relationship, further determine the marking position of the review result in the to-be-audited document. The review result is obtained for a specific content segment in the structured reorganized document, and these content segments have their original positions in the to-be-audited document. Using the mapping relationship established earlier, the position information of the content segment corresponding to the review result in the structured reorganized document can be accurately converted to the position information in the to-be-audited document, thereby determining the specific marking position of the review result in the to-be-audited document.

[0059] Based on the determined marking position, mark the review result in the to-be-audited document. The marking method can be selected according to actual needs, for example, high-light display, adding comments, inserting special markers, etc. can be used to intuitively display the review result at the corresponding position of the to-be-audited document.

[0060] In one embodiment, the model library includes a large language model, a fine-tuned review model, a rule-based symbolic reasoning model, and a structure analysis model.

[0061] The model library contains multiple models. The large language model, with its powerful language understanding and generation capabilities accumulated through massive data training, can deeply grasp the semantic connotation, contextual logical relationship, and overall context information of the document text, and performs outstandingly in processing natural language description-rich and semantically complex document content, such as understanding viewpoints, emotional tendencies, etc. in long dissertations, providing a comprehensive semantic foundation for review.

[0062] The fine-tuned review model is a fine-tuned and optimized model for specific review tasks or fields. It integrates knowledge and review rules in specific fields and can more accurately identify common errors, problems or key information in documents in that field, such as improper use of legal terminology, logical contradictions in clauses, etc. in legal document review, to improve the professionalism and accuracy of review.

[0063] The rule-based symbolic reasoning model makes inference and judgment on the document according to the pre-set explicit rules and logical relations, is suitable for processing document content with strict format specification and clear logical structure, such as mathematical formula derivation and programming code logic verification, and can ensure the rigor and reliability of the review of such documents by strictly following the rules for reasoning.

[0064] The structure analysis model is used for analyzing the structural hierarchy of the document, can accurately identify the structural elements such as title, paragraph and chapter of the document, understand the subordinate and parallel relationship between the parts, provide a clear structural framework for review, and help the reviewer to grasp the overall architecture of the document from a macro perspective, and also assist other models to perform more accurate review within a specific structural unit.

[0065] In one embodiment, the performing the plurality of review tasks according to the review task graph configured with the task routing further includes: during the execution of the review task, calling two review models of the same category from the model library to review the content segment required by the review task respectively, determining a first output result and a second output result; determining a deviation of the first output result and the second output result, in a case where the deviation is less than a preset deviation threshold, taking the first output result as a review result of the review task; in a case where the deviation is greater than or equal to the preset deviation threshold, triggering a manual review process.

[0066] In the actual execution process of the document review task, a multi-model collaboration and deviation checking mechanism is adopted to ensure the accuracy and reliability of the review result. Specifically, when the review task starts running, two review models of the same category are selected from the model library with rich functions and various types. Although the two models belong to the same category, they may be based on different training data, algorithm optimization strategies or parameter settings. They will respectively carry out independent review on the content segment required by the current review task, and respectively output a first output result and a second output result by using their own algorithm logic and knowledge system for in-depth analysis of the content.

[0067] The two output results are carefully calculated and evaluated for deviation. The deviation calculation aims to measure the consistency between the two model output results, and quantifies the differences in viewpoints, conclusions, scores, etc. In this process, a deviation threshold is pre-set, which is determined based on a large amount of historical data, business requirements and review accuracy requirements. When the calculated deviation is less than the pre-set deviation threshold, it means that the output results of the two models are highly consistent and have high credibility. At this time, the first output result is taken as the final review result of the review task, ensuring the efficient progress of the review process.

[0068] When the deviation is greater than or equal to the preset deviation threshold, it indicates that the output results of the two models are greatly different, and there may be model understanding deviation, data anomaly, or task complexity that causes the model to be difficult to reach an agreement. An artificial review process is triggered.

