A system and method for intelligent processing and verification of claims documents

By combining modular decomposition and reconstruction with OCR recognition and a multimodal large model, the problems of classifier failure, OOM crash, and untraceable conclusions in claims document review were solved, achieving efficient and accurate document processing and traceability of claims conclusions.

CN123118271APending Publication Date: 2026-11-17GROUPAMA AVIC PROPERTY INSURANCE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611370153.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, claims document review suffers from problems such as overlapping titles interfering with classification, causing traditional visual or text classifiers to fail, long document sorting being prone to OOM crashes, semantic breaks and confusion of multi-source data in text fragment retrieval, and untraceable review conclusions.

Method used

The system employs a claims document splitting and reconstruction module for multi-threaded concurrent reading and splitting, uses hash fingerprinting for deduplication, combines OCR and visual multimodal large model for document subject type recognition, uses a legal litigation reorganization and sorting module for multi-task feature recognition, constructs a tree outline structure for index reconstruction, and uses a vertical expert intelligent agent to generate claims reports, thereby achieving source tracing and location of conclusions.

Benefits of technology

It improved the accuracy of document classification, reduced the risk of memory overflow, avoided semantic breaks and data confusion in documents, and significantly improved the traceability and review efficiency of claims settlement conclusions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118271A_ABST
    Figure CN123118271A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent claims document processing and verification system and method, belonging to the field of artificial intelligence technology. By using a claims document splitting and reconstruction module, a claims document classification and deduplication module, a legal litigation reorganization and sorting module, a reconstruction hierarchical index module, and a retrieval and verification report generation module in cooperation, it can accurately identify the main document type, effectively avoid misclassifying medical records as invoices due to the priority of stacked document titles appearing in front of the text flow, significantly improve the accuracy of document classification in complex stacking scenarios, reduce context length and memory overhead, prevent memory overflow, and ensure the accuracy of page number sequence and sorting logic. It also reduces claims confusion and illusions caused by heterogeneous data from different injured persons, different policies, and multiple hospitalizations. Furthermore, the verification conclusion can achieve a linked highlighting and tracing mechanism based on the outline index, supporting precise positioning at the row level and one-click two-sided verification, significantly improving the traceability of the review conclusion and the efficiency of the review.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent claims document processing and claims settlement system and method. Background Technology

[0002] In existing technologies, when reviewing personal injury claims, business personnel generally use manual methods to access the image system, reviewing each image one by one, relying on experience to visually select the target documents needed for review, and then examining the specific content of each document for the purpose of claim assessment. The formation of the claim assessment conclusion depends entirely on manual reading of the original images and manual text recording, and there is no automated document classification, content structure extraction, or conclusion visualization and traceability mechanism.

[0003] Currently, although some scattered intelligent document classification or text retrieval assistance tools have appeared on the market, their technical solutions generally rely on public cloud big model interfaces for remote data interaction, lacking physical security barriers in local private computing environments. In terms of data retrieval and generation, existing technologies generally adopt a simple single-point RAG (Retrieval-augmented Generation) mechanism, but the obtained classification results, retrieved corresponding segments, generated corresponding compensation conclusions are isolated from the underlying original images, and no linkage highlighting positioning mechanism based on image physical coordinates has been established.

[0004] The current technology also has corresponding shortcomings, such as: (1) Multiple title stacking interferes with classification, causing traditional visual or text classifiers to fail: In actual claims photography, there are often complex scenarios with multiple documents stacked and physically occluded, such as multiple small invoices stacked on the discharge record. If the traditional classifier only relies on single-page text extraction, since the invoice header is at the beginning of the text stream, it will cause serious title conflict interception failure, causing the entire medical record to be misjudged as an invoice. Simply using a convolutional neural network classifier will also cause image features to fail due to background interference and layout mixing, which is very easy to generate classification noise; (2) Long document sorting is limited by the upper limit of the model context token, which is easy to generate OOM (Out of Memory). OfMemory (Memory Overflow) Crash: Judicial judgments and other multi-page documents are often in a completely out-of-order state when they enter the system. If dozens or even hundreds of pages of text are sent into the large model for global sorting at once, it is very easy to exceed the context token limit of the model and cause memory overflow. Or, due to the loss of intermediate information in the long text reasoning of the large model, the sorting logic is seriously inaccurate and the page number sequence is disordered; (3) Semantic breaks and confusion of multi-source data in text fragment retrieval: The traditional text fragmentation mechanism forcibly cuts the document into blocks according to fixed characters, resulting in the complete medical financial table data and continuous document content being fragmented. The edges are truncated. When multiple documents are mixed for retrieval, heterogeneous data of different injured persons, different policy payout limits and multiple hospitalizations are indiscriminately spread into the large model, causing serious claims confusion and illusion when the large model performs financial actuarial calculations; (4) The conclusion lacks physical binding with the original image, and the review conclusion is not traceable: The conclusion formed by the existing AI claims system is only presented as isolated structured text, which has no connection with the original image drawing. The claims conclusion is in a black box state. The review personnel or quality inspection link need to manually search and check in hundreds of images again, which cannot achieve precise positioning at the row level and one-click double-sided verification.

[0005] Therefore, how to provide an effective technical solution to address the problems in existing technologies, such as multiple title stacking interfering with classification, causing traditional visual or text classifiers to fail, easily resulting in OOM crashes, semantic breaks and confusion of multi-source data in text fragment retrieval, and the inability to trace review conclusions, has become an urgent technical challenge to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent claims document processing and claims settlement system and method to solve the above-mentioned problems existing in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an intelligent claims document processing and claims settlement system, including a claims document splitting and reconstruction module, a claims document classification and deduplication module, a legal litigation reorganization and sorting module, a reconstruction hierarchical index module, and a claims settlement report generation module; The claims document splitting and reconstruction module is used to obtain claims case packages, perform multi-threaded concurrent reading and splitting of the claims case packages to obtain multiple single-page vector documents, perform affine transformation on the multiple single-page vector documents, extract fingerprints from the single-page image binary data obtained after the affine transformation to obtain a hash fingerprint set, and reconstruct the page order of the multiple single-page vector documents to obtain claims parsing data. The claims document classification and deduplication module is used to extract the binary byte stream of claims parsing data, calculate the hash value of the binary byte stream, compare the hash value with the hash fingerprint set, and input claims parsing data with different hash values ​​and hash fingerprint sets into the cascade classification channel for pre-verification of claims parsing data. If the pre-verification passes, the mapping result is output, the OCR text in the mapping result is extracted, and the OCR text is judged by the classification screening rules. If it meets the classification screening rules, the classification result and main title are obtained. The mapping result is physically analyzed using a visual multimodal large model to obtain the document body type, and the main title text is identified and extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, the corresponding document type is obtained. The claims document classification and deduplication module is also used to calculate the text similarity ratio based on the document type using a matching algorithm, and to deduplicate the claims parsing data of different document types based on the similarity ratio to obtain claims tags; The legal litigation reorganization and sorting module is used to perform multi-task feature recognition on the claim parsing data corresponding to the claim label of legal litigation using a large model, obtain the recognition result, perform logical verification on the recognition result based on three underlying physical hard rules, obtain the verification result, and perform adaptive path diversion on the verification result to obtain the document data. The reconstructed hierarchical index module is used to extract page data from the claims analysis data, perform structured aggregation of page data and claims tags, construct a tree outline structure using a depth-first search algorithm based on the aggregation results, and stitch cross-page tables together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, an outline index is obtained. The claim review report generation module is used to process claim cases based on document type, claim tags, document data, and outline index, using corresponding vertical expert agents to obtain one or more claim review opinions. When multiple claim review opinions are obtained, they are marked and deduplicated to obtain the final claim review opinion.

[0008] In one possible design, a parsing mapping traceability module is also included. This module is used to obtain the final claim assessment opinion, index and deduplicate the final claim assessment opinion, and reconstruct it to obtain a normalized reference data object. Based on the normalized reference data object, position fusion logic is used to extract the physical width of the original image and the physical width of the current target rendering view from the position metadata corresponding to the reference data object. Based on the physical width of the original image and the physical width of the current target rendering view, absolute pixel coordinates are obtained. The vertical scroll offset of the page is calculated based on the absolute pixel coordinates. The target highlighted area is smoothly centered based on the vertical scroll offset of the page to achieve data traceability.

