A Trusted Extraction and Verification Method Based on Document Object Composite Evidence Anchors

Abstract

This invention relates to the field of artificial intelligence document processing technology, and particularly to a reliable extraction and verification method based on composite evidence anchors for document objects. This method parses the document to be processed into a set of document objects, generates a composite evidence anchor for each document object, performs object-level retrieval according to user tasks, and assembles a minimum sufficient context package with evidence anchors. The output of the large model is split into atomic facts and each fact is bound to a candidate evidence anchor. Based on the evidence anchors, the original text objects are relocated, and a deterministic program performs joint verification of semantics, numerical values, object type, structural path, coordinates, and hash, obtaining a comprehensive consistency score and a failure vector. If the verification fails, neighborhood expansion, object-level re-retrieval, structural relocation, or manual review is performed based on the failure vector, and the anchor state is updated. Finally, the verified results, evidence location information, and audit logs are output. This invention can improve the location accuracy, auditability, and error correction capabilities of complex mixed-format document processing results.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence document processing technology, and in particular to a reliable extraction and verification method based on composite evidence anchors of document objects. Background Technology

[0002] With the development of large language models and visual language models, document-oriented technologies such as automatic parsing, information extraction, question answering, summarizing, and proofreading are increasingly being applied in scenarios such as enterprise office work, legal contracts, financial auditing, scientific research and publishing, government document processing, and engineering and technical document management. Unlike ordinary natural language text, real-world business documents typically contain body paragraphs, heading levels, tables, charts, formulas, footnotes, annotations, headers and footers, attachment references, and cross-page continuation structures. The model needs to understand not only the semantics of the text but also the document's layout structure, object relationships, and source location.

[0003] This type of document processing scenario has the following technical unique characteristics. First, the document content has obvious object-oriented and hierarchical characteristics. The same semantic fact may be distributed among the title, body text, table cells, and attachment clauses, making it difficult to maintain its original structural relationship using only ordinary text fragments. Second, the facts in the document have physical location attributes. For example, a certain amount may be located in the third row and fourth column of a table on page 5, a certain experimental conclusion may be located in the caption below a chart, and a certain formula variable definition may be located in the explanatory paragraph after the formula. These location attributes directly affect the efficiency of fact verification and manual review. Third, the document processing results usually need to be auditable, especially in contract review, financial report verification, tender document response, and technical report review. Users not only focus on the content of the answer but also require quick access to the original text to verify whether the answer is based on evidence. Fourth, scanned documents, converted documents, and multi-version documents may contain problems such as OCR recognition errors, incorrect page reading order, cross-page table splicing errors, and partial content tampering. If the model output lacks an underlying evidence verification mechanism, source parsing errors can easily be amplified into errors in the final conclusion.

[0004] In existing technologies, one approach uses OCR recognition, layout analysis, and text extraction techniques to convert PDF, image, or Word documents into text blocks, which are then input into a large model for summarization, question answering, or field extraction. While this approach can improve the efficiency of automatic processing of unstructured documents, it typically only retains paragraph-level text in complex document layouts. This makes it difficult to accurately describe the correspondence between table cells, formulas, figure captions, cross-page objects, and the original coordinates, and also makes it difficult to determine whether the original positioning information remains valid after document version changes.

[0005] Another approach employs retrieval enhancement generation techniques, dividing documents into several text segments and using vector retrieval, keyword retrieval, or hybrid retrieval to recall relevant segments, before generating responses using a large language model. For example, patent CN119988588A discloses a multimodal document retrieval enhancement generation method based on a large model, involving document parsing, recursive segmentation, vectorization, and response generation based on the large model; patent CN120256648A discloses a retrieval enhancement processing method based on a large model, involving knowledge document parsing and segmentation, multi-path recall, and ranking processing. While these approaches can alleviate the problem of large models lacking external knowledge, they typically use text blocks or knowledge segments as retrieval and citation units, making it difficult to perform document object-level evidence binding and automatic verification for each generated fact.

[0006] Another approach involves attaching source citations to the model's response or matching answer fragments with original text fragments using text similarity after the response is generated, thus achieving a certain degree of source tracing. For example, patent CN120561252B discloses a method for tracing the source of evidence in medical model responses. It semantically segments the response content into multiple fine-grained fragments and then matches them with knowledge text to obtain source tracing evidence. This type of approach can improve the interpretability of the response, but its source tracing objects are usually knowledge text or text fragments, making it difficult to simultaneously verify the physical coordinates, logical structure, object type, text offset, content integrity, and version status of document objects.