[0069] Through this multi-model cooperation and deviation checking combined with artificial intervention mechanism, the accuracy and consistency of the review results can be maximized while ensuring the review efficiency, meeting the needs of various complex document review tasks.

[0070] In an embodiment, after determining the review result of the structured reorganized document, the method further includes: obtaining a feedback result of the review result; and adjusting a node weight in the review task graph based on the feedback result.

[0071] After obtaining the feedback result, the node weight in the review task graph is adjusted based on the feedback result. Each node in the review task graph represents a review task, and the node weight reflects the importance of the review task in the entire review process, the execution difficulty, or the influence on the final review result.

[0072] For example, for a review task on the critical path of the review process, even if the feedback result shows that the review result quality is high, it may still need to maintain a high weight because it has an important influence on the execution of subsequent tasks, to ensure the stability and reliability of the entire review process.

[0073] In an embodiment, after obtaining the feedback result of the review result, the method further includes: constructing a user review preference vector based on the feedback result; and embedding the user review preference vector into a prompt word template to which a review model in the model library belongs.

[0074] After obtaining the feedback result of the review result, the user's needs are further explored based on the feedback to construct a user review preference vector, which can accurately represent the user's individualized inclination in review, such as the degree of attention to different types of errors, the emphasis on specific field content, etc. Then, the constructed user review preference vector is embedded into the prompt word template corresponding to the review model in the model library, so that the model can combine user preferences for more user demand-oriented processing in subsequent review, improving the fit between the review result and the user's expectations.

[0075] Optionally, in order to improve the adaptability of the model to individual review preferences, a review preference representation model can be constructed by periodically aggregating user feedback behavior samples. This process consists of three levels: The term preference model: extract historical samples of user modification and retention behavior for specific terms, establish a "term -> replacement strategy" vector index, and use it to construct a term dictionary weight; Expression style model: analyze the user's preference for the passive / active, sentence structure(long / short), conjunction usage, and text / annotation style during the suggestion modification process, and construct a syntactic vector space projection; Review strategy tolerance model: statistics on the user's acceptance rate and confidence tolerance interval for different types of suggestions(such as ambiguous expression, logical jump, missing reference), and establish a personalized confidence threshold mapping table.

[0076] Feedback data is not only used for training models, but also for reconstructing "review task graphs" and optimizing review process strategies. The following analysis is performed on feedback samples: Task redundancy determination: if the user continuously ignores a certain type of suggestion(such as "sentence optimization" suggestions), the task node can be marked as a "low weight node" or "optional execution node" according to the feedback frequency, reducing the burden on the model; Dependence path adjustment: if the user often makes reverse modifications to downstream task suggestions after a certain task output(such as opposing term application suggestions after term explanation), automatically adjust the edge weight in the task graph or adjust the node execution order; Task node fusion / splitting: for users who frequently repeat the same type of feedback in multiple suggestion items, extract common behavior logic and generate new task node templates to achieve "semantic similar suggestion aggregation".

[0077] According to the high-confidence feedback samples collected in the feedback pool(such as the user's modified content and the semantic distance of the original suggestion), the samples are added to the model training data. Two continuous learning paths are supported: Fine-tuning training path: use LoRA or QLoRA to fine-tune the large model, only update the specific task layer or expression style generation header, support periodic version rollback and version difference evaluation; Prompt optimization path: feedback samples are used to strengthen the Prompt template, such as adding "user preference expression style constraints", "emphasizing term explanation trace references", "reducing style suggestion output", etc. to improve the robustness of the Prompt.

[0078] With Prompt-A / B testing environment, different Prompt versions can be run simultaneously, and their acceptance rate and misjudgment rate are recorded to automatically converge to the optimal Prompt template combination for the current user / team.

