A multi-technology integrated intelligent document processing method and system

Abstract

This invention discloses a multi-technology integrated intelligent document processing method and system for performing artificial intelligence technology. Through multi-technology integration and innovation, a high-precision, low-cost, and professional intelligent document processing system is constructed, solving the problems of low accuracy, high cost, and insufficient professionalism of existing products. It effectively improves the efficiency of judicial document processing, significantly improves the quality of judicial services, and reduces litigation costs. At the same time, the final processed real target document is split into privacy and non-privacy content, and two semi-pseudo-documents are generated, each containing privacy and non-privacy content. The two semi-pseudo-documents and the corresponding real target document are combined to form a double-encrypted hierarchical document library, realizing a double-layer privacy protection for the real target document.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent document processing method and system that integrates multiple technologies. Background Technology

[0002] Currently, document processing products on the market mainly include traditional OCR recognition software (such as ABBYY FineReader and Hanvon OCR) and basic AI document processing tools (such as Baidu Intelligent Document Processing and Tencent Cloud Document Recognition). The main functional structures of these products include: document scanning and recognition modules, text extraction modules, format conversion modules, and simple template matching functions. Traditional OCR products primarily rely on a single recognition engine, achieving text recognition through steps such as image preprocessing, character segmentation, feature extraction, and pattern recognition; AI document processing tools, on the other hand, add basic semantic understanding and information extraction functions to the OCR framework.

[0003] For example, Chinese patent CN120597846B discloses a method and system for automatic generation and multi-dimensional review of standard documents based on a large model. The method includes: establishing a distributed database of multi-source documents and analyzing heterogeneous texts through natural language processing to obtain a standardized knowledge network; extracting indicator elements based on the standardized knowledge network and forming a structured parameter library through verification and validation; constructing a template library based on the structured parameter library, and analyzing user needs by combining semantic matching to automatically generate a standard document outline. This invention improves the efficiency and quality of document generation, reduces the cost of manual review, and enhances the comprehensiveness and accuracy of review.

[0004] However, in the current market environment of accelerated digital transformation and increasing demand for specialization, existing products have significant shortcomings: In terms of technical limitations, single OCR engines are not accurate enough when handling complex layouts, handwritten text, and seal interference, generally below 90%, which cannot meet the accuracy requirements of professional documents such as legal documents; in terms of cost efficiency, there is a lack of intelligent resource scheduling mechanisms, and the cost of using AI computing power is high, which is difficult for small and medium-sized enterprises to afford; in terms of specialization, existing general-purpose products cannot adapt to the special format requirements of the legal industry, such as the element-based complaint standard required by the Supreme People's Court, resulting in a large amount of manual secondary processing; the processing flow is fragmented, with each functional module operating independently, lacking collaborative optimization, and the overall processing efficiency is low.

[0005] Therefore, there is an urgent need to build a high-precision, low-cost, and professional intelligent document processing system through the integration and innovation of multiple technologies to solve the technical bottlenecks of traditional products. Summary of the Invention

[0006] The core of this invention lies in constructing a high-precision, low-cost, and professional intelligent document processing system to solve the problems of low precision, high cost, and insufficient professionalism of existing products. This system improves the efficiency of judicial document processing by more than 80%, significantly improves the quality of judicial services, and reduces litigation costs. Simultaneously, the system separates the processed real target document into privacy and non-privacy content, generating two semi-pseudo-documents, each containing both privacy and non-privacy content. These two semi-pseudo-documents and the corresponding real target document are combined to form a double-encrypted, tiered document library, achieving dual-layer privacy protection for the real target document.

[0007] To solve the above problems, the present invention adopts the following technical solution.

[0008] A multi-technology integrated intelligent document processing method includes the following steps:

[0009] S1. Receive multiple initial documents to be processed through the input layer and preprocess them;

[0010] S2. Perform OCR intelligent recognition on the content of each document through the recognition layer: Select the optimal engine combination through intelligent scheduling algorithm to perform high-precision text recognition and result fusion;

[0011] S3. The analysis layer performs deep semantic analysis and information extraction on the identified content, and intelligently fills in the missing parts to obtain multiple semi-finished document data.

[0012] S4. The semi-finished document data is subjected to quality inspection, logical verification and error correction through the processing layer. Low-quality semi-finished document data is filtered out and included in the manual review queue. The remaining semi-finished document data that meets the quality requirements is optimized and then enters step S5.

[0013] S5. For each semi-finished document data, a target document conforming to the standard format is generated through the output layer to obtain multiple real target documents, and knowledge graphs and metadata annotations are selectively constructed based on the real target documents.