[0007] Furthermore, document data extraction solutions for specific scenarios have begun to utilize intermediate document representations, table row and column coordinates, page numbers, and candidate numerical evidence. For example, patent CN121880434A discloses a data extraction method that converts a financial report sub-document into an intermediate representation containing table and paragraph elements, and verifies and fuses candidate numerical evidence; US20250077554A1 discloses the use of position embedding in document information extraction, which can output the bounding box containing the text that leads to the generation of the answer. The above solutions illustrate that page numbers, bounding boxes, candidate numerical evidence, and citation output have gradually become common technical elements in intelligent document processing, but there is still a lack of composite evidence anchor state management, atomic fact-level mandatory binding, and multi-dimensional failure repair closed loops for complex documents of various types.

[0008] Existing technologies suffer from at least the following shortcomings. First, the granularity of evidence sources is relatively coarse, typically remaining at the level of documents, pages, text blocks, or single bounding boxes. They cannot reliably locate document objects such as table cells, formula areas, figure captions, annotations, or paragraphs spanning multiple pages. Users still need to manually search for the original text when reviewing the model output, resulting in low auditing efficiency. Second, the model output lacks atomic fact-level constraints. Multiple facts are often combined into a single natural language response, making it difficult for the system to determine which original text objects support each amount, date, subject, indicator, conclusion, and clause citation. This makes it difficult to detect local illusions or errors. Third, existing source tracing verification focuses primarily on semantic similarity or numerical confidence, lacking a joint verification mechanism that considers physical coordinates, structural paths, object types, text offsets, content multi-summary hashing, and version status. This makes it difficult to distinguish between different technical issues such as text content modification, layout re-encoding, OCR re-recognition, structural path drift, and document version replacement. Fourth, existing systems typically only indicate that the answer is unreliable or require regeneration after verification failure. They lack a rule-based closed-loop processing mechanism that selects neighborhood retrieval based on the failure vector, object-level re-retrieval, structural path relocation, OCR region re-identification, evidence anchor status update, and audit log recording. As a result, similar errors recur in subsequent tasks. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, this invention provides a trusted extraction and verification method based on document object composite evidence anchors.

[0010] The technical solution of this invention is as follows: A trusted extraction and verification method based on document object composite evidence anchors includes the following steps: S1, parse the document to be processed into a collection of document objects O={o1,o2,...,o...} n}, each document object o i At least include object text c i Object type t i Page p i coordinate frame b i , structural path x i and reading order r i ; S2, for each document object o i Generate composite evidence anchor A i Composite evidence anchor A i It includes at least document identifier, version identifier, page number, rectangular coordinate frame, polygon coordinate, document object type, logical structure path, native object path, text offset, multi-digest hash, page object Merkle root, and evidence anchor state. The multi-digest hash includes text content hash, structure path hash, geometric location hash, and object image region-aware hash, and each digest is calculated independently. S3, retrieve candidate document objects from the document object set O according to the user task, and assemble the candidate document objects and their composite evidence anchors into a minimum sufficient context package; S4, the large model generates structured results based on the minimum sufficient context package, and then decomposes the structured results into a set of atomic facts F={f1,f2,...,f m}, and for each atomic fact f j Bind one or more candidate composite evidence anchors; S5, based on candidate composite evidence anchor A i Relocate the original object o i Read the current object text, coordinates, structural path, native object path, object type, and digest hashes of the original object, and use a deterministic program to process the atomic facts f. j Perform a joint check of semantic consistency, numerical consistency, object type consistency, structural path consistency, coordinate validity, and hash integrity to obtain the overall consistency score C(f). j A i and failure vector E j = <e sem ,e num ,e type ,e struct ,e geo ,e hash >The self-assessed confidence level of the large model output is not used as the basis for passing the verification; S6, when the overall consistency score C(f) j A i When the value is below the confidence threshold τ or a severe failure dimension exists, the failure vector E is used as the basis for determining the failure level. j Perform neighborhood expansion retrieval, object-level re-retrieval, structure relocation, OCR region re-identification or manual review, and update the evidence anchor status of the corresponding composite evidence anchor; S7 outputs the location information, anchor status, overall consistency score, and audit log of the atomic facts that pass the verification and their composite anchor evidence. It marks the status of atomic facts that fail the verification but does not output them as credible conclusions.

[0011] Furthermore, to better implement this invention, the parsing in S1 specifically involves: performing page-level preprocessing on the document to be processed, including page rotation correction, skew correction, resolution normalization, noise removal, and page number recognition; reading the original structure of the parsable electronic document, wherein the original structure includes paragraphs, runs, tables, comments, footnotes, and style tags in OpenXML; calling the OCR recognition model and layout analysis model on scanned images or image-type documents to obtain text lines, paragraph blocks, table areas, image areas, formula areas, and header / footer areas; the object type t iIncludes headings, body paragraphs, list items, tables, table rows, table cells, images, figure captions, formulas, variable explanations, footnotes, comments, headers, footers, and attachment references; the coordinate frame b i Rectangular coordinates (x) min ,y min ,x max ,y max For irregular regions, use polygon coordinates q. i Indicates; the structural path x i The hierarchical path of the object in the document tree.