[0009] In one possible design, the claims case package is read and split concurrently using multiple threads to obtain multiple single-page vector documents. An affine transformation is then performed on these single-page vector documents, and fingerprint extraction is performed on the binary data of the resulting single-page images to obtain a hash fingerprint set. Finally, the page order of the multiple single-page vector documents is reconstructed to obtain the claims parsing data, including: For a single multi-page PDF file in a claims case package, extract the total number of pages in the single multi-page PDF file, and parse the single multi-page PDF file to obtain a single-page vector document and the original logical code index; A page-level concurrent thread pool is constructed based on the total number of pages, and the page-level concurrent thread pool includes multiple rendering sub-threads; In each rendering sub-thread, an affine transformation matrix is ​​used to scale and render a single-page vector document to obtain multiple JPEG images. Read the single-page binary data of the JPEG image and use the MD5 digest algorithm to calculate the hash fingerprint of the single-page binary data to obtain the hash fingerprint of each rendering sub-thread. Based on multiple hash fingerprints, obtain a hash fingerprint set. The page order of multiple JPEG images is rearranged based on the preset original logical page number index to obtain the claims parsing data.

[0010] In one possible design, the hash value is compared with a hash fingerprint set. Claims parsing data whose hash values ​​differ from the hash fingerprint set are input into a cascading classification channel for pre-verification. If the pre-verification passes, the mapping result is output, including: The hash value is compared with the hash fingerprint set. If the hash value is the same as the hash fingerprint in the hash fingerprint set, the claim parsing data is marked as a duplicate file and subsequent operations are terminated. If the hash value is different from the hash fingerprint in the hash fingerprint set, the claims parsing data is input into the cascaded classification channel to use OCR to recognize the claims parsing data, obtain OCR text, and perform physical feature comparison verification on the OCR text. If the physical feature comparison verification passes, the first mapping result is obtained. If the physical feature comparison verification fails, but the preset invoice feature words exist in the OCR text, then based on the exclusive filtering rules and compliance review, the invoice-specific large model is used to filter and review the OCR text to obtain the second mapping result. Output either the first mapping result or the second mapping result as the mapping result.

[0011] In one possible design, the OCR text in the mapping result is extracted, and a preliminary classification screening rule is applied to the OCR text. If it meets the preliminary classification screening rule, the classification result and main title are obtained. A visual multimodal large model is used to perform physical analysis on the mapping result to obtain the document body type. The mapping result is then used to recognize the title text and obtain a recognition Boolean value. If the recognition Boolean value is 1, the corresponding main title text is extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to a preset similarity threshold, the corresponding document type is obtained, including: The mapping results are used to extract text, resulting in OCR text. The OCR text is initially classified and screened to determine whether it belongs to a pure image page with no / effective text. If it does not belong to the category, the similarity between the title text in the OCR text and the preset core category keyword corpus is calculated. If the similarity score is greater than the preset similarity threshold, the classification result and the main title in the title text are output. If the mapping results are not subject to initial classification and screening rules, a visual multimodal large model is used to perform physical analysis on the page structure of the mapping results to obtain the document body type and identify and extract the main title text in the mapping results. Calculate the similarity between the main title text and the preset corpus of keywords for each core category. If the similarity is greater than or equal to the preset similarity threshold, the corresponding document type is obtained.

[0012] In one possible design, after obtaining the corresponding document type, the following is also included: Based on document type, extract the scattered images and single pages of PDF from the claims analysis data to obtain a summary of pages to be processed; The pages to be processed are aggregated and input into a concurrent queue. In the concurrent queue, the aggregated pages to be processed are packaged according to preset packaging rules. Based on strong exclusivity and adversarial rules, a shared global large model inference thread pool is used to input the packaged aggregated pages to be processed into a large language model for classification, resulting in a standard JSON classification array. The standard JSON classification array is used as the classification result. During the classification process, based on legal litigation interception rules, claims parsing data containing preset judge-specific terms or legal procedure mediation features in the packaged aggregated pages to be processed are forcibly prioritized as claims tags for legal litigation, so as to perform multi-task feature recognition on the claims parsing data corresponding to legal litigation.

[0013] In one possible design, a large model is used to perform multi-task feature recognition on the claims parsing data tagged as legal litigation, obtaining the recognition results. Logical verification of the recognition results is then performed based on three underlying physical hard rules, yielding the verification results. Adaptive path splitting is then applied to the verification results to obtain the document data, including: For claims analysis data tagged as legal litigation, preset characters from the claims analysis data are extracted as footer features. The footer features and recognition Boolean values ​​are then structurally assembled to obtain the assembled data. The assembled data is input into a large model for multi-task feature recognition to obtain recognition results, which include the calibration status of the homepage information. Logical verification of the recognition results is performed based on three underlying physical hard rules, which are used to extract page numbers, correct the first page, and filter documents to obtain the verification results. Based on the calibration status of the homepage information in the verification results, adaptive path diversion is performed on the legal litigation single-page image to obtain the document data.

[0014] In one possible design, the claims label includes the medical record cover sheet, discharge certificate, surgical record, imaging examination report, outpatient / emergency room record, legal proceedings, identity information, disability / death certificate, accident report, policy information, medical invoices, and other invoices.

[0015] In one possible design, the vertical expert agent includes case and underwriting identification experts, medical record front page experts, outpatient / emergency room record experts, inpatient discharge record experts, imaging report experts, disability and death assessment experts, and legal litigation specialists; The cases and underwriting experts were used to analyze the accident determination reports in the claims case package and extract the accident liability ratio. Based on the accident liability ratio, type of insurance, deductible, and reimbursement ratio, the scope of insurance liability was obtained. The medical record homepage expert is used to extract information and days from the claims package to obtain the patient's basic information and actual number of days hospitalized; The outpatient / emergency room record expert is used to sort and organize the medical records in the claims package based on a timeline to obtain a sequence of medical cards; The hospital discharge record expert is used to record and extract information from the claims package to obtain the discharge diagnosis and medical advice; The image report expert is used to rearrange the CT, DR, and / or MRI reports in the claims package to obtain rearranged image reports; The disability and death assessment experts are used to extract the number, assessment institution, disability level, lost work, nursing care, and nutritional period conclusions from the judicial appraisal reports in the claims package; The legal and litigation specialist is used to classify cases in the claims package into mediation agreements, rulings, or judgments based on title characteristics.

[0016] Secondly, the present invention provides an intelligent processing and verification method for claims documents, including: The process involves acquiring a claims case package, performing multi-threaded concurrent reading and splitting of the claims case package to obtain multiple single-page vector documents, performing affine transformations on the multiple single-page vector documents, extracting fingerprints from the binary data of the single-page images obtained after the affine transformation to obtain a hash fingerprint set, reconstructing the page order of the multiple single-page vector documents, and obtaining the claims parsing data. Extract the binary byte stream of the claims parsing data and calculate its hash value. Compare the hash value with the hash fingerprint set. Input the claims parsing data with different hash values ​​and hash fingerprint sets into the cascaded classification channel for pre-verification. If the pre-verification passes, output the mapping result. Extract the OCR text from the mapping result and perform a preliminary classification screening rule on the OCR text. If it meets the preliminary classification screening rule, obtain the classification result and main title. Use a visual multimodal large model to perform physical analysis on the mapping result to obtain the document body type and identify and extract the main title text. Input the main title or main title text into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, obtain the corresponding document type. Based on document type, a matching algorithm is used to calculate the text similarity ratio. Based on the similarity ratio, duplicate claims data of different document types are deduplicated to obtain claims tags. The large model is used to perform multi-task feature recognition on the claims parsing data corresponding to the claims label of legal litigation to obtain the recognition results. The recognition results are logically verified based on three underlying physical hard rules to obtain the verification results. The verification results are then adaptively routed to obtain the document data. The claims analysis data is extracted from the page data, and the page data and claims tags are aggregated in a structured manner. Based on the aggregation results, a tree outline structure is constructed using a depth-first search algorithm, and cross-page tables are stitched together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, an outline index is obtained. Based on document type, claim tags, document data, and outline index, the corresponding vertical expert AI agent is used to process the claim case package for claim review, resulting in one or more claim review opinions. When multiple claim review opinions are obtained, they are marked and deduplicated to obtain the final claim review opinion.

[0017] Thirdly, the present invention provides a computer device comprising a memory, a processor, and a transceiver connected in sequence and communication, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the intelligent processing and claims settlement method for claims documents as described in the second aspect above.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, perform the intelligent processing and claims settlement method for claims documents as described in the second aspect above.

[0019] Fifthly, the present invention provides a computer program product containing instructions that, when the instructions are executed on a computer, cause the computer to perform the intelligent processing and claims settlement method for claims documents as described in the second aspect above.