[0014] S6. For each real target document, its content is split into a private content and a non-private content through a privacy protection layer, thereby obtaining multiple private content and multiple non-private content. For the same real target document divided into private content A and non-private content B, another non-private content is randomly selected and combined with private content A to form the first semi-pseudo-document. Then, another private content is randomly selected and combined with non-private content B to form the second semi-pseudo-document.

[0015] S7. Combine each real target document and its two corresponding semi-pseudo-documents into a tiered document library, which serves as the final processing result of the corresponding initial document. Randomly select one of the semi-pseudo-documents as the explicit document and do not perform encryption processing. Select the other semi-pseudo-document as the first-level hidden document, perform first-level encryption processing and generate a first-level key. Select the real document as the second-level hidden document, perform second-level encryption processing and generate a second-level key.

[0016] Furthermore, the process includes the following step S8: when an unauthorized person attempts to obtain the first-level hidden document in the hierarchical document library by illegal means, the second-level hidden document is automatically destroyed and an alarm is triggered; when an authorized person enters the first-level key and the second-level key, the private content A and the non-private content B can be extracted from the two semi-pseudo-documents, and a real document can be synthesized again. Then, step S7 is performed again to generate a new, secure hierarchical document library.

[0017] A multi-technology integrated intelligent document processing system is used to execute the above-mentioned intelligent document processing method. It includes an input layer, a recognition layer, an analysis layer, a processing layer, an output layer, and a privacy protection layer. The input layer includes a document input module, and the recognition layer includes a multi-engine OCR fusion module, which integrates multiple OCR recognition engines.

[0018] Furthermore, the analysis layer includes an AI intelligent analysis module, an intelligent template matching module, and a checkbox processing module. The AI ​​intelligent analysis module achieves intelligent scheduling through a computing power distribution engine and performs deep semantic analysis and information extraction on document content. The intelligent template matching module is used for case type identification, intelligent template matching, and automatic content filling. The checkbox processing module adopts a dual-engine mechanism that combines a rule engine and an AI model to intelligently identify and judge the status of checkboxes.

[0019] Furthermore, the processing layer includes a computing power distribution and scheduling module, a parallel processing module, and a quality control module. The computing power distribution and scheduling module is used to manage various AI resources, select intelligent routes, balance the load, and optimize costs. The parallel processing module realizes the parallel execution of multiple tasks and performance optimization through multi-threading and asynchronous processing technology. The quality control module is used to perform quality checks, logical verification, and error correction on the processing results.

[0020] Furthermore, the output layer includes a document generation module, a semantic understanding and knowledge graph construction module, and a classification and metadata annotation module. The document generation module generates real target documents that conform to standard formats based on the processing results. The semantic understanding and knowledge graph construction module is used to extract key semantic units from the document content and construct a dynamic knowledge graph.

[0021] Furthermore, the privacy protection layer includes a feature segmentation module, a true / false fusion module, a document package generation module, and a multi-level encryption module. The feature segmentation module is used to segment the real target document, the true / false fusion module is used to generate a semi-pseudo-document, the document package generation module is used to combine the real target document and the semi-pseudo-document into a hierarchical document library, and the multi-level encryption module is used to selectively encrypt the data in the hierarchical document library.

[0022] Furthermore, the privacy protection layer also includes a security monitoring and processing module and a secondary generation module. The security monitoring and processing module is used to monitor whether the tiered document library has been viewed illegally and to destroy the second-layer hidden document when necessary. The secondary generation module is used to extract the private content A and non-private content B from the two semi-pseudo-documents and regenerate the second-layer hidden document.

[0023] Compared with the prior art, the advantages of this invention are:

[0024] (1) This solution integrates and innovates multiple technologies to process documents sequentially through input → recognition → analysis → matching → scheduling → parallel processing → processing → quality control → generation → knowledge graph, thus constructing a high-precision, low-cost, and professional intelligent document processing system. This solves the problems of low precision, high cost, and insufficient professionalism of existing products. By improving the efficiency of judicial document processing by more than 80%, it significantly improves the quality of judicial services, reduces litigation costs, and makes it easier for ordinary people to obtain professional legal services. It has important social significance and economic value.