[0079] Considering that feedback data itself may contain document sensitive information, behavior logs are encrypted and stored with permission management. The following security strategies are supported: Users can enable "local feedback mode", all behavior samples are saved on the terminal and do not participate in model training; Support for anonymous feedback processing for aggregated behavior modeling; The administrator can control whether the team model introduces individual feedback samples and set a specific feedback type whitelist (such as only introducing "term replacement suggestion" feedback); All model retraining versions have an auditable version number and rollback mechanism to ensure the traceability and security of behavior learning.

[0080] To ensure stable operation under different security requirements, the application designs a deployment and security control module that plays a global protection role in the overall process: In the initialization phase, users can configure the operation mode, including local deployment, enterprise private cloud, edge node cluster, etc. All model execution modules support local calling (such as using ONNX / LoRA quantization models) or sandbox running to ensure that data does not go online and models cannot be accessed reversely.

[0081] During data transfer, audit logs are automatically generated for each model inference, including calling time, input summary, model version, output result summary, etc. "Model calling traceable link" can be optionally enabled for data security audit and compliance review.

[0082] In addition, "active desensitization strategy" is supported: users can configure which paragraphs, keywords, and meta fields do not participate in inference (such as classified customer names and key code segments), which will be replaced with placeholders for structural review, or trigger only rule engine judgment processing. All data cache storage is encrypted, and storage duration, automatic cleaning, and other strategies can be set.

[0083] The following describes the multi-model scheduling and structural semantic fusion document review device provided by the application. The multi-model scheduling and structural semantic fusion document review device described below can be referred to in conjunction with the multi-model scheduling and structural semantic fusion document review method described above.

[0084] As shown in Figure 3 The device comprises: The reorganization module 310 is configured to rearrange the content in the document to be reviewed based on the semantic hierarchical structure of the document to be reviewed to obtain a structured reorganized document. The review task graph construction module 320 is configured to construct a review task graph of the document to be reviewed, the review task graph being constructed based on a plurality of review tasks of the document to be reviewed as nodes and an upstream and downstream dependency relationship between each review task as an edge; and the review module 330 is configured to configure a task route for each review task in the review task graph, and execute the plurality of review tasks based on the review task graph configured with the task route, determine a review result of the structured reorganized document, and the task route is used to call a review model corresponding to a review task from a model library and obtain a content segment required to be reviewed by the review task in the process of executing the review task.

[0085] The document review device provided by the application can rearrange the content of the document to be reviewed based on the semantic hierarchical structure to obtain a structured reorganized document, construct a review task graph taking review tasks as nodes and a dependency relationship between tasks as edges, and configure a task route for each review task, so that the corresponding review model can be accurately called and the required content segment can be obtained when the review task is executed. In the face of a large number of intensive documents, an automatic document review process is realized, the review efficiency is effectively improved, and the accuracy and consistency of the review result are guaranteed by means of structured processing and accurate model calling, thereby fully meeting the needs of businesses for high efficiency, high accuracy and high consistency.

[0086] In one embodiment, the review module 330 is specifically configured to: The determination process of the task route includes: determining a content segment required to be reviewed by the review task from the structured reorganized document; performing semantic feature analysis on the content segment to determine a semantic analysis result of the content segment, determining a review model matched with the semantic analysis result from the model library based on the semantic analysis result, and determining execution path information of the matched review model; constructing the task route based on the index information of the content segment and the execution path information.

[0087] In one embodiment, the review task graph construction module 320 is specifically configured to: The construction of the review task graph of the document to be reviewed includes: determining a plurality of review tasks of the document to be reviewed based on a task target of the document to be reviewed; constructing a directed acyclic graph of the plurality of review tasks taking the plurality of review tasks as nodes and an upstream and downstream dependency relationship between each review task as an edge, and taking the directed acyclic graph as the review task graph.

[0088] In one embodiment, the review module 330 is specifically configured to: The determining the review result of the structured reorganized document further includes: determining a mapping relationship between the document to be audited and the structured reorganized document; determining a marking position of the review result in the document to be audited based on the mapping relationship; marking the review result in the document to be audited based on the marking position.