[0012] Furthermore, to better implement the present invention, the parsing of table objects, formula objects, and chart objects in S1 specifically involves: identifying the header level, merged cells, row and column coordinates, page continuation relationships, and unit fields for table objects, and generating an independent document object for each table cell; identifying formula expressions, variable names, variable explanation paragraphs, and reference text near the formula for formula objects; and identifying chart titles, captions, axis names, legends, and data series or trend descriptions for chart objects.

[0013] Furthermore, in order to better realize the present invention, the composite evidence anchor A in S2 i Its multi-digest hash is specifically as follows: , Where D represents the document identifier, V represents the document version identifier, and P represents the document version identifier. i Indicates the page number where the object is located, B i Represents a rectangular coordinate frame, Q i T represents the coordinates of a polygon or quadrilateral. i Indicates the document object type, X i R represents the logical structure path. i Indicates the path to the native object, U i S represents the text offset of an object's text within its parent object or native document structure. i Indicates the anchor state of evidence; text content hash H text =Hash(N(c i )), where N(c i This represents the text after performing Unicode normalization, whitespace normalization, and punctuation normalization on the object text; the structure path hash H struct =Hash(T i ||X i ||R i ); Geometric location hash H geo =Hash(P i ||B i ||Q i Image region-aware hashing Himg =pHash(Image(P i B i Page object Merkle root M page H of multiple objects on the same page text H struct and H geo It is calculated from the bottom up.

[0014] Furthermore, in order to better implement the present invention, the evidence anchor state S in S2 i Specifically, this includes: unverified, located, semantically passed, numerically passed, structurally drifted, coordinate invalid, text hash invalid, image hash invalid, suspected OCR error, awaiting manual review, repaired, and obsolete; the evidence anchor status is updated after each retrieval, generation, verification, and repair, and the status transition time, triggering task, failure vector, and repair action are recorded; and multiple composite evidence anchors are allowed to form an evidence anchor group G. k ={A a A b In a table spanning multiple pages, multiple cells, formulas and their variable explanation paragraphs, charts and their captions each form an evidence anchor group, and the evidence anchor groups are set to a uniform state.

[0015] Furthermore, to better implement the present invention, the retrieval in S3 specifically involves calculating the comprehensive retrieval score R(q,o) of the document object according to the following formula. i And select candidate document objects based on the comprehensive retrieval score: , Among them, S sem Indicates the semantic similarity between a user's question and the object's text or object summary; S struct This indicates the degree of match between the chapters, clauses, tables, or attachments in a user question and the object structure path; S type Indicates the degree of matching between the user task type and the object type; S geo S represents the page proximity between the target object and the matched object; hist This represents the historical credibility of the object being successfully verified in similar tasks; α, β, γ, δ, and η are non-negative weights, and α+β+γ+δ+η=1.

[0016] Furthermore, to better implement the present invention, the minimum sufficient context package in S3 specifically includes candidate object text, object type, page number, structural path, adjacent objects, table row and column relationships, formula variable explanations, and figure annotations, and the total number of tokens does not exceed a preset upper limit; for tasks that require cross-object understanding, the neighboring objects are expanded according to the inclusion edges, reading order edges, cross-page continuation edges, figure and text reference edges, and formula variable edges between document objects to form a local evidence subgraph.

[0017] Furthermore, to better implement this invention, the binding of atomic facts and evidence anchors in S4 specifically involves: requiring the large model to output atomic facts through constraint hints, structured output patterns, or function call interfaces, with each atomic fact f... j The corresponding structured record should include at least the fact text, fact type, normalized value, unit, confidence level, candidate evidence anchor list, and reason for generation. Binding rules are executed according to fact type: monetary facts are bound to table cells or paragraph evidence anchors containing currency, amount field name, or price table position; date facts are bound to paragraphs, table cells, or annotation evidence anchors containing date text and contextual roles; subject facts are bound to contract header, signature area, business name field, or attachment subject description evidence anchors; clause facts are bound to clause number, title path, and adjacent paragraph evidence anchors; conclusion facts are bound to evidence anchor groups consisting of data cells, formula objects, and conclusion paragraphs. When there is a conflict between multiple candidate evidence anchors, the support is calculated according to the evidence anchor status, object type matching degree, version identifier, hash integrity, and business priority, and the evidence anchor with low support is marked as a conflict candidate.