[0025] (2) The real target document is split into privacy and non-privacy content, and two semi-pseudo-documents are generated, each containing privacy and non-privacy content. The two semi-pseudo-documents and the corresponding real target document are combined to form a double-encrypted hierarchical document library. Unauthorized personnel cannot obtain the complete real target document P by obtaining one of the semi-pseudo-documents alone. Even if they obtain the two semi-pseudo-documents, since they do not know which part of the semi-pseudo-document corresponds to the content of document P, it is difficult for unauthorized personnel to accurately merge the two semi-pseudo-documents to obtain the real target document P. Therefore, double privacy protection for the real target document P is achieved. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention;

[0027] Figure 2 This is a block diagram of the system of the present invention;

[0028] Figure 3 This is a block diagram of the privacy protection layer of the present invention;

[0029] Figure 4 This is a flowchart illustrating the use of the privacy protection layer of the present invention;

[0030] Figure 5 This is a block diagram of the hierarchical document library of the present invention. Detailed Implementation

[0031] The technical solutions will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0032] First implementation method:

[0033] Please see Figures 1 to 5 A multi-technology integrated intelligent document processing system includes an input layer, a recognition layer, an analysis layer, a processing layer, and an output layer. This intelligent document processing system enables the following multi-technology integrated intelligent document processing methods.

[0034] A multi-technology integrated intelligent document processing method includes the following steps:

[0035] S1. Receive multiple initial documents to be processed through the input layer and preprocess them;

[0036] Specifically, the input layer includes a document input module that can automatically recognize various document formats such as JPG, PNG, PDF, BMP, and TIFF, and determine their types, such as distinguishing between scanned documents, photographs, and electronic documents. It then performs image enhancement preprocessing on low-quality images, including denoising, sharpening, and contrast adjustment, using an adaptive image enhancement algorithm that dynamically adjusts the preprocessing parameters based on image quality indicators. Furthermore, it can automatically detect document layout structure, identifying areas such as titles, body text, tables, and seals. This process is based on a deep learning layout analysis model with an accuracy rate of 98%. Additionally, the input layer supports batch upload operations, enabling concurrent uploads of multiple documents and queue management, and supports large file resume uploads and progress tracking.

[0037] S2. Perform OCR intelligent recognition on the content of each document through the recognition layer: Select the optimal engine combination through intelligent scheduling algorithm to perform high-precision text recognition and result fusion;

[0038] Specifically: The recognition layer includes a multi-engine OCR fusion module, which integrates multiple OCR recognition engines, such as mainstream engines like Baidu Cloud OCR, PaddleOCR, Tesseract, and Tencent Cloud OCR. It can dynamically select the optimal engine based on factors such as image features, historical performance, and cost budget through intelligent scheduling algorithms. When recognition fails, it automatically adjusts parameters and re-recognizes, up to three times. In addition to the optimal engine, a backup engine is also set up. By calculating the confidence score for the recognition result, when the confidence score is <0.7, the backup engine is automatically activated to perform secondary recognition and merge the results, marking low-confidence areas.

[0039] The intelligent scheduling algorithm is as follows: Score(Engine) = w acc ×Accuracy+w speed ×Speed+w cost ×Cost; where: Score is the final score, Accuracy represents the accuracy rate, Speed ​​represents the running speed, and Cost represents the cost, w acc w speed w cost Each of these has its own weight, which can be dynamically adjusted adaptively based on image quality. For example, when image quality is high, speed is prioritized, such as w. speed =0.5, w acc =0.3, w cost =0.2; When image quality is low, accuracy is prioritized, such as w acc =0.7, w speed =0.1, w cost =0.2.

[0040] S3. The analysis layer performs deep semantic analysis and information extraction on the identified content, and intelligently fills in the missing parts to obtain multiple semi-finished document data.

[0041] Specifically, the analysis layer includes an AI intelligent analysis module, an intelligent template matching module, and a checkbox processing module. The AI ​​intelligent analysis module integrates multiple large model APIs, such as GPT-4, Claude, Wenxin Yiyan, and Tongyi Qianwen, and achieves intelligent scheduling through a computing power distribution engine. It performs deep semantic analysis and information extraction on document content, deeply analyzing document semantics to understand legal terminology, causal relationships, and time series. It structurally extracts key elements such as party information, case facts, litigation claims, and evidence lists, and identifies legal relationships between entities (plaintiff-defendant relationship, creditor-debtor relationship, etc.). Through mechanisms such as text coverage detection and entity consistency verification, it reduces AI illusions, which refer to information gaps. This can lead to logical breaks or factual errors in the content generated by the model. By analyzing the completeness and coverage of the text, we can effectively ensure that the model covers all necessary information when generating content, thereby reducing this phenomenon. In the above process, a three-layer progressive extraction (entity → structure → supplement) is mainly adopted to gradually refine the information granularity. Through a multi-round self-correction mechanism, correction prompts are dynamically generated for verification errors, with a maximum of 3 rounds of iterative correction. The confidence score in the 0-1 interval is calculated for each extracted field. Through the Few-shot learning technique, a small number of examples (usually 3-5 samples) are injected into the prompts to improve the accuracy of the AI ​​model. Its core mechanism lies in using the context learning ability of a large language model to capture task rules.