[0089] In one embodiment, the review module 330 is further specifically used for: determining that the model library includes a large language model, a fine-tuned review model, a rule-based symbolic reasoning model, and a structure analysis model.

[0090] In one embodiment, the review module 330 is further specifically used for: The executing the plurality of review tasks based on the review task graph configured with the task routing further includes: In the process of executing the review task, calling two review models of the same category from the model library to review the content segment required to be reviewed by the review task, respectively, to determine a first output result and a second output result; determining a deviation of the first output result and the second output result, and in a case where the deviation is less than a preset deviation threshold, taking the first output result as a review result of the review task; in a case where the deviation is greater than or equal to the preset deviation threshold, triggering a manual review process.

[0091] In one embodiment, the review module 330 is further specifically used for: The determining the review result of the structured reorganized document further includes: obtaining a feedback result on the review result; adjusting a node weight in the review task graph based on the feedback result.

[0092] In one embodiment, the review module 330 is further specifically used for: The obtaining a feedback result on the review result further includes: constructing a user review preference vector based on the feedback result; embedding the user review preference vector into a prompt word template to which a review model in the model library belongs.

[0093] Figure 4 An example of an entity structure diagram of an electronic device is shown in FIG. 1. Figure 4As shown, the electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 complete mutual communication through the communications bus 440. The processor 410 can invoke a logical instruction in the memory 430 to execute a multi-model scheduling and structure semantic fusion document review method, which includes: based on a semantic hierarchical structure of a to-be-reviewed document, rearranging content in the to-be-reviewed document to obtain a structured reorganized document; constructing a review task graph of the to-be-reviewed document, the review task graph being constructed based on a plurality of review tasks of the to-be-reviewed document as nodes and an upstream and downstream dependency relationship between the review tasks as edges; configuring a task route for each review task in the review task graph, and based on the review task graph configured with the task route, executing the plurality of review tasks to determine a review result of the structured reorganized document, the task route being used to call a review model corresponding to a review task from a model library and obtain a content segment required to be reviewed by the review task from the structured reorganized document in the process of review task execution.

[0094] In addition, the logical instruction in the memory 430 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0095] On the other hand, the present application also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, when the computer program is executed by a processor, the computer can execute the multi-model scheduling and structure semantic fusion document review method provided by the above-mentioned method, the method includes: based on the semantic hierarchical structure of the to-be-reviewed document, rearranging the content in the to-be-reviewed document to obtain a structured reorganized document; construct a review task graph of the to-be-reviewed document, the review task graph being constructed based on a plurality of review tasks of the to-be-reviewed document as nodes and an upstream and downstream dependency relationship between each review task as an edge; configure a task route for each review task in the review task graph, and based on the review task graph configured with the task route, execute the plurality of review tasks to determine a review result of the structured reorganized document, the task route being used to call a review model corresponding to a review task from a model library and obtain a content segment required to be reviewed by the review task from the structured reorganized document in a process of review task execution.

[0096] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the document review method of multi-model scheduling and structure semantic fusion provided by each method described above, the method comprising: based on a semantic hierarchical structure of a to-be-reviewed document, rearranging content in the to-be-reviewed document to obtain a structured reorganized document; construct a review task graph of the to-be-reviewed document, the review task graph being constructed based on a plurality of review tasks of the to-be-reviewed document as nodes and an upstream and downstream dependency relationship between each review task as an edge; configure a task route for each review task in the review task graph, and based on the review task graph configured with the task route, execute the plurality of review tasks to determine a review result of the structured reorganized document, the task route being used to call a review model corresponding to a review task from a model library and obtain a content segment required to be reviewed by the review task from the structured reorganized document in a process of review task execution.