[0018] Furthermore, to better implement the present invention, the joint verification in S5 is specifically as follows: the overall consistency score C(f) is calculated according to the following formula. j A i ): , Among them, C sem This indicates the semantic consistency between the factual text and the original text; C num Indicates the degree of matching after normalization of numerical values, units, dates, proportions, or formula calculation results; C type Indicates the degree of matching between the fact type and the document object type; C struct Indicates the degree of consistency between the structure path and the native object path; C geo Indicates whether the anchor coordinates of the evidence are locatable on the current page and fall within the corresponding object area; C hash H represents text H struct H geo H img and M pageCombinatorial integrity; λ1 to λ6 are weight parameters, and λ1+λ2+λ3+λ4+λ5+λ6=1; when C(f j A i If the atomic fact is greater than or equal to the confidence threshold τ and there is no serious failure dimension, it is determined that the atomic fact is supported by the evidence anchor; otherwise, it proceeds to the failure cause determination. The failure vector E j Each component takes the value of 0, 1, or 2, where 0 indicates pass, 1 indicates slight abnormality, and 2 indicates severe abnormality.

[0019] Furthermore, to better implement the present invention, the failure cause determination in S5 and the failure repair in S6 are specifically as follows: if H text Changes and H geo If essentially unchanged, it is judged as suspected content modification; if H geo Changes and H text If consistent, it is judged as a suspected layout rearrangement; if H img Change and H text Changes are identified as suspected OCR or changes in scan quality; if M page Changes are identified as page-level version changes; if the object pointed to by the structural path does not exist, it is identified as structural drift; accordingly, when semantic consistency is insufficient but coordinates, structure, and hash are valid, a neighborhood expansion search is performed on the same chapter, the same table row and column, adjacent figure captions, or formula variable interpretation objects; when numerical consistency is insufficient, the original text object is reread and numerical normalization, unit conversion, table cell positioning correction, or formula recalculation is performed; when object types do not match, paragraph routing is switched to table cell routing, formula routing, or annotation routing; when structural drift occurs, the original object path, heading level, adjacent object hash, and reading order are used for repositioning; coordinates are invalid or hashes are inconsistent. When a document is found to have changed version, content, or been reformatted, a prompt is issued, and relevant evidence anchors are regenerated or discarded. If an OCR error or layout parsing error is determined, the system returns to the original page image area based on the coordinate frame in the evidence anchor, and performs cropping, enlargement, binarization, noise reduction, rotation correction, or multi-OCR engine re-recognition on that area. If the re-recognition result is consistent with the context structure, the document object text and evidence anchor hash are updated; otherwise, a manual review task is output. An audit log is generated for each failed repair process, including the user task, candidate evidence anchors, failure score, failure reason, repair action, object text before and after repair, evidence anchor after repair, and final credibility status.

[0020] The beneficial effects of this invention are: (1) Since a composite evidence anchor containing physical coordinates, logical structure, native path, multi-summary hash and status information is generated for each document object such as title, paragraph, table cell, formula, figure caption and annotation, the model output can directly return to the original page and object area, and can determine whether the object has become invalid due to version change, reflow or OCR change. Therefore, the evidence positioning granularity can be improved from ordinary text block to verifiable document object level.

[0021] (2) Since the model output is broken down into independently verifiable atomic facts and each atomic fact is required to be bound to one or more composite evidence anchors, it is possible to discover local errors in the amount, date, subject, clause number and conclusion judgment item by item, thereby reducing the risk that local illusions in long natural language responses are masked by the whole.

[0022] (3) Due to the use of deterministic joint verification of semantics, numerical values, object type, structural path, coordinates and multi-digest hash, the system can not only determine whether the answer is consistent with the semantics of the original text, but also distinguish different technical problems such as unit conversion error, table cell mismatch, structural path drift, page reflow, document version change, local content tampering and OCR area drift, thereby improving the credibility of the processing results of complex mixed text documents.

[0023] (4) Since the system can form a closed loop from error discovery to error repair by triggering neighborhood expansion, object-level re-retrieval, structure relocation, OCR region re-identification, evidence anchor update or manual review tasks based on the failure vector and evidence anchor status after verification failure, it can reduce the recurrence of the same type of error in subsequent tasks.

[0024] (5) Since the output results include the original text excerpt, page number, coordinate frame, structure path, native object path, consistency score, multi-summary verification status and audit log, enterprise users can establish a traceable, explainable and verifiable large model document processing flow in scenarios such as contract review, financial verification, technical report review and tender document response.