[0042] The intelligent template matching module, based on a hierarchical template system, realizes case type recognition, intelligent template matching, and automatic content filling, that is: it can automatically determine that a pleading belongs to multiple standard types such as private lending, divorce disputes, and traffic accidents. The recognition process mainly relies on a two-stage classification algorithm. The first stage is the rapid screening of TF-IDF keywords (retaining the top-3 candidates), and the second stage is the deep semantic classification of the LLM (with an accuracy of 95%+); deeply matches the extracted structured data with template placeholders, supports the filling of various element types such as text, date, amount, checkbox, and table. This process mainly relies on an incremental filling mechanism to intelligently complete missing fields through context reasoning; it can also perform format adjustment through a formatting engine, such as date conversion: converting "2024-01-15" to "January 15, 2024", amount conversion: converting "1234.56" to "One thousand two hundred and thirty-four yuan and fifty-six cents", and array expansion: automatically generating a numbered list for multiple litigation requests; in addition, it also has functions for template library management and placeholder parsing, maintains a standardized Word template library, supports template version control and dynamic updates, parses placeholder tags in the template, and supports nested paths and formatting rules.

[0043] The checkbox processing module adopts a dual-engine mechanism combining a rule engine and an AI model to perform intelligent recognition and status judgment on checkboxes, that is: automatically identify the position of checkbox elements in the document, and determine whether the checkbox is checked. The role of the rule engine is to perform rapid recognition and judgment based on image features (pixel density, edge detection). The AI model mainly uses a deep learning model to accurately identify fuzzy situations. When the confidence level of the rule engine is greater than 0.8, it is directly adopted; otherwise, the AI model is called; it can also combine document semantic understanding, analyze the text content associated with the checkbox, and infer the state that the checkbox should be selected to achieve auxiliary judgment; perform intelligent error tolerance for situations such as handwritten checks and blurred images, and collect manual correction feedback to continuously optimize the recognition model.

[0044] S4. Perform quality inspection, logical verification, and error correction on the semi-finished document data through the processing layer, screen out low-quality semi-finished document data, and include it in the manual review queue. The remaining semi-finished document data with compliant quality enters step S5 after quality optimization;

[0045] Specifically: The processing layer includes a computing power distribution and scheduling module, a parallel processing module, and a quality control module. The computing power distribution and scheduling module is used to manage various AI resources, select intelligent routes, balance load, and optimize costs. That is, it uniformly manages resource information such as multiple API keys, quotas, and rates; selects the optimal API based on task type, priority, and resource availability; dynamically allocates requests to different resources to avoid single-point overload; prioritizes low-cost solutions while meeting performance requirements; and automatically switches to backup APIs when the main API fails to ensure service availability. It also monitors API call volume and remaining quotas in real time and provides early warnings.

[0046] The multi-dimensional scoring algorithm adopted by the computing power distribution and scheduling module is as follows: Score = w1 × Performance + w2 × Cost + w3 × Availability, where: Score is the final score, Performance represents performance, Cost represents cost, and Availability represents availability. w1, w2, and w3 are their respective weights, which can be dynamically adjusted according to specific scenarios. For example, when the budget is tight, w2 is set higher (e.g., 0.6), while w1 and w3 are set lower (e.g., 0.3), prioritizing APIs with lower fees. When high precision is required, w1 is set higher (e.g., 0.7), while w2 and w3 are set lower (e.g., 0.3), prioritizing APIs with stable performance and fast response. By unifying these three different dimensions into a single score, this module can rank all available API resources based on a single, comparable value, thereby selecting the optimal API for the current scenario. The module can also dynamically adjust the request rate based on historical success rates. When the success rate of a certain API starts to decline, the system will automatically and gradually reduce the rate of sending requests to that API and use cached results for duplicate requests to save costs.

[0047] The parallel processing module achieves parallel execution and performance optimization of multiple tasks through multi-threading and asynchronous processing technologies. It possesses task decomposition capabilities, breaking down complex tasks into independent sub-task units. These sub-tasks, such as placeholder parsing, information extraction, and template matching, are executed concurrently using a thread pool. Specifically, it can process areas like party information, litigation requests, and factual grounds in parallel. During batch uploads, multiple documents are processed simultaneously, and the results of each parallel task are then collected and merged. Compared to serial processing, this achieves an average speedup of 60%-80%. During this parallel processing, timeout limits are set for each sub-task to prevent prolonged blocking. Time-consuming API requests are handled asynchronously and non-blockingly (while waiting for a time-consuming operation to complete, the currently executing thread is not blocked, allowing it to execute other tasks, thus maximizing thread resource utilization). The thread pool is reusable, avoiding the overhead of frequently creating and destroying threads.