[0097] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0098] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A document review method integrating multi-model scheduling and structural semantics, characterized in that: include: Based on the semantic hierarchical structure of the document to be reviewed, the contents of the document to be reviewed are rearranged to obtain a structured reorganized document; Constructing a review task graph for the document to be reviewed, wherein the review task graph is constructed based on multiple review tasks for the document to be reviewed as nodes and upstream and downstream dependency relationships between the review tasks as edges; A task route is configured for each review task in the review task graph, and based on the review task graph configured with the task route, the multiple review tasks are executed to determine the review results of the structured reorganized document. The task route is used to call the review model corresponding to the review task from the model library and obtain the content fragments that need to be reviewed for the review task from the structured reorganized document during the execution of the review task.

2. The document review method integrating multi-model scheduling and structural semantics according to claim 1 is characterized in that: The process of determining the task routing includes: determining, from the structured reorganized document, content segments that need to be reviewed for the review task; Performing semantic feature analysis on the content segment to determine a semantic analysis result of the content segment, and based on the semantic analysis result, determining a review model that matches the semantic analysis result from the model library, and determining execution path information of the matching review model; The task route is constructed based on the index information of the content segment and the execution path information.

3. The document review method combining multi-model scheduling and structural semantics according to claim 1 is characterized in that: The step of constructing the review task graph of the document to be reviewed includes: Determining a plurality of review tasks for the document to be reviewed based on the task objectives of the document to be reviewed; With multiple review tasks as nodes and upstream and downstream dependencies between the review tasks as edges, a directed acyclic graph of the multiple review tasks is constructed, and the directed acyclic graph is used as the review task graph.

4. The document review method combining multi-model scheduling and structural semantics according to claim 1 is characterized in that: After determining the review result of the structured reorganized document, the method further includes: Determining a mapping relationship between the document to be reviewed and the structured reorganized document; Based on the mapping relationship, determining the marking position of the review result in the document to be reviewed; Based on the marked position, the review result is marked in the document to be reviewed.

5. The document review method of multi-model scheduling and structural semantic fusion according to claim 1 is characterized in that: The model library includes a large language model, a fine-tuning review model, a rule-based symbolic reasoning model, and a structural analysis model.

6. The document review method of multi-model scheduling and structural semantic fusion according to claim 1 is characterized in that: The step of executing the plurality of review tasks based on the review task graph configured with task routing further includes: During the execution of the review task, two review models of the same category are called from the model library to review the content segments required for review in the review task respectively, and determine a first output result and a second output result; determining a deviation between the first output result and the second output result, and if the deviation is less than a preset deviation threshold, using the first output result as a review result of the review task; When the deviation is greater than or equal to a preset deviation threshold, a manual review process is triggered.

7. The document review method of multi-model scheduling and structural semantic fusion according to claim 1 is characterized in that: After determining the review result of the structured reorganized document, the method further includes: Obtaining feedback on the review results; Based on the feedback results, the node weights in the review task graph are adjusted.

8. The document review method of multi-model scheduling and structural semantic fusion according to claim 7 is characterized in that: After obtaining feedback on the review result, the method further includes: Based on the feedback results, construct a user review preference vector; The user review preference vector is embedded into the prompt word template of the review model in the model library.

9. A document review device integrating multi-model scheduling and structural semantics, characterized in that: include: a reorganization module, configured to rearrange the contents of the document to be reviewed based on the semantic hierarchical structure of the document to be reviewed to obtain a structured reorganized document; A review task graph construction module is used to construct a review task graph for the document to be reviewed, wherein the review task graph is constructed based on multiple review tasks of the document to be reviewed as nodes and upstream and downstream dependencies between the review tasks as edges; The review module is used to configure task routing for each review task in the review task graph, and based on the review task graph configured with task routing, execute the multiple review tasks and determine the review results of the structured reorganized document. The task routing is used to call the review model corresponding to the review task from the model library and obtain the content fragments that need to be reviewed for the review task from the structured reorganized document during the execution of the review task.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the document review method of multi-model scheduling and structural semantic fusion as described in any one of claims 1 to 8 is implemented.

Images