[0020] The beneficial effects of this invention are as follows: (1) This invention uses a claims document splitting and reconstruction module to perform multi-threaded concurrent reading and splitting of claims case packages, and performs page order reconstruction on the obtained multiple single-page vector documents to obtain claims parsing data, so as to achieve physical separation and deduplication of documents. The claims document classification and deduplication module is used to input claims parsing data with different hash values ​​and hash fingerprint sets into the cascade classification channel to perform pre-verification of claims parsing data. The OCR text that passes the verification is judged by the classification initial screening rules, so as to accurately identify the document main type, effectively avoid misjudging the case as an invoice due to the priority of the stacked invoice title appearing in front of the text stream, and significantly improve the accuracy of document classification in complex stacking scenarios. (2) The claim parsing data with the claim tag "legal litigation" is identified by the legal litigation reorganization and sorting module. Then, the three underlying physical hard rules are used to perform logical verification on the identification results to obtain the verification results. The verification results are then subjected to adaptive path diversion to obtain the document data. This avoids sending dozens or even hundreds of pages of text into the large model at once to perform global sorting, thereby reducing the context length and memory overhead, preventing memory overflow, and ensuring the accuracy of page number timing and sorting logic. (3) By reconstructing the hierarchical index module, page data is extracted from the claims parsing data. The page data and claims tags are aggregated in a structured manner. Based on the aggregation results, a tree-shaped outline structure is constructed using a depth-first search algorithm. The tree-shaped outline structure is then stitched together across pages to obtain the outline index. This forms a hierarchical index without disrupting the semantic continuity of the document, avoiding table truncation and semantic breaks caused by fixed character blocks. Furthermore, based on the claims tags and the tree-shaped outline structure, multi-source data isolation is achieved, reducing claims confusion and illusions caused by the flattening of heterogeneous data from different injured persons, different policies, and multiple hospitalizations. (4) The present invention also uses a vertical expert intelligent agent to process the claims case package through the search and claim report generation module to obtain one or more claim opinions. Multiple claim opinions are marked and deduplicated to obtain the final claim opinion. This enables the claim conclusion to achieve a linkage highlighting and tracing mechanism based on the outline index, and supports precise positioning at the row level and one-click dual-sided verification, which significantly improves the traceability of the review conclusion and the efficiency of the review. Attached Figure Description

[0021] Figure 1 This is a block diagram of the intelligent claims document processing and claims settlement system provided in an embodiment of the present invention; Figure 2 The flowchart illustrates the intelligent processing and claims settlement method for claims documents provided in this embodiment of the invention. Detailed Implementation

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0023] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0024] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0025] Example: like Figure 1 As shown, the first aspect of this embodiment provides an intelligent claims document processing and claims settlement system, including a claims document splitting and reconstruction module, a claims document classification and deduplication module, a legal litigation reorganization and sorting module, a reconstruction hierarchical index module, and a claims settlement report generation module; The claims document splitting and reconstruction module is used to obtain claims case packages, perform multi-threaded concurrent reading and splitting of the claims case packages to obtain multiple single-page vector documents, perform affine transformation on the multiple single-page vector documents, extract fingerprints from the single-page image binary data obtained after the affine transformation to obtain a hash fingerprint set, and reconstruct the page order of the multiple single-page vector documents to obtain claims parsing data. The claims document classification and deduplication module is used to extract the binary byte stream of claims parsing data, calculate the hash value of the binary byte stream, compare the hash value with the hash fingerprint set, and input claims parsing data with different hash values ​​and hash fingerprint sets into the cascade classification channel for pre-verification of claims parsing data. If the pre-verification passes, the mapping result is output, the OCR text in the mapping result is extracted, and the OCR text is judged by the classification screening rules. If it meets the classification screening rules, the classification result and main title are obtained. The mapping result is physically analyzed using a visual multimodal large model to obtain the document body type, and the main title text is identified and extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, the corresponding document type is obtained. The claims document classification and deduplication module is also used to calculate the text similarity ratio based on the document type using a matching algorithm, and to deduplicate the claims parsing data of different document types based on the similarity ratio to obtain claims tags; The legal litigation reorganization and sorting module is used to perform multi-task feature recognition on the claim parsing data corresponding to the claim label of legal litigation using a large model, obtain the recognition result, perform logical verification on the recognition result based on three underlying physical hard rules, obtain the verification result, and perform adaptive path diversion on the verification result to obtain the document data. The reconstructed hierarchical index module is used to extract page data from the claims analysis data, perform structured aggregation of page data and claims tags, construct a tree outline structure using a depth-first search algorithm based on the aggregation results, and stitch cross-page tables together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, an outline index is obtained. The claim review report generation module is used to process claim cases based on document type, claim tags, document data, and outline index, using corresponding vertical expert agents to obtain one or more claim review opinions. When multiple claim review opinions are obtained, they are marked and deduplicated to obtain the final claim review opinion.

[0026] It should be noted that after the claims document classification and deduplication module obtains the corresponding document type, it also performs regular expression cleaning on the claims parsing data of different document types to remove punctuation, spaces, and non-Chinese characters, and to remove typesetting noise. Based on the historical classified text library of the same predicted category, the Ratcliff-Obershelp algorithm is used to calculate the text similarity ratio of the cleaned claims parsing data. A preset similarity threshold is used to compare the calculated similarity. If it is greater than the preset similarity threshold, it is judged as text-level duplication. The data judged as text-level duplication is deduplicated to obtain the final claims label. In this embodiment, the preset similarity threshold for fixed type documents is 0.98, and the preset similarity threshold for narrative documents is 0.95, so as to avoid false deduplication caused by the overlap of common feature words such as patient names and hospital names between different document types. The principle of the Ratcliff-Obershelp algorithm is to evaluate the similarity between two strings by calculating the length of the longest common subsequence (LCS).

[0027] In a preferred embodiment, the system further includes a parsing mapping traceability module. This module is used to obtain the final claims assessment opinion, index and deduplicate the final claims assessment opinion, and reconstruct it to obtain a normalized reference data object. Based on the normalized reference data object, position fusion logic is used to extract the physical width of the original image and the physical width of the current target rendering view from the position metadata corresponding to the reference data object. Based on the physical width of the original image and the physical width of the current target rendering view, absolute pixel coordinates are obtained. The vertical scroll offset of the page is calculated based on the absolute pixel coordinates. The target highlighted area is smoothly centered based on the vertical scroll offset of the page to achieve data traceability. After obtaining the final claims assessment opinion, the system also establishes an association between the claims assessment conclusion in the final claims assessment opinion and the corresponding evidence data. Based on the position metadata of the evidence data, a mapping relationship is established between the claims assessment conclusion and the position of the document data in the claims case package to achieve traceability between the conclusion and the evidence, support automated processing of the entire claims case process, and significantly improve the process integrity of the system.

[0028] In practice, regular expressions are used to globally search and parse the reference marker strings in the final claim assessment report. First, the source document representation and document display index corresponding to the marker string are read. When multiple consecutive reference markers are detected mapping to the same source document, index deduplication is automatically performed, reconstructing the discrete marker strings into a unified single-node reference structure and generating a normalized reference data object. Then, in response to the positioning instructions of the reference data object, the position fusion logic is invoked to process the position metadata bound to the reference. Specifically, the position fusion logic groups the discrete line-level position data in memory according to the search index, and automatically performs interval recombination and boundary fusion on the line-level rectangular coordinates with the same source document identifier and consecutive line numbers, merging them into a multi-line coordinate interval object representing a complete paragraph text area, which is then written to the rendering stack to avoid gaps during coordinate rendering. Subsequently, the physical width of the image during the original OCR parsing and the physical width of the current target rendering view are extracted from the position metadata, and a geometric space scaling transformation is performed. First, the scaling scale is calculated, expressed as follows: In the formula, To scale the bar, Render the physical width of the view for the current target. Given the physical width of the image during original OCR parsing, retrieve the physical coordinates of all original line-level bounding boxes corresponding to the reconstructed paragraph intervals. The coordinates are... Perform a matrix multiplication on the original physical coordinates and the scaling scale to obtain the absolute pixel coordinates under the target rendering view. The expression is: Since the original OCR coordinates are obtained based on the physical pixels of an image with a fixed resolution, and the physical pixel size of the display viewport of each terminal has dynamic uncertainty, this embodiment calculates the scale K and performs dot product mapping to normalize and convert the fixed physical image coordinates into accurate absolute pixel coordinates under any target rendering viewport, ensuring geometric consistency of rendering across resolution terminals. After obtaining the absolute pixel coordinates, the absolute pixel coordinates are input into the rendering context of the target canvas, and the graphics drawing interface is called to generate a vector semi-transparent highlight mask in the corresponding coordinate interval. At the same time, the system uses the mapped coordinates... and The coordinates automatically calculate the vertical scroll offset of the page and drive the view scrolling component to automatically and smoothly center-align the target highlighted area, achieving accurate data traceability from the reasoning conclusion to the location of the original document.