[0048] The quality control module is used to perform quality checks, logical verification, and error correction on the processing results. For example, it verifies whether all required fields are filled, checks whether the formats of ID numbers, mobile phone numbers, and amounts are compliant, checks the consistency of information such as amounts, dates, and parties involved in the document, ensures that the document meets the basic requirements of legal documents, automatically corrects common errors (such as date formats and amount thousandths), calculates the document quality score (0-1) based on a multi-level quality scoring model, adds annotations to low-confidence fields and detected problems, and automatically sends low-quality documents to the manual review queue.

[0049] The multi-level quality scoring model is as follows: Q total =0.4×Q completeness +0.3×Q accuracy +0.2×Q consistency +0.1×Q format Q total Q represents the overall score of the document. completeness Representing completeness, accounting for 40%, it measures the missing rate or coverage of data fields. (Q) accuracy Representing accuracy, accounting for 30%, this involves comparing key fields against authoritative data sources to verify error rates and assess whether the information in the document is correct, accurate, and reliable. consistency Representing consistency, this assesses whether the document content remains consistent throughout, and whether there are any contradictions or conflicts with established standards. (Q) format It represents the standardization of formatting, assessing whether the document's layout, style, structure, and visual presentation meet the standards and requirements.

[0050] Decision-making rule: Q total When Q is ≥0.90, the document quality is acceptable; when Q is ≤0.75... total <0.90, document quality is moderate, manual review is recommended; Q total When the value is less than 0.75, the document quality is poor and manual processing is necessary.

[0051] S5. For each semi-finished document data, a target document conforming to the standard format is generated through the output layer to obtain multiple real target documents, and knowledge graphs and metadata annotations are selectively constructed based on the real target documents.

[0052] Specifically, the output layer includes a document generation module, a semantic understanding and knowledge graph construction module, and a classification and metadata annotation module. The document generation module generates real target documents that conform to standard formats based on the processing results. It mainly includes: precise manipulation of Word document elements based on the python-docx library, preserving the template's font, font size, line spacing, page margins and other format settings, supporting complex elements such as tables, images, headers and footers, watermarks, etc., replacing content while keeping the template format unchanged, dynamically showing / hiding document blocks according to data content, supporting export to multiple formats such as DOCX, PDF, and HTML, supporting one-click generation of multiple documents, automatically saving historical versions, and supporting rollback and comparison.

[0053] The semantic understanding and knowledge graph construction module extracts key semantic units from document content using natural language processing technology to construct a dynamic knowledge graph. It can identify key concepts, entities, and relationships in policy documents, achieve cross-document correlation analysis, establish semantic associations and reference relationships between different documents, generate a policy evolution map, present the relationship network between policies, construct a time series knowledge graph based on publication time, effectiveness level, etc., and can also realize knowledge graph updates, automatically updating the associated nodes and relationships when new official documents are published.

[0054] The recommendations for knowledge graphs rely on graph neural network modeling, using GNNs to represent the complex network structure of entities and relationships;

[0055] The influence or relevance of a new official document on an existing knowledge graph is quantitatively assessed using a relevance calculation formula: U = (N...) new / N total )×α+(R new / R total )×(1-α), where: N new N represents the number of newly added nodes (i.e., the number of new concepts or entities extracted from new documents that do not exist in the existing knowledge graph). total R represents the total number of nodes (i.e., the total number of all nodes in the existing knowledge graph). new R represents the number of newly generated relations (i.e., the number of new relations extracted from new documents that do not exist in the existing knowledge graph). totalThe total number of relations is represented by α (i.e., the total number of relations in the entire existing knowledge graph). α represents the node weight coefficient (0≤α≤1), which can be adjusted. If α is close to 1, it indicates that the formula focuses more on the contribution of newly added nodes, and introducing new concepts (such as new policy areas or new entity types) is more important than establishing new connections. Conversely, if α is close to 0, the formula focuses more on the contribution of newly added relations, indicating that establishing new connections between existing concepts (such as new policy references or new logical relationships) better reflects the influence of documents. If α=0.5, it indicates that the addition of nodes and relations is equally important.

[0056] The Analytic Hierarchy Process (AHP) is used to construct an evaluation model and assign weights to its various indicators. Specifically, the evaluation model consists of three key indicators (release time, level of effectiveness, and issuing department) and their weights. The purpose is to calculate a comprehensive score for each policy to quantify its importance or influence. The evaluation model formula is P = α × T + β × L + γ × D, where T represents the release time score (i.e., the newer the release time, the higher the score), α represents the release time weight, and α = 0.3, L represents the level of effectiveness score (e.g., administrative regulations > departmental rules > normative documents), β represents the level of effectiveness weight, and β = 0.5, D represents the issuing department's authority score, and γ represents the issuing department weight, and γ = 0.2.