[0025] (6) Since the system only outputs the atomic facts that have passed the verification as the trusted result and performs status labeling or repair processing on the facts that have not passed the verification, it can reduce the probability of illusion propagation in large models in complex document extraction and question answering, reduce the scope of manual review, and improve the controllability of the document intelligent processing system in high-risk business scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the document object parsing and composite evidence anchor generation process of the present invention; Figure 3This is a schematic diagram of the object-level retrieval and minimum sufficient context assembly process of the present invention; Figure 4 This is a schematic diagram of the state machine for multidimensional consistency verification, failure vector determination, and repair in this invention; Figure 5 This is a schematic diagram of the system module architecture of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0028] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0029] See Figure 1 The method of this invention comprises seven stages: document object parsing, composite evidence anchor generation, object-level retrieval and context assembly, atomic fact generation and evidence anchor binding, multi-dimensional consistency verification, failure repair and closed-loop update, and output and audit presentation, corresponding to steps S1 to S7 respectively. See also Figure 2 The document object parsing stage outputs a set of document objects O, and the composite evidence anchor generation stage outputs a composite evidence anchor A for each document object. i Organize page objects by page, Merkle root M page See also Figure 3 The object-level retrieval process is based on the comprehensive retrieval score R(q,o). i After selecting candidate objects, expand the neighborhood along containment edges, reading order edges, page-crossing continuation edges, image / text reference edges, and formula variable edges to assemble the minimum sufficient context package. See also Figure 4 The multidimensional consistency verification process involves a deterministic procedure that calculates the overall consistency score C(f). j A i And output the failure vector E. j Repair state machine based on E j The component values ​​are selected to determine the corresponding repair action, driving the evidence anchor state to transition between unverified, located, semantically passed, numerically passed, structurally drifted, coordinate invalid, text hash invalid, image hash invalid, suspected OCR error, awaiting manual review, repaired, and obsolete. See also Figure 5In terms of system implementation, this invention can be deployed as a local document processing service, an enterprise knowledge base service, a browser plugin, an office software plugin, or a cloud API service; the document object library, evidence anchor index, vector index, and audit log can be stored separately or merged in a database that supports full-text search, vector search, and structured query; the large model can be a locally deployed model, a private model, or a cloud model, and the evidence anchor generation module and consistency verification module run independently as deterministic programs to reduce the impact of the instability of the large model on the final credible result.

[0030] Example 1: Reliable Questions about Payment Terms in Contract Documents This example uses a 12-page PDF contract as the processing object. The contract includes the main text terms, an attached price list, scanned signature pages, and two annotations. The user asks: "What are the total price, payment terms, and late payment penalty percentages of this contract, and on which page are they based?" The system first performs page-level parsing of the contract, identifying 16 title objects, 184 main text paragraph objects, 3 table objects, 126 table cell objects, 2 annotation objects, and 1 signature image object.

[0031] During the composite evidence anchor generation stage, the system generates evidence anchor A for the "Total Contract Price" cell in the price table attached to page 5. price Its page number is 5, the coordinate frame is (112, 436, 482, 472), the polygon coordinates are the four vertices of the table cell, the object type is "table cell", the logical structure path is "Attachment 1 / Price List / Total Row / Contract Total Price Column", the native object path is the PDF page object and the table row and column indexes, and the text offset is the 0th to 13th characters within the cell. The system calculates H respectively. text H struct H geo and H img And combine the hashes of all objects on page 5 into a page Merkle root M page The system generates evidence anchor A for the payment deadline paragraph in section 7.2 on page 4. pay Its coordinate frame is (86, 216, 515, 267), the object type is "Text Paragraph", and the structure path is "Article 7 / Section 7.2 / Payment Deadline". The system generates evidence anchor A for the breach of contract paragraph 12.3 on page 9. penalty Its structural path is "Article 12 / Section 12.3 / Liability for Overdue Breach of Contract".

[0032] During the object-level retrieval phase, the system identifies three user intents: "total price," "payment period," and "late payment penalty percentage." For the "total price" intent, the object type matching degree S of the table cell object... type Set the semantic similarity to 0.95. semSet the structural path matching degree S to 0.88. struct Taking 0.82, in this embodiment, we take α=0.4, β=0.3, γ=0.3, δ=0, and η=0. Therefore, R=0.4×0.88+0.3×0.82+0.3×0.95=0.883, resulting in a comprehensive retrieval score R of 0.883. For the intent of "payment period," the comprehensive retrieval score R of paragraph 7.2 is 0.912; for the intent of "penalty percentage," the comprehensive retrieval score R of paragraph 12.3 is 0.904. The system assembles the above three objects, along with their adjacent headings, table headers, and clause contexts, into a minimum sufficient context package, which is then input into the large language model.

[0033] The large language model generates three atomic facts: F1 is "The total contract price is RMB 3,850,000", the fact type is amount, the normalized value is 3,850,000, the unit is yuan, and the candidate evidence anchor is A. price F2 states "The payment period is within 30 calendar days after Party B issues a valid invoice," the fact type is time limit, the normalization value is 30, the unit is calendar days, and the candidate anchor of evidence is A. pay F3 states "Overdue payment penalty is calculated at 0.05% per day of the outstanding amount," with the fact type being percentage, a normalized value of 0.0005, and the unit being daily. The candidate anchor of evidence is A. penalty .