[0029] In a preferred embodiment, the claims case package is read and split using multi-threaded concurrent methods to obtain multiple single-page vector documents. An affine transformation is performed on these single-page vector documents, and fingerprint extraction is performed on the binary data of the single-page images obtained after the affine transformation to obtain a hash fingerprint set. Finally, page order reconstruction is performed on the multiple single-page vector documents to obtain claims parsing data, including: For a single multi-page PDF file in a claims case package, extract the total number of pages in the single multi-page PDF file, and parse the single multi-page PDF file to obtain a single-page vector document and the original logical code index; A page-level concurrent thread pool is constructed based on the total number of pages, and the page-level concurrent thread pool includes multiple rendering sub-threads; In each rendering sub-thread, an affine transformation matrix is ​​used to scale and render a single-page vector document to obtain multiple JPEG images. Read the single-page binary data of the JPEG image and use the MD5 digest algorithm to calculate the hash fingerprint of the single-page binary data to obtain the hash fingerprint of each rendering sub-thread. Based on multiple hash fingerprints, obtain a hash fingerprint set. The page order of multiple JPEG images is rearranged based on the preset original logical page number index to obtain the claims parsing data.

[0030] It should be noted that the expression for a page-level concurrent thread pool is: In the formula, For page-level concurrent thread pools, It is a minimum value function. The preset concurrency limit constant is set to 8 in this embodiment. This represents the total number of pages.

[0031] In practice, the system first acquires the user-input claim case package, then extracts the total number of pages in a single multi-page PDF file within the claim case package, and dynamically establishes a page-level concurrent thread pool based on the total number of pages. This page-level concurrent thread pool includes multiple rendering sub-threads to eliminate handle lock contention when multiple threads concurrently render the same file. Each rendering sub-thread independently loads its own document object via a file path, achieving physical handle isolation. Subsequently, each rendering sub-thread uses an affine transformation matrix to scale and render the single-page vector document. In this embodiment, the scaling factor is 1.5 times to ensure the text edge clarity required for subsequent OCR recognition while controlling memory usage. Finally, the rendered JPEG image is directly saved to local storage. Only metadata is transmitted between microservices to achieve zero-copy image transmission at the local storage layer. The system reads the single-page image binary data of the JPEG image and calculates a hash fingerprint using the MD5 digest algorithm. The hash fingerprint is a 128-bit hash value, and its expression is... In the formula, For JPEG image stream objects, To read the binary byte stream data of a JPEG image, To call the standard MD5 message digest function, To output the calculation result as a 32-bit hexadecimal string, which is a 128-bit hash value, A hash fingerprint is used to achieve image-level deduplication and metadata association. After each rendering sub-thread completes asynchronously, the single-page image data and its corresponding original logical page number index are stored in a memory buffer list. After all single-page rendering tasks of the current PDF are completed, the pages are rearranged in ascending order in memory according to the original logical page number index. This restores the physical page order of the original document with extremely low computational overhead. The original logical page number index is used to represent the positive integer sequence number of the page in the initial natural arrangement order in the original PDF file, such as 1 to N.

[0032] In a preferred embodiment, the hash value is compared with a hash fingerprint set. Claims analysis data whose hash values ​​and hash fingerprint sets differ are input into the cascaded classification channel for pre-verification of the claims analysis data. If the pre-verification passes, the mapping result is output, including: The hash value is compared with the hash fingerprint set. If the hash value is the same as the hash fingerprint in the hash fingerprint set, the claim parsing data is marked as a duplicate file and subsequent operations are terminated. If the hash value is different from the hash fingerprint in the hash fingerprint set, the claims parsing data is input into the cascaded classification channel to use OCR to recognize the claims parsing data, obtain OCR text, and perform physical feature comparison verification on the OCR text. If the physical feature comparison verification passes, the first mapping result is obtained. If the physical feature comparison verification fails, but the preset invoice feature words exist in the OCR text, then based on the exclusive filtering rules and compliance review, the invoice-specific large model is used to filter and review the OCR text to obtain the second mapping result. Output either the first mapping result or the second mapping result as the mapping result.

[0033] It should be noted that the calculated hash value is compared with the hash fingerprint set. If the calculated hash value is the same as any existing hash value in the hash fingerprint set, the claim file is directly marked as a duplicate file, and the associated original document ID is recorded, terminating subsequent processing to avoid unnecessary calculations and save computational resources. For files that do not pass the physical deduplication, the claim parsing data is input into the cascade classification channel to use OCR to recognize the claim parsing data. Physical feature comparison verification is performed on the recognized OCR text. If the verification passes, the first preset result is obtained. If the verification fails, but the recognized OCR text contains preset invoice feature words, the OCR text is filtered and reviewed using an invoice-specific large model based on exclusive filtering rules and compliance review to obtain the second mapping result. The preset invoice feature words include unified social credit code, invoice date, fee receipt, general receipt, or electronic invoice, etc.

[0034] Furthermore, in the cascading classification channel, physical feature comparison and verification are performed on the OCR text based on regular expression detection logic. Specifically, this includes: checking whether the OCR text simultaneously contains invoice identifiers such as unified social credit code, invoice date, payment receipt, general invoice, or electronic invoice; using a preset regular expression to match the invoice number in the OCR text to check whether it contains keywords such as invoice code, invoice number, or invoice number, followed by a pure numeric number of no less than 6 digits; if it contains both medical and the aforementioned payment receipt features, it is directly mapped to medical invoice fees; if it only meets the general invoice features and does not contain the word "medical," it is mapped to other invoice fees; if all three verification steps are satisfied, it is determined to be a legitimate invoice and the mapping result is output.

[0035] If the above three-step verification fails, but the OCR text contains characteristic words resembling invoices such as payment receipts, settlement documents, and official seals, then a special large-scale invoice detection model is triggered. This involves applying exclusive filtering rules and compliance review to the OCR text based on a preset invoice financial audit prompt word template. If the verification passes, the text is confirmed as a formal invoice and then mapped for output; otherwise, it returns to the general classification process. For example, the exclusive filtering rule can be to forcibly filter documents containing insurance policies and detailed lists as non-invoices. The compliance review can be to verify whether the invoice issuer is a medical institution listed in the medical register. In this embodiment, the invoice financial audit prompt word template is set to 0.5 sampling temperature.

[0036] In a preferred embodiment, the OCR text in the mapping result is extracted, and the OCR text is subjected to a preliminary classification screening rule. If it meets the preliminary classification screening rule, the classification result and main title are obtained. The mapping result is physically analyzed using a visual multimodal large model to obtain the document body type. The mapping result is then subjected to title text recognition to obtain a recognition Boolean value. If the recognition Boolean value is 1, the corresponding main title text is extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to a preset similarity threshold, the corresponding document type is obtained, including: The mapping results are used to extract text, resulting in OCR text. The OCR text is initially classified and screened to determine whether it belongs to a pure image page with no / effective text. If it does not belong to the category, the similarity between the title text in the OCR text and the preset core category keyword corpus is calculated. If the similarity score is greater than the preset similarity threshold, the classification result and the main title in the title text are output. If the mapping results are not subject to initial classification and screening rules, a visual multimodal large model is used to perform physical analysis on the page structure of the mapping results to obtain the document body type and identify and extract the main title text in the mapping results. Calculate the similarity between the main title text and the preset corpus of keywords for each core category. If the similarity is greater than or equal to the preset similarity threshold, the corresponding document type is obtained.

[0037] In practice, the OCR text is initially classified using certain rules. If the OCR text is a pure image page with no / valid text, it is categorized as "Other Invoice Fees." If not, the similarity between the title text in the OCR text and the pre-set keyword corpus for each core category is calculated. If the similarity score is greater than or equal to 0.85, the corresponding classification result is obtained, and the main title in the OCR text is extracted. The pre-set keyword corpus for each core category includes medical record homepages, discharge certificates, and surgical records, etc. If the OCR text is not included in the classification, the classification is determined by the following rules: If the initial classification screening rules are used, a visual multimodal large model is used to perform physical analysis of the page structure to determine the main type of the document and identify and extract the main title text at the top of the page. The main type of the document includes certificates, text documents, or non-text documents. After the valid main title text is extracted, it is input into the TF-IDF pre-detection vector space. The main title text is then matched with the keyword corpus of each core category. If the matching score is greater than or equal to 0.85, the document type to which it belongs is output, thus achieving a second reduction in computational cost.