[0057] The classification and metadata annotation module is designed to create a meta-learning-driven few-shot classifier. Using the MAML (Model-Agnostic Meta-Learning) algorithm, an efficient classification model can be quickly trained with a small amount of labeled data, i.e., an effective classification model can be trained with only 5-10 labeled samples. The MAML algorithm enables the ability to quickly adapt to new categories, integrates document visual layout, text semantics, and format features, automatically generates metadata tags for unlabeled documents, intelligently selects the most valuable samples to request manual annotation, and continuously optimizes the classification model as the labeled data increases.

[0058] The MAML meta-learning algorithm consists of two phases: the meta-training phase and the rapid adaptation to new tasks phase.

[0059] I. Meta-training Phase: Retrieve tasks (task T) one by one from the task distribution system and execute subsequent operations. First step (intra-task update): θ' = θ - α L_T(θ); Second step (meta-update): Golden starting point θ = initial parameter θ - β L_T(θ');

[0060] The first step aims to simulate the rapid adaptation process of the model on the new task T. First, the current "general" model (with initial parameters θ) is applied to task T. Then, the loss L_T(θ) is calculated using a small number of support set samples from task T, and the gradient of this loss with respect to the initial parameters θ is calculated. L_T(θ), this gradient indicates the direction in which the model adjusts its parameters on task T. The model updates along this direction with a step size of α, resulting in a set of parameters θ' temporarily optimized for task T. Note that θ' is a temporary variable that is only valid in the current task T and cannot be directly used as the final model. It is only used to evaluate "the rate at which the model adapts to a new task if it starts learning from θ".

[0061] The second step aims to update the "general" initial parameters θ themselves, allowing them to adapt to all tasks more quickly in the future. First, the θ' obtained in the previous step is applied to the query set samples of task T, and the loss L_T(θ') is calculated. This loss measures the final performance of the temporary model θ' on task T. The gradient of this loss with respect to the original parameters θ is then calculated. L_T(θ') is then used to update the initial parameters θ along the direction of this gradient (with a step size of β). By repeatedly executing the cycle of "simulation adaptation → performance evaluation → initial parameter optimization" on different tasks, MAML eventually learns a "golden starting point" parameter θ. Starting from this point, the model can adapt to new tasks extremely quickly.

[0062] II. Quickly Adapting to New Tasks:

[0063] Input: The "golden starting point" parameter θ learned in the meta-training phase; a completely new task that may never have been encountered during training; and this task has only 5-10 labeled samples. The task process completely reuses the "in-task update" step from the meta-training phase, with the formula θ' = θ - α. L_newtask(θ) is used to calculate the gradient using a small number of samples from the new task, and then updated to obtain the final model θ' suitable for this new task.

[0064] This application, through the innovative integration of multiple technologies, sequentially processes documents through input, recognition, analysis, matching, scheduling, parallel processing, quality control, generation, and knowledge graph generation. It constructs a high-precision, low-cost, and professional intelligent document processing system, addressing the problems of low accuracy, high cost, and insufficient professionalism in existing products. By improving the efficiency of judicial document processing by over 80%, it significantly improves the quality of judicial services, reduces litigation costs, and makes it easier for ordinary people to access professional legal services, thus possessing significant social and economic value.

[0065] Second implementation method:

[0066] Based on the first embodiment, this implementation adds a privacy protection layer and steps S6, S7, and S8 following step S5 to improve the storage security of the obtained real target document, as detailed below:

[0067] S6. For each real target document, its content is split into a private content and a non-private content through a privacy protection layer, thereby obtaining multiple private content and multiple non-private content. For the same real target document (let's call it document P) divided into private content A and non-private content B, another non-private content is randomly selected and combined with private content A to form the first semi-pseudo-document. Then, another private content is randomly selected and combined with non-private content B to form the second semi-pseudo-document.

[0068] The privacy protection layer includes a feature segmentation module and a real-fake document fusion module. The feature segmentation module is used to segment the real target document, and the real-fake document fusion module is used to generate a semi-fake document.

[0069] Privacy content A may be the names, contact information, addresses, etc. of the people in the real target document, while non-privacy content B may be the details of the dispute, litigation claims, etc. in the real target document;

[0070] Through the operation in step S6, the first semi-pseudo-document is composed of the private content A of document P and the non-private content of another real target document P1, and the second semi-pseudo-document is composed of the non-private content B of document P and the private content of another real target document P2. P, P1, and P2 are three different real target documents. For the two semi-pseudo-documents corresponding to document P, unauthorized personnel cannot obtain the complete real target document P by obtaining either semi-pseudo-document alone. Even if they obtain both semi-pseudo-documents, since they do not know which part of the semi-pseudo-document corresponds to the content of document P, unauthorized personnel cannot accurately merge the two semi-pseudo-documents to obtain the real target document P. Therefore, dual privacy protection for the real target document P is achieved.