[0034] During the consistency verification phase, the system relocates A. price In the corresponding table cell, normalize "3,850,000 yuan" to 3,850,000 yuan. num Returns 1.00; the fact type "Amount" matches the object type "Table Cell", C type Take 0.96; both the structure path and the native object path can be located, C struct Set to 1.00; the coordinate region can be repositioned on page 5, C geo Take 1.00; H text H struct H geo and M page All are consistent, C hash The value is set to 1.00; the overall consistency score C is 0.974, which is higher than the confidence threshold τ = 0.85. The system will assign A... price The status changed from "Unverified" to "Value Passed". The overall consistency scores for F2 and F3 were 0.943 and 0.938 respectively, both passing the verification. The system ultimately output three reliable answers, highlighting the total cell of the price list on page 5, item 7.2 on page 4, and item 12.3 on page 9 in the user interface.

[0035] Example 2: Reliable Extraction of Table Values ​​and Text Conclusions from Test Reports This embodiment uses an 18-page product testing report as the processing object. Page 8 of the report contains a performance test data table, page 10 contains calculation formulas, and page 11 contains a conclusion paragraph. The user requests the system to extract whether the "average response time meets the requirement of not exceeding 200 milliseconds, and provide the basis." The system first identifies 10 rows of test records in the performance test table, each row including the sample number, test round, response time, and test ambient temperature.

[0036] The system generates evidence anchor A for each of the 10 table cells in the response time column. t1 To A t10 And establish evidence anchor group G for these evidence anchors. time Assume the values ​​in this column are 185, 191, 188, 196, 190, 187, 193, 189, 192, and 186, all in milliseconds. The system also generates evidence anchor A for the formula object on page 10. formula The formula object text is "average response time T". avg =(ΣT i ) / n”, its variable explanation paragraph explains T i Let be the response time for the i-th test, and n be the number of tests. The system also generates evidence anchor A for the conclusion paragraph on page 11. conclusion The original text states that "the average response time of the sample under the specified test environment is less than 200 milliseconds, which meets the requirements."

[0037] The large language model generates atomic facts F4 based on context packets: "The average response time is 189.7 milliseconds"; F5: "The average response time is less than 200 milliseconds"; F6: "The sample meets the response time requirement". The system binds F4 to the evidence anchor group G. time and A formula Bind F5 to G time A formula And the evidence anchor of the paragraph where the indicator threshold is located, bind F6 to G. time A formula and A conclusion The system performs a numerical consistency check on F4, summing the 10 response time cells to obtain 1897. Dividing this by the sample size of 10 yields a result of 189.7 milliseconds, consistent with the normalized F4 value. Therefore, C... num Set the value to 1.00. The system performs a threshold comparison on F5, and 189.7 < 200, which is valid. For F6, a conclusion support verification is performed, confirming that the conclusion paragraph on page 11 is consistent with the calculation result.

[0038] In this embodiment, if the large language model initially outputs "average response time is 198.7 milliseconds", semantic consistency may still be high, but numerical consistency C will be low. numThe score will drop to 0.25, and the failure vector will be <0,2,0,0,0,0>. At this point, even if the other five dimensions are all perfect, the overall consistency score C=0.875 is still higher than τ=0.85, but due to the existence of the severe failure dimension (e... num =2), based on the dual criteria of "C≥τ and no serious failure dimension", the system still determines that this fact fails the validation. The system will not output this fact as a reliable result, but will instead trigger numerical failure repair, rereading A. t1 To A t10 The corresponding table cells were processed and the formula was executed to obtain the correct value of 189.7 milliseconds. Then, the large language model was asked to perform a partial rewrite. Through this process, the core of the system is not ordinary table extraction, but rather the use of evidence anchor groups and deterministic calculations to discover and correct numerical illusions in the large model.

[0039] Example 3: OCR error localization and evidence anchor update in the tender document This example uses a scanned tender document as the processing object. Page 23 of the document contains a delivery cycle commitment table. Due to a slight tilt in the scanned image and obscuring of the seal, the initial OCR result misidentifies "45 calendar days" as "4S calendar days". The user inquires: "What is the promised delivery cycle in the tender document?" The system initially retrieves table cell A on page 23. delivery The atomic fact F7, "Delivery cycle is 4S calendar days," is generated by the large language model.

[0040] The consistency check module found that the fact type of F7 was a time limit, but the normalized numeric parsing failed, and the character "S" could not be converted into a number; at the same time, A delivery H geo Consistent with historical records, H struct Consistent with the table's row and column paths, while H text With H img All errors occurred, with a failure vector of <0,2,0,0,0,2>, indicating that there might be an OCR character recognition error in this area rather than an object location error. Based on this, the system will... delivery The status is updated to "suspected OCR error", and the region is cropped from the original page image according to the evidence anchor coordinate box (326,512,458,548). The cropped image is then magnified by 1.5 times, tilted and binarized, and the second OCR engine is called to re-identify.