[0038] In a preferred embodiment, after obtaining the corresponding document type, the method further includes: Based on document type, extract the scattered images and single pages of PDF from the claims analysis data to obtain a summary of pages to be processed; The pages to be processed are aggregated and input into a concurrent queue. In the concurrent queue, the aggregated pages to be processed are packaged according to preset packaging rules. Based on strong exclusivity and adversarial rules, a shared global large model inference thread pool is used to input the packaged aggregated pages to be processed into a large language model for classification, resulting in a standard JSON classification array. The standard JSON classification array is used as the classification result. During the classification process, based on legal litigation interception rules, claims parsing data containing preset judge-specific terms or legal procedure mediation features in the packaged aggregated pages to be processed are forcibly prioritized as claims tags for legal litigation, so as to perform multi-task feature recognition on the claims parsing data corresponding to legal litigation.

[0039] It should be noted that scattered images and single-page PDFs must be classified using a large language model for deep semantic classification. This embodiment employs a batch concurrency optimization mechanism. First, the pages to be processed are aggregated into a concurrent queue, then grouped into blocks of five pages each. A shared global large model inference thread pool is used to concurrently submit these blocks to the large language model. The prompts force the large language model to receive multiple document texts simultaneously and output a single, normalized standard JSON classification array, significantly reducing system latency. An example of a standard JSON classification array is ["Classification A", "Classification B"]. During the large language model classification, strong exclusion and adversarial rules are applied to scattered images and single-page PDFs, and rules are applied to those containing judge-specific rules. Documents with legal procedure mediation characteristics are subject to legal litigation interception rules. They are forcibly prioritized as claims under the legal litigation category and intercepted and inserted into an independent legal sorting pool for secondary AI page reordering. Examples of strong exclusion and adversarial rules include making decisions based on the details of the main text when multiple titles overlap and conflict. Untitled continuation pages are marked as system continuation pages. Judge-specific terms include, but are not limited to, "This court holds," "Adopts," "Disadvantages," "First instance court found the facts," "Appeal dismissed," and "Upholds the original judgment." Legal procedure mediation characteristics include clauses signed by the parties and confirmation of liability, such as officially issued traffic accident mediation agreements, people's mediation agreements, and traffic accident settlement agreements.

[0040] Furthermore, in this embodiment, for PDF files, the classification status of the previous page is processed and maintained according to the page number order of the PDF file. When the main title text is not recognized by the visual multimodal large model, i.e., its Boolean value is false, and the effective text of the cleaned page is greater than 1000 characters, the system determines that the current page is a continuation page without a main title. At this time, the classification status of the previous page is read, and a set inclusion check is performed in the pre-configured set of allowed continuation page categories. The pre-configured set of allowed continuation page categories includes, but is not limited to, seven types of long documents such as medical record homepage, medical discharge certificate, medical surgical record, accident determination letter, insurance policy information, legal proceedings, and disability / death certificate. If the classification status of the previous page exists in the set of allowed continuation page categories, the current continuation page without main title text directly and automatically inherits the category label of the previous page without repeatedly calling the large language model. If the classification status of the previous page is not in the set of allowed continuation page categories, for example, if the previous page is an invoice or outpatient / emergency record, the inheritance is forcibly interrupted, and the current page classification is downgraded and reset to other invoice fees.

[0041] In a preferred embodiment, a large model is used to perform multi-task feature recognition on the claims parsing data corresponding to the claims category labeled as legal litigation, obtaining the recognition result. The recognition result is then logically verified based on three underlying physical hard rules, yielding the verification result. Adaptive path splitting is applied to the verification result to obtain the document data, including: For claims analysis data tagged as legal litigation, preset characters from the claims analysis data are extracted as footer features. The footer features and recognition Boolean values ​​are then structurally assembled to obtain the assembled data. The assembled data is input into a large model for multi-task feature recognition to obtain recognition results, which include the calibration status of the homepage information. Logical verification of the recognition results is performed based on three underlying physical hard rules, which are used to extract page numbers, correct the first page, and filter documents to obtain the verification results. Based on the calibration status of the homepage information in the verification results, adaptive path diversion is performed on the legal litigation single-page image to obtain the document data.

[0042] It should be noted that the legal litigation reorganization and sorting module only reorders single-page or multi-page images corresponding to the claim label "legal litigation". First, it extracts each single-page image from the claim parsing data and extracts the last 100 characters of the text as a footer feature. Then, it structurally assembles the metadata of each module to obtain the assembled data. The metadata of each module includes a unique page identifier, a preview of the first 1000 characters of the OCR text, the footer feature, and a Boolean value indicating whether the page contains the main title text. These are sent in batches to the large model for multi-task feature recognition to obtain the recognition results. Then, three underlying physical hard rules are used to logically verify the recognition results to obtain the verification results. Based on the calibration status of the homepage information in the verification results, the legal litigation single-page images are adaptively routed to obtain the document data.

[0043] In practice, adaptive path routing includes path A and path B. Path A is mainly for page number order lookup (sort by page numbers). When the identified homepage contains valid initial physical page numbers, path A is triggered. Path B is mainly for global semantic outline reordering (sort document pages). When the identified homepage does not contain any valid page numbers, or when a large area of ​​physical page numbers is missing in the candidate single page and cannot be directly aligned, path B is triggered.

[0044] In this embodiment, the execution logic of path A is as follows: starting from the starting page number of the designated homepage, the search for the target page number is performed iteratively and incrementally in the pool of unclaimed candidate pages. The current page number is incremented by 1 in each iteration. If no page matching the target page number is found in the current iteration, the search terminates and the loop exits. If a unique page matching the target page number is found, that page is directly accepted. The plain Chinese text of that page is extracted for hard character matching. If the text contains signature words such as "judge," "clerk," "presiding judge," or "signature," then that page is determined to be the final page of the current document. If two or more matching pages are found... When physical page number conflicts occur between pages with the same page number, a large text model is invoked for disambiguation. The large text model reads the last 500 characters of the previous verified page and performs contextual sentence connection and semantic coherence analysis with the first 500 characters of all conflict candidate pages. By evaluating the grammatical and logical rationality of the concatenated pages, the page with the highest semantic coherence is selected as the target page. If disambiguation fails, the first candidate page is selected as a fallback, and the signature characters are compared. After each successful alignment, the matching page is added to the verified list and removed from the unclaimed candidate pool until the end page marker is detected, at which point the loop exits.

[0045] Specifically, the execution logic of path B is as follows: call a descriptor construction function to process unordered candidate pages, wherein for pages with OCR text of 1000 characters or less, the full text is retained, and for pages with OCR text of more than 1000 characters, only the first and last 500 characters and footer features are retained, then calculate the total character count of all descriptors and select a shunting channel, the shunting channels include a global one-time sorting channel and an adaptive sliding window progressive sorting channel; in the global one-time sorting channel, if the total character count is less than 50,000 characters, perform global logical sorting on all descriptors within a single large model call, the large model analyzes the text outline and character continuation relationship, and directly outputs the sorted page identifier array; in the adaptive sliding window progressive sorting channel, if the total character count is greater than or equal to 50,000 characters, split the descriptor sequence into several sub-blocks according to the character size, extract the last 500 characters of the last page of the verified stack as a sliding alignment reference pointer, in the large model call, require the model to find the unique next page that seamlessly connects with the reference pointer in semantic structure in the current sub-block, after successful positioning, perform a physical dequeue operation, pop this page from the to-be-sorted candidate pool, update the sliding alignment reference pointer and construct a new sub-block to enter the next round of search, until a signature feature is detected or the candidate pool is emptied.

[0046] In this embodiment, the three underlying physical hard rules include regular physical page number extraction and overwriting, first page misjudgment correction, and blacklist document filtering, wherein the regular physical page number extraction and overwriting specifically uses four preset regular expressions to physically extract page numbers from footer features, the four preset regular expressions include the horizontal line page number regex r'-\s (\d+)\s -' for matching page numbers in the form of "- 5 -" or "-5-"; the Chinese page number regex r'第\s (\d+)\s 页' for matching page numbers in the form of "第5页"; the slash ratio page number regex r'(\d+)\s / \s \d+' for matching page numbers in the form of "5 / 10"; the isolated digit regex r'(?<!\d)([1-9]\d?)(?!\d)\s The $' flag is used to match isolated page numbers at the end. If the page number extracted by the large model is empty but the regular expression successfully extracts a page number, it is filled with the regular expression page number. If the page number extracted by the large model is inconsistent with the regular expression page number, and the regular expression page number is within the valid range of 1 to 99, it is forcibly overwritten, and the regular expression page number is used as the standard. The homepage misjudgment correction is as follows: if the large model determines that a page is the homepage, but the page contains a false boolean value for whether the page contains the main title text, or the actual page number of the page after the physical page number regular expression extraction and overwriting rules is greater than 2, it is determined that the large model misjudged the homepage, and the homepage flag is forcibly overwritten and reset to false. The blacklist document filtering is as follows: if the large model determines that a page is the homepage, but the identified sub-document type belongs to the blacklist documents that do not require page order reordering in claims such as summons, appeals, complaints, answers, post-judgment notices, service receipts, applications, and powers of attorney, the homepage flag is forcibly overwritten and reset to false.