[0071] S7. Each real target document and its two corresponding semi-pseudo-documents are combined to form a tiered document library (i.e., one real target document corresponds to one tiered document library), which is used as the final processing result of the corresponding initial document. Furthermore, one of the semi-pseudo-documents is randomly selected as the explicit document and is not encrypted. The other semi-pseudo-document is selected as the first-level hidden document, which is encrypted at level one and a level one key is generated. The real document is selected as the second-level hidden document, which is encrypted at level two and a level two key is generated.

[0072] The privacy protection layer also includes a document package generation module and a multi-level encryption module. The document package generation module is used to form a hierarchical document library from real target documents and semi-pseudo-documents, and the multi-level encryption module is used to selectively encrypt the data in the hierarchical document library.

[0073] The tiered document library is viewed sequentially: explicit documents, first-level hidden documents, and second-level hidden documents. Explicit documents can be viewed directly, while the first and second-level hidden documents require corresponding level 1 and level 2 keys to be viewed sequentially. This is because only authorized personnel possess the level 1 and level 2 keys and can normally view all three levels of documents in the tiered document library. When an unauthorized person views the tiered document library, they directly view the explicit document (i.e., one of the semi-pseudo-documents) and are unaware of the first and second-level hidden documents behind it. The explicit document only contains either private content A or non-private content B. The unauthorized person cannot obtain the complete complaint content from the explicit document and also does not know which part of the explicit document contains the private or non-private content that they need to steal (i.e., the content belonging to the real target document P). This achieves the first line of protection against the complaint.

[0074] Step S8: When an unauthorized person attempts to obtain the first-level hidden document in the hierarchical document library by illegal means, the second-level hidden document is automatically destroyed and an alarm is triggered; when an authorized person enters the first-level key and the second-level key, the private content A and the non-private content B can be extracted from the two semi-pseudo-documents, and a real document can be synthesized again. Then, step S7 is performed again to generate a new and secure hierarchical document library.

[0075] The privacy protection layer also includes a security monitoring and processing module and a secondary generation module. The security monitoring and processing module is used to monitor whether the tiered document library is viewed illegally and to destroy the second-layer hidden document when necessary. The secondary generation module is used to extract the privacy content A and non-privacy content B from the two semi-pseudo-documents and regenerate the second-layer hidden document.

[0076] When an unauthorized person becomes aware of the existence of the first-layer hidden document and attempts to illegally obtain it, the privacy protection layer automatically destroys the second-layer hidden document (i.e., the real target document). This makes it difficult for the unauthorized person to directly obtain the second-layer hidden document even if they later discover its existence. Furthermore, when the unauthorized person successfully breaks the first-level encryption mechanism and obtains the first-layer hidden document, they cannot accurately merge the two semi-pseudo-documents to obtain the real target document P because they do not know which part of the explicit document and the first-layer hidden document (i.e., the two semi-pseudo-documents) corresponds to the real target document P. This achieves a second layer of protection for the complaint.

[0077] In addition, after authentication, authorized users can accurately extract the private content A and non-private content B from the two semi-pseudo-documents by combining the primary key and the secondary key, and regenerate the real target document P. This effectively reduces the risk of losing the final processing result, and reassembles the real target document P and the corresponding two semi-pseudo-documents into a tiered document library to continue to maintain dual privacy protection for the document content.

[0078] The above description is merely a preferred embodiment of the present invention; it encompasses all the protection scope of the present invention. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solutions and improved concepts of the present invention, should be covered within the protection scope of the present invention.