[0041] The re-identification result was "45 calendar days". The system verified this result against the table row header "Delivery Period" and column header "Commitment Value", confirming that the result was reasonable within the context. The system then updated the text of the table cell object and recalculated H. text and H imgThe evidence anchor version identifier was updated from V1 to V2, and the anchor status was changed from "suspected OCR error" to "fixed". The large language model regenerates atomic facts F based on the updated objects. 7prime "The delivery cycle is 45 calendar days," and the consistency score C is 0.963. The audit log records the original OCR results, cropping coordinates, image enhancement strategies, re-identification results, hashes before and after the evidence anchor update, state transitions, and the final trusted state.

[0042] As can be seen from this embodiment, the present invention can not only trace the source of the model's answer after the fact, but also reverse locate the original document object when the answer is found to be unreliable, and perform local repair for OCR or layout parsing errors, thereby avoiding the propagation of source identification errors to the final question and answer results.

[0043] Example 4: Simulated Comparative Test of Batch Verification of Multiple Document Types This embodiment is a simulated test scenario used to illustrate the technical effects of the present invention. The test set is assumed to include 100 contracts, 50 test reports, and 30 financial reports, generating a total of 3600 atomic facts through extraction or question-and-answer processing. These facts include 900 monetary facts, 620 date-related facts, 480 subject-related facts, 700 clause-related facts, 500 calculation-related facts, and 400 conclusion-related facts. The tests employed a standard text block RAG scheme, a page number and coordinate reference-only scheme, and the present invention's composite evidence anchor scheme, with the results manually reviewed as a comparison benchmark.

[0044] In the plain text block RAG scheme, the field-level traceability accuracy is 78.4%, the numerical fact error detection rate is 72.1%, and the average manual review time is 42 seconds per fact. In the page number and coordinate reference-only scheme, the field-level traceability accuracy is 84.6%, the numerical fact error detection rate is 81.3%, and the average manual review time is 28 seconds per fact. After adopting the composite evidence anchor scheme of this invention, the atomic fact-level evidence location accuracy is 94.8%, the numerical fact error detection rate is 96.2%, the structural drift and version change detection rate is 92.5%, and the average manual review time is reduced to 11 seconds per fact. The above values ​​are typical test values ​​in the simulation embodiment, used to illustrate the improvement of the trusted document processing effect of composite evidence anchor, multi-digest hashing, and failure repair state machine.

[0045] In this simulation test, the ordinary text block RAG scheme could not distinguish between model illusions and source document parsing errors; the page number and coordinate reference scheme alone could improve the efficiency of manual location, but it could not accurately determine whether the evidence was invalid when the document was rearranged or the OCR results changed; this invention uses H text H struct H geo H imgand M page The combination of these features allows for the classification of errors into content modifications, layout reflows, structural shifts, OCR errors, or version changes, and automatically selects the corresponding repair actions. Therefore, it offers higher stability and auditability in batch processing of complex mixed-format documents.

[0046] In practical implementation, the method of the present invention can be flexibly configured and optimized in the following aspects: Regarding document parsing models, in addition to using general OCR models and general layout analysis models, specialized parsing models trained for contracts, invoices, test reports, scientific research papers, or tender documents can also be used; for native Word documents, the OpenXML structure can also be read directly to reduce OCR errors.

[0047] Regarding the representation of evidence anchor coordinates, in addition to using rectangular coordinate frames, polygon coordinates, text line sequence coordinates, table row and column coordinates, or native document XML paths can also be used. When cross-page positioning is required, multiple coordinate areas can be combined into evidence anchor groups, and a unified state can be set for the evidence anchor groups.

[0048] Regarding multi-digest hashing algorithms, in addition to SHA-256, SM3, BLAKE3, locality-sensitive hashing, or perceptual hashing can also be used; text hashing, structural hashing, geometric hashing, image perceptual hashing, and page Merkle root can be calculated separately and used in combination to distinguish between content modifications, layout reflows, structural shifts, and OCR changes.

[0049] Regarding retrieval weights, α, β, γ, δ, and η can be dynamically adjusted based on the task type. For example, field extraction tasks can increase the weight of object type matching, clause question-and-answer tasks can increase the weight of structural path matching, and scanned document review tasks can increase the weight of page proximity matching and historical credibility.

[0050] Regarding the consistency threshold τ, different values ​​can be set according to the application risk level. For example, it can be set to 0.75 for ordinary summary tasks, 0.85 for contract review tasks, and 0.90 or higher for financial audit or medical report review tasks; the system can also set different thresholds for monetary amounts, dates, and conclusion-related facts.