[0047] In a preferred embodiment, the claim label includes the medical record cover page, discharge certificate, surgical record, imaging examination report, outpatient / emergency room record, legal proceedings, identity information, disability / death certificate, accident report, policy information, medical invoices, and other invoices.

[0048] In a preferred embodiment, the hierarchical index reconstruction module is specifically used to extract page data, which includes OCR text, character coordinates, paragraph outline levels, and HTML / Markdown layout data for each page. This data is then combined with claim tags for structured aggregation. First, a depth-first search algorithm is used to construct a tree-like outline structure. During the traversal of page data and claim tags, cross-page table stitching checks are performed on adjacent pages of the same type. When HTML tags with identical outline paths and table column width attributes are detected on two adjacent pages... When using nodes, physically remove the end of the previous page. Closing tags and the beginning of the next page

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] Figure 2

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] Guide labels and repeating header rows The structure is then stitched and merged to restore it into a continuous, unified long table, resulting in a complete table to prevent data misalignment and table structure damage caused by page breaks. Based on the complete table, a tree outline is obtained. Furthermore, the nodes of the tree outline, the complete table, the corresponding original slices, and the claim tags are written as indivisible core metadata into the knowledge base database and search module. The knowledge base database and search module can be, for example, Elasticsearch or a vector database. The original paragraphs on the tree outline and the merged complete table nodes are written in batches to the Elasticsearch index as indivisible whole nodes. Each node has a unique outline level and contextual metadata. If the leaf node text of the tree outline is greater than or equal to 2000 characters, a sliding window slice is used, with a slice size of 500 characters and an overlap of 20 characters. This is used to calculate its physical starting offset within the original large node. The complete table is divided into sub-tables according to the Token upper limit, and the unique identifier of the parent large node is forcibly implanted into the small node slices to form pointer binding. Furthermore, initialize a local area network reused HTTP connection pool, maintain 20 long connections, allow a maximum of 50 concurrent connections, automatically execute 3 exponential backoff retries, and simultaneously start two fully parallel asynchronous threads. The first thread batch-enters the original text's large nodes and micro-slices with hierarchical outline trees and parent-child binding relationships into the Elasticsearch index; the second thread batch-calls the feature vector proxy interface, calculates the feature vectors of each micro-slice concurrently, batch-imports them into a distributed vector database, and writes them to the Redis cache after successful upload. In a preferred embodiment, the vertical expert agent includes a case and underwriting assessment expert, a medical record panel expert, an outpatient / emergency room record expert, an inpatient discharge record expert, an imaging report expert, a disability / death assessment expert, and a legal litigation specialist. The case and underwriting assessment expert is used to analyze the accident determination report in the claims package and extract the accident liability ratio. Based on the accident liability ratio, insurance type, deductible, and reimbursement ratio, the scope of insurance liability is obtained. The medical record panel expert is used to extract information and days from the claims package to obtain the patient's basic information and actual hospitalization days. The outpatient / emergency room record expert is used to... The inter-axis system sorts and arranges the medical records in the claims package to obtain a sequence of medical cards; the inpatient discharge record expert extracts records from the claims package to obtain discharge diagnoses and medical orders; the imaging report expert rearranges CT, DR, and / or MRI reports in the claims package to obtain rearranged imaging reports; the disability and death assessment expert extracts the number, assessment institution, disability level, lost work time, nursing care, and nutritional period conclusions from the judicial appraisal reports in the claims package; and the legal litigation specialist classifies the cases in the claims package into mediation agreements, rulings, or judgments based on title characteristics.It should be noted that the medical record homepage expert is also used to extract multiple admission records, reorder them in ascending order according to the actual admission time, and automatically correct the logical error in the original scanned document where all records were printed as "Number of Hospitalizations: 1" due to input errors. The hospitalization and discharge record expert is also used when the large model executes specific truncation instructions. After parsing specific medication or exercise orders, once "department head", "physician's signature", "stamp", or a friendly reminder appears in the text, the large model must immediately stop extraction and directly truncate. All tail templates and signature noise after the truncation point are silently discarded to achieve physical truncation and noise reduction of medical orders. The vertical expert intelligent agent also includes a surgical procedure expert and a financial invoice actuary. The surgical procedure expert is used to generate a structural card independently for each surgery based on the principle of one procedure per surgery, accurately analyze the name and quantity of the implanted bone in the fracture, and prevent the merging of surgeries on different days. The financial invoice actuary is used to perform double deduplication based on the combination of invoice code and number, only calculate formal tax invoices, and extract the payer, invoice date, overall payment, personal payment, and total invoice amount. In practice, the legal litigation specialist's processing logic includes rapid bypass determination, concurrent processing of multiple documents, cross-boundary sliding continuation extraction, and adaptive truncation and summarization. Rapid bypass determination specifically means that if a case contains only one document or is classified as a mediation agreement, the system directly bypasses the process, using the original text as the extraction result without large-scale model processing. Concurrent processing of multiple documents specifically means that when faced with multiple judgments and rulings to be extracted, a native multi-threaded concurrency mechanism is activated, with the number of threads set to the smaller of the total number of documents and 5. Each thread independently and in parallel calls the large-scale model. Cross-boundary sliding continuation extraction specifically means that when a single document exceeds 30,000 characters, it is divided into multiple sub-blocks. During extraction, the extraction conclusion of the previous block is used as the context content input for the current block, ensuring that clauses at the edges of cross-blocks are aligned and spliced. Adaptive truncation and summarization specifically means that if the extracted and spliced ​​text exceeds 40,000 characters, the large-scale model is directly truncated and summarized, and the Markdown extraction report is returned as a legal report. If it is within 40,000 characters, the locally deployed large-scale model is called to generate a highly structured claims assessment opinion. It is important to note that, due to the massive release of LLM summarization and refinement requests during concurrent execution of multiple vertical expert agents, which is limited by the physical GPU memory limit of the local computing cluster, a local GPU memory hard limit control lock needs to be configured. This local GPU memory hard limit control lock is used as a global concurrent thread safety control semaphore. Based on the number of local physical cards and the size of GPU memory, the maximum number of concurrent summarization calls is hard-limited to 8. Each parallel expert node must compete for this semaphore lock before initiating a summarization call. Threads that fail to compete are queued in physical memory in a non-blocking manner. This addresses the GPU memory overflow problem caused by excessive concurrency from a hardware and software combined perspective, protecting hardware load safety while ensuring local server throughput.Preferably, if the length of the judgment or ruling exceeds a preset character threshold (in this embodiment, the preset character threshold is 30,000 characters), a greedy grouping mechanism based on RAG block boundaries and a progressive context state passing mechanism is adopted within the legal litigation specialist team. Specifically, this includes greedy grouping of RAG block boundaries, progressive context state passing, and seamless splicing of core clauses. Greedy grouping of RAG block boundaries refers to binning the input text blocks to ensure that a single RAG text block is not forcibly truncated in the middle. Simultaneously, a grouping upper limit is set, and any excess is smoothly merged into the last group, ensuring the integrity of the document content. Completeness; Progressive transfer of contextual state refers to initiating a cascading extraction process for extremely long documents divided into multiple groups. Starting from the second group of text, the extracted results generated in the previous group are automatically and dynamically injected into the system prompts of the current group as known context. Seamless splicing of core clauses refers to using a model to combine the core facts extracted in the previous group with the text of the current group to determine the cross-block continuity of core result paragraphs such as "Judgment as follows" or "Ruling as follows," and performing word-by-word original text copying and splicing updates to ensure that judgment clauses in extremely long documents do not experience sentence breaks, repetitions, or content truncation when crossing block boundaries. Finally, after each vertical expert agent outputs a sub-report containing exclusive tags, the sub-reports from multiple sections are standardized and spliced ​​together. After each objective fact, calculated amount, and diagnostic conclusion in the response, a high-precision literature tracing mark is forcibly added.The intelligent claims processing and verification system provided in the first aspect of this embodiment adds source anchors after the response and uses a row-level rectangular bounding box coordinate and scale scaling mechanism on the front end to redraw a semi-transparent rectangular mask on the Canvas coating. This enables one-click viewport highlighting from the large model conclusion data to the specific text lines of the original scanned document. Reviewers do not need to manually search for images, reducing review and verification time by more than 80% and eliminating the claims black box. This embodiment also deploys various models and expert agents privately in a localized general-purpose server cluster within the insurance company's local area network. No public network interaction occurs throughout the entire process, fundamentally avoiding the legal compliance risks of leaking patient ID cards and sensitive medical records to the public cloud. It also completely eliminates public network latency and bandwidth fluctuations, reducing interface latency by more than 90% compared to cloud APIs, achieving financial-grade data privacy physical isolation and ultra-low latency on the local area network. Furthermore, it sets... The local physical memory hard-limit control lock hard-aligns the software's high concurrency with the physical graphics card's safe load, ensuring that the system will never crash due to graphics card overflow during high-concurrency case handling while guaranteeing the high throughput limit of the local area network's physical graphics cards, greatly improving the industrial-grade availability of the private deployment system; through adaptive sliding window progressive sorting channels and chain-style sliding splicing algorithms, it solves the bottleneck of traditional large models being limited by context token restrictions and unable to handle large-page documents in a single global sort, achieving accurate restoration of the time sequence of out-of-order judicial documents, with processing capabilities at an industry-leading level; through hierarchical restoration of outline trees, using node vector retrieval, a persistent mechanism that brings out the context from the original outline, and parallel execution of multiple vertical domain expert agents, each agent performs its own function, physically isolating medical data and financial accounts, and using machine deterministic reasoning to eliminate the confusion and illusions caused by traditional large models in policy payouts or drug projects.As shown, the second aspect of this embodiment provides a method for intelligent processing and verification of claims documents, which can be executed, but is not limited to, by a computer device or virtual machine with certain computing resources, such as a personal computer or smartphone, or by a virtual machine; the method for intelligent processing and verification of claims documents includes, but is not limited to, the following steps: S1. Obtaining a claims case package, performing multi-threaded concurrent reading and splitting of the claims case package to obtain multiple single-page vector documents, performing an affine transformation on the multiple single-page vector documents, and extracting fingerprints from the single-page image binary data obtained after the affine transformation to obtain a hash fingerprint set, and performing... Page order reconstruction yields claims analysis data; S2. Extract the binary byte stream of the claims analysis data and calculate its hash value. Compare the hash value with the hash fingerprint set. Input claims analysis data with different hash values ​​and hash fingerprint sets into the cascaded classification channel for pre-verification. If the pre-verification passes, output the mapping result. Extract the OCR text from the mapping result and perform initial classification screening on the OCR text. If it meets the initial classification screening rules, obtain the classification result and main title. Use a visual multimodal large model to perform physical analysis on the mapping result to obtain the document body type and identify and extract the main title text. S1. Input the main title or main title text into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, the corresponding document type is obtained. S2. Based on the document type, use the matching algorithm to calculate the text similarity ratio. Based on the similarity ratio, deduplicate the claim parsing data of different document types to obtain the claim label. S3. Use a large model to perform multi-task feature recognition on the claim parsing data corresponding to the claim label of legal litigation to obtain the recognition result. Logically verify the recognition result based on three underlying physical hard rules to obtain the verification result. Adaptively split the verification result to obtain the document data. S5. Extract page data from the claims analysis data, perform structured aggregation of page data and claims tags, construct a tree outline structure using a depth-first search algorithm based on the aggregation results, and stitch cross-page tables together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, obtain an outline index. S6. Based on document type, claims tags, document data, and the outline index, use the corresponding vertical expert agent to process the claims case package for claims review, obtaining one or more claims review opinions. When multiple claims review opinions are obtained, they are marked and deduplicated to obtain the final claims review opinion. This embodiment's third aspect provides a computer device, including a memory, a processor, and a transceiver connected in sequence. The memory stores a computer program, the transceiver sends and receives messages, and the processor reads the computer program to execute the claims document intelligent processing and claims review method as described in the second aspect.Specifically, the memory may include, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Flash Memory, First Input First Output (FIFO), and / or First Input Last Output (FILO), etc.; the processor may include, but is not limited to, a microprocessor of the STM32F105 series. Furthermore, the computer device may also include, but is not limited to, a power module, a display screen, and other necessary components. The working process, working details, and technical effects of the aforementioned computer device provided in the third aspect of this embodiment can be found in the intelligent processing and verification method for claims documents described in the second aspect, and will not be repeated here. The fourth aspect of this embodiment provides a computer-readable storage medium, that is, the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the intelligent processing and verification method for claims documents as described in the second aspect is performed. The computer-readable storage medium refers to a carrier for storing data, and may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The working process, details, and technical effects of the aforementioned computer-readable storage medium provided in the fourth aspect of this embodiment can be found in the intelligent processing and verification method for claims documents as described in the second aspect, and will not be repeated here. The fifth aspect of this embodiment provides a computer program product, including a computer program or instructions. When executed by a computer, the computer program or instructions are used to implement the intelligent processing and verification method for claims documents as described in the second aspect. The working process, details, and technical effects of the aforementioned computer program product provided in this embodiment can be found in the intelligent processing and verification method for claims documents as described in the second aspect, and will not be repeated here. Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart claims document processing and claims settlement system, characterized in that, This includes modules for splitting and reconstructing claims documents, classifying and deduplicating claims documents, reorganizing and sorting legal litigation documents, reconstructing hierarchical indexes, and generating claims reports. The claims document splitting and reconstruction module is used to obtain claims case packages, perform multi-threaded concurrent reading and splitting of the claims case packages to obtain multiple single-page vector documents, perform affine transformation on the multiple single-page vector documents, extract fingerprints from the single-page image binary data obtained after the affine transformation to obtain a hash fingerprint set, and reconstruct the page order of the multiple single-page vector documents to obtain claims parsing data. The claims document classification and deduplication module is used to extract the binary byte stream of claims parsing data, calculate the hash value of the binary byte stream, compare the hash value with the hash fingerprint set, and input claims parsing data with different hash values ​​and hash fingerprint sets into the cascade classification channel for pre-verification of claims parsing data. If the pre-verification passes, the mapping result is output, the OCR text in the mapping result is extracted, and the OCR text is judged by the classification screening rules. If it meets the classification screening rules, the classification result and main title are obtained. The mapping result is physically analyzed using a visual multimodal large model to obtain the document body type, and the main title text is identified and extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, the corresponding document type is obtained. The claims document classification and deduplication module is also used to calculate the text similarity ratio based on the document type using a matching algorithm, and to deduplicate the claims parsing data of different document types based on the similarity ratio to obtain claims tags; The legal litigation reorganization and sorting module is used to perform multi-task feature recognition on the claim parsing data corresponding to the claim label of legal litigation using a large model, obtain the recognition result, perform logical verification on the recognition result based on three underlying physical hard rules, obtain the verification result, and perform adaptive path diversion on the verification result to obtain the document data. The reconstructed hierarchical index module is used to extract page data from the claims analysis data, perform structured aggregation of page data and claims tags, construct a tree outline structure using a depth-first search algorithm based on the aggregation results, and stitch cross-page tables together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, an outline index is obtained. The claim review report generation module is used to process claim cases based on document type, claim tags, document data, and outline index, using corresponding vertical expert agents to obtain one or more claim review opinions. When multiple claim review opinions are obtained, they are marked and deduplicated to obtain the final claim review opinion.

2. The intelligent claims document processing and claims settlement system according to claim 1, characterized in that, It also includes a parsing mapping traceability module, which is used to obtain the final claim assessment opinion, index, deduplicate, and reconstruct the final claim assessment opinion to obtain a normalized reference data object. Based on the normalized reference data object, the module uses position fusion logic to extract the physical width of the original image and the physical width of the current target rendering view from the position metadata corresponding to the reference data object. Based on the physical width of the original image and the physical width of the current target rendering view, the module obtains the absolute pixel coordinates, calculates the vertical scroll offset of the page based on the absolute pixel coordinates, and smoothly centers and aligns the target highlighted area based on the vertical scroll offset of the page to achieve data traceability.

3. The intelligent claims document processing and claims settlement system according to claim 1, characterized in that, The claims case package is read and split using multi-threaded concurrent methods to obtain multiple single-page vector documents. An affine transformation is performed on these single-page vector documents, and fingerprint extraction is performed on the binary data of the resulting single-page images to obtain a hash fingerprint set. Finally, the page order of the multiple single-page vector documents is reconstructed to obtain the claims parsing data, including: For a single multi-page PDF file in a claims case package, extract the total number of pages in the single multi-page PDF file, and parse the single multi-page PDF file to obtain a single-page vector document and the original logical code index; A page-level concurrent thread pool is constructed based on the total number of pages, and the page-level concurrent thread pool includes multiple rendering sub-threads; In each rendering sub-thread, an affine transformation matrix is ​​used to scale and render a single-page vector document to obtain multiple JPEG images. Read the single-page binary data of the JPEG image and use the MD5 digest algorithm to calculate the hash fingerprint of the single-page binary data to obtain the hash fingerprint of each rendering sub-thread. Based on multiple hash fingerprints, obtain a hash fingerprint set. The page order of multiple JPEG images is rearranged based on the preset original logical page number index to obtain the claims parsing data.

4. The intelligent claims document processing and claims settlement system according to claim 1, characterized in that, The hash value is compared with the hash fingerprint set. Claims parsing data with different hash values ​​and hash fingerprint sets are input into the cascading classification channel for pre-verification of the claims parsing data. If the pre-verification passes, the mapping result is output, including: The hash value is compared with the hash fingerprint set. If the hash value is the same as the hash fingerprint in the hash fingerprint set, the claim parsing data is marked as a duplicate file and subsequent operations are terminated. If the hash value is different from the hash fingerprint in the hash fingerprint set, the claims parsing data is input into the cascaded classification channel to use OCR to recognize the claims parsing data, obtain OCR text, and perform physical feature comparison verification on the OCR text. If the physical feature comparison verification passes, the first mapping result is obtained. If the physical feature comparison verification fails, but the preset invoice feature words exist in the OCR text, then based on the exclusive filtering rules and compliance review, the invoice-specific large model is used to filter and review the OCR text to obtain the second mapping result. Output either the first mapping result or the second mapping result as the mapping result.

5. The intelligent claims document processing and claims settlement system according to claim 1, characterized in that, Extract the OCR text from the mapping results and perform initial classification screening. If the OCR text meets the initial classification screening rules, the classification result and main title are obtained. A visual multimodal large model is used to perform physical analysis on the mapping results to obtain the document body type. Title text recognition is then performed on the mapping results to obtain a Boolean value. If the Boolean value is 1, the corresponding main title text is extracted. The main title or main title text is input into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to a preset similarity threshold, the corresponding document type is obtained, including: The mapping results are used to extract text, resulting in OCR text. The OCR text is initially classified and screened to determine whether it belongs to a pure image page with no / effective text. If it does not belong to the category, the similarity between the title text in the OCR text and the preset core category keyword corpus is calculated. If the similarity score is greater than the preset similarity threshold, the classification result and the main title in the title text are output. If the mapping results are not subject to initial classification and screening rules, a visual multimodal large model is used to perform physical analysis on the page structure of the mapping results to obtain the document body type and identify and extract the main title text in the mapping results. Calculate the similarity between the main title text and the preset corpus of keywords for each core category. If the similarity is greater than or equal to the preset similarity threshold, the corresponding document type is obtained.

6. The intelligent claims document processing and claims settlement system according to claim 1, characterized in that, After obtaining the corresponding document type, the following is also included: Based on document type, extract the scattered images and single pages of PDF from the claims analysis data to obtain a summary of pages to be processed; The pages to be processed are aggregated and input into a concurrent queue. In the concurrent queue, the aggregated pages to be processed are packaged according to preset packaging rules. Based on strong exclusivity and adversarial rules, a shared global large model inference thread pool is used to input the packaged aggregated pages to be processed into a large language model for classification, resulting in a standard JSON classification array. The standard JSON classification array is used as the classification result. During the classification process, based on legal litigation interception rules, claims parsing data containing preset judge-specific terms or legal procedure mediation features in the packaged aggregated pages to be processed are forcibly prioritized as claims tags for legal litigation, so as to perform multi-task feature recognition on the claims parsing data corresponding to legal litigation.

7. The intelligent claims document processing and claims settlement system according to claim 5, characterized in that, A large model is used to perform multi-task feature recognition on claims data tagged as legal litigation, yielding recognition results. Logical verification of these results is then performed based on three underlying physical hard rules, resulting in verification results. Adaptive path splitting is applied to these verification results to obtain document data, including: For claims analysis data tagged as legal litigation, preset characters from the claims analysis data are extracted as footer features. The footer features and recognition Boolean values ​​are then structurally assembled to obtain the assembled data. The assembled data is input into a large model for multi-task feature recognition to obtain recognition results, which include the calibration status of the homepage information. Logical verification of the recognition results is performed based on three underlying physical hard rules, which are used to extract page numbers, correct the first page, and filter documents to obtain the verification results. Based on the calibration status of the homepage information in the verification results, adaptive path diversion is performed on the legal litigation single-page image to obtain the document data.

8. The intelligent processing and verification system for claims documents according to claim 1, characterized in that, The claims labels include the medical record cover page, discharge certificate, surgical record, imaging examination report, outpatient / emergency room record, legal proceedings, identity information, disability / death certificate, accident report, policy information, medical invoices, and other invoices.

9. The intelligent processing and verification system for claims documents according to claim 8, characterized in that, The vertical expert AI includes experts in case and underwriting characterization, medical record front page, outpatient / emergency room records, inpatient discharge records, imaging reports, disability and death assessment, and legal and litigation specialists. The cases and underwriting experts were used to analyze the accident determination reports in the claims case package and extract the accident liability ratio. Based on the accident liability ratio, type of insurance, deductible, and reimbursement ratio, the scope of insurance liability was obtained. The medical record homepage expert is used to extract information and days from the claims package to obtain the patient's basic information and actual number of days hospitalized; The outpatient / emergency room record expert is used to sort and organize the medical records in the claims package based on a timeline to obtain a sequence of medical cards; The hospital discharge record expert is used to record and extract information from the claims package to obtain the discharge diagnosis and medical advice; The image report expert is used to rearrange the CT, DR, and / or MRI reports in the claims package to obtain rearranged image reports; The disability and death assessment experts are used to extract the number, assessment institution, disability level, lost work, nursing care, and nutritional period conclusions from the judicial appraisal reports in the claims package; The legal and litigation specialist is used to classify cases in the claims package into mediation agreements, rulings, or judgments based on title characteristics.

10. A method for intelligent processing and verification of claims documents, applied to the system according to any one of claims 1 to 9, characterized in that, include: The process involves acquiring a claims case package, performing multi-threaded concurrent reading and splitting of the claims case package to obtain multiple single-page vector documents, performing affine transformations on the multiple single-page vector documents, extracting fingerprints from the binary data of the single-page images obtained after the affine transformation to obtain a hash fingerprint set, reconstructing the page order of the multiple single-page vector documents, and obtaining the claims parsing data. Extract the binary byte stream of the claims parsing data and calculate its hash value. Compare the hash value with the hash fingerprint set. Input the claims parsing data with different hash values ​​and hash fingerprint sets into the cascaded classification channel for pre-verification. If the pre-verification passes, output the mapping result. Extract the OCR text from the mapping result and perform a preliminary classification screening rule on the OCR text. If it meets the preliminary classification screening rule, obtain the classification result and main title. Use a visual multimodal large model to perform physical analysis on the mapping result to obtain the document body type and identify and extract the main title text. Input the main title or main title text into the TF-IDF pre-detection vector space for similarity matching. If the similarity score is greater than or equal to the preset similarity threshold, obtain the corresponding document type. Based on document type, a matching algorithm is used to calculate the text similarity ratio. Based on the similarity ratio, duplicate claims data of different document types are deduplicated to obtain claims tags. The large model is used to perform multi-task feature recognition on the claims parsing data corresponding to the claims label of legal litigation to obtain the recognition results. The recognition results are logically verified based on three underlying physical hard rules to obtain the verification results. The verification results are then adaptively routed to obtain the document data. The claims analysis data is extracted from the page data, and the page data and claims tags are aggregated in a structured manner. Based on the aggregation results, a tree outline structure is constructed using a depth-first search algorithm, and cross-page tables are stitched together to obtain a complete table. Based on the tree outline structure, the complete table, and the claims tags, an outline index is obtained. Based on document type, claim tags, document data, and outline index, the corresponding vertical expert AI agent is used to process the claim case package for claim review, resulting in one or more claim review opinions. When multiple claim review opinions are obtained, they are marked and deduplicated to obtain the final claim review opinion.