Claims

1. A multi-technology integrated intelligent document processing method, characterized in that: Includes the following steps: S1. Receive multiple initial documents to be processed through the input layer and preprocess them; S2. Perform OCR intelligent recognition of the content of each document through the recognition layer: Select the optimal engine combination through the intelligent scheduling algorithm to perform high-precision text recognition and result fusion. The intelligent scheduling algorithm realizes the dynamic selection of the optimal engine based on image features, historical performance and cost budget. When recognition fails, the parameters are automatically adjusted and recognition is repeated. S3. The analysis layer performs deep semantic analysis and information extraction on the identified content, and intelligently fills in the missing parts to obtain multiple semi-finished document data. S4. The processing layer performs quality checks, logical verification, and error correction on the semi-finished document data. A document quality score is calculated based on a multi-level quality scoring model, as follows: Q total =0.4×Q completeness +0.3×Q accuracy +0.2×Q consistency +0.1×Q format , where Q total Q represents the overall score of the document. completeness Representing completeness, Q accuracy Representing accuracy, Q consistency Representing consistency, Q format The format and standardization of the documents are used to filter out low-quality semi-finished documents and include them in the manual review queue. The remaining semi-finished documents that meet the quality standards are then optimized and proceed to step S5. S5. For each semi-finished document data, a target document conforming to the standard format is generated through the output layer to obtain multiple real target documents, and knowledge graphs and metadata annotations are selectively constructed based on the real target documents. S6. For each real target document, its content is split into a private content and a non-private content through a privacy protection layer, thereby obtaining multiple private content and multiple non-private content. For the same real target document divided into private content A and non-private content B, another non-private content is randomly selected and combined with private content A to form the first semi-pseudo-document. Then, another private content is randomly selected and combined with non-private content B to form the second semi-pseudo-document. S7. Combine each real target document and its two corresponding semi-pseudo-documents into a tiered document library, which serves as the final processing result of the corresponding initial document. Randomly select one of the semi-pseudo-documents as the explicit document and do not perform encryption processing. Select the other semi-pseudo-document as the first-level hidden document, perform first-level encryption processing and generate a first-level key. Select the real document as the second-level hidden document, perform second-level encryption processing and generate a second-level key.

2. The intelligent document processing method integrating multiple technologies according to claim 1, characterized in that: The tiered document library is viewed in the following order: visible documents, first-level hidden documents, and second-level hidden documents are viewed sequentially.

3. The intelligent document processing method integrating multiple technologies according to claim 1, characterized in that: The process also includes the following steps, S8: When an unauthorized person attempts to obtain the first-level hidden document in the hierarchical document library by illegal means, the second-level hidden document is automatically destroyed and an alarm is triggered; when an authorized person enters the first-level key and the second-level key, the private content A and the non-private content B are extracted from the two semi-pseudo-documents, and a real document is synthesized again. Then, step S7 is performed again to generate a new, secure hierarchical document library.

4. A multi-technology integrated intelligent document processing system, used to execute the intelligent document processing method according to claim 3, characterized in that: It includes an input layer, a recognition layer, an analysis layer, a processing layer, an output layer, and a privacy protection layer. The input layer includes a document input module, and the recognition layer includes a multi-engine OCR fusion module, which integrates multiple OCR recognition engines.

5. The intelligent document processing system integrating multiple technologies according to claim 4, characterized in that: The analysis layer includes an AI intelligent analysis module, an intelligent template matching module, and a checkbox processing module. The AI ​​intelligent analysis module achieves intelligent scheduling through a computing power distribution engine and performs deep semantic analysis and information extraction on document content. The intelligent template matching module is used for case type identification, intelligent template matching, and automatic content filling. The checkbox processing module adopts a dual-engine mechanism combining a rule engine and an AI model to intelligently identify and judge the status of checkboxes.

6. The intelligent document processing system integrating multiple technologies according to claim 4, characterized in that: The processing layer includes a computing power distribution and scheduling module, a parallel processing module, and a quality control module. The computing power distribution and scheduling module is used to manage various AI resources, select intelligent routes, balance load, and optimize costs. The parallel processing module uses multi-threading and asynchronous processing technology to achieve parallel execution of multiple tasks and performance optimization. The quality control module is used to perform quality checks, logical verification, and error correction on the processing results.

7. The intelligent document processing system integrating multiple technologies according to claim 4, characterized in that: The output layer includes a document generation module, a semantic understanding and knowledge graph construction module, and a classification and metadata annotation module. The document generation module generates real target documents that conform to standard formats based on the processing results. The semantic understanding and knowledge graph construction module is used to extract key semantic units from the document content and construct a dynamic knowledge graph.

8. The intelligent document processing system integrating multiple technologies according to claim 4, characterized in that: The privacy protection layer includes a feature segmentation module, a real-fake document fusion module, a document package generation module, and a multi-level encryption module. The feature segmentation module is used to segment the real target document. The real-fake document fusion module is used to generate a semi-fake document. The document package generation module is used to combine the real target document and the semi-fake document into a hierarchical document library. The multi-level encryption module is used to selectively encrypt the data in the hierarchical document library.

9. The intelligent document processing system integrating multiple technologies according to claim 8, characterized in that: The privacy protection layer also includes a security monitoring and processing module and a secondary generation module. The security monitoring and processing module is used to monitor whether the tiered document library is viewed illegally and to destroy the second-layer hidden document when necessary. The secondary generation module is used to extract the privacy content A and non-privacy content B from the two semi-pseudo-documents and regenerate the second-layer hidden document.

10. The intelligent document processing system integrating multiple technologies according to claim 9, characterized in that: The privacy content A includes the names, contact information, and addresses of the individuals in the actual target document.

Images