[0051] Regarding large language models, in addition to using general cloud-based large models, locally deployed models, industry-fine-tuned models, or multi-model integration solutions can also be used. Regardless of the model used, the evidence anchor generation, multi-dimensional consistency verification, and failure repair modules can all run as independent deterministic processing links.

[0052] Regarding output formats, in addition to natural language responses, output can also include JSON, tables, audit reports, highlighted PDF previews, office software annotations, or review tasks from enterprise process systems; each output format can carry evidence anchors and consistency scores.

[0053] Regarding the evidence anchor state machine, in addition to using states such as unverified, located, semantically passed, numerically passed, structural drift, coordinate failure, text hash failure, image hash failure, suspected OCR error, pending manual review, repaired, and abandoned, states such as "pending compliance review," "pending financial review," and "pending legal review" can also be added according to industry scenarios. However, each state should be triggered by the failure vector, verification result, or manual review result, rather than being directly determined by the self-evaluation of the large language model.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the present invention.

Claims

1. A trusted extraction and verification method based on document object composite evidence anchors, characterized in that, Includes the following steps: S1, parse the document to be processed into a collection of document objects O={o1,o2,...,o...} n }, each document object o i At least include object text c i Object type t i Page p i coordinate frame b i , structural path x i and reading order r i ; S2, for each document object o i Generate composite evidence anchor A i Composite evidence anchor A i It includes at least document identifier, version identifier, page number, rectangular coordinate frame, polygon coordinate, document object type, logical structure path, native object path, text offset, multi-digest hash, page object Merkle root, and evidence anchor state. The multi-digest hash includes text content hash, structure path hash, geometric location hash, and object image region-aware hash, and each digest is calculated independently. S3, retrieve candidate document objects from the document object set O according to the user task, and assemble the candidate document objects and their composite evidence anchors into a minimum sufficient context package; S4, the large model generates structured results based on the minimum sufficient context package, and then decomposes the structured results into a set of atomic facts F={f1,f2,...,f m }, and for each atomic fact f j Bind one or more candidate composite evidence anchors; S5, based on candidate composite evidence anchor A i Relocate the original object o i Read the current object text, coordinates, structural path, native object path, object type, and digest hashes of the original object, and use a deterministic program to process the atomic facts f. j Perform a joint check of semantic consistency, numerical consistency, object type consistency, structural path consistency, coordinate validity, and hash integrity to obtain the overall consistency score C(f). j A i and failure vector E j = <e sem ,e num ,e type ,e struct ,e geo ,e hash >The self-assessed confidence level of the large model output is not used as the basis for passing the verification; S6, when the overall consistency score C(f) j A i When the value is below the confidence threshold τ or a severe failure dimension exists, the failure vector E is used as the basis for determining the failure level. j Perform neighborhood expansion retrieval, object-level re-retrieval, structure relocation, OCR region re-identification or manual review, and update the evidence anchor status of the corresponding composite evidence anchor; S7 outputs the location information, anchor status, overall consistency score, and audit log of the atomic facts that pass the verification and their composite anchor evidence. It marks the status of atomic facts that fail the verification but does not output them as credible conclusions.

2. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The parsing in S1 specifically involves: performing page-level preprocessing on the document to be processed, including page rotation correction, skew correction, resolution normalization, noise removal, and page number recognition; reading the original structure of parsable electronic documents, which includes paragraphs, runs, tables, comments, footnotes, and style tags in OpenXML; and calling the OCR recognition model and layout analysis model on scanned images or image-type documents to obtain text lines, paragraph blocks, table areas, image areas, formula areas, and header / footer areas; the object type t i Includes headings, body paragraphs, list items, tables, table rows, table cells, images, figure captions, formulas, variable explanations, footnotes, comments, headers, footers, and attachment references; the coordinate frame b i Rectangular coordinates (x) min ,y min ,x max ,y max For irregular regions, use polygon coordinates q. i Indicates; the structural path x i The hierarchical path of the object in the document tree.

3. The trusted extraction and verification method based on document object composite evidence anchors according to claim 2, characterized in that, The parsing of table objects, formula objects, and chart objects in S1 specifically involves: identifying the header level, merged cells, row and column coordinates, page continuation relationships, and unit fields for table objects, and generating an independent document object for each table cell; For formula objects, identify the formula expression, variable name, variable explanation paragraph, and reference text near the formula; for chart objects, identify the chart title, caption, axis name, legend, and data series or trend description.

4. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The composite evidence anchor A in S2 i Its multi-digest hash is specifically as follows: Where D represents the document identifier, V represents the document version identifier, and P represents the document version identifier. i Indicates the page number where the object is located, B i Represents a rectangular coordinate frame, Q i T represents the coordinates of a polygon or quadrilateral. i Indicates the document object type, X i R represents the logical structure path. i Indicates the path to the native object, U i S represents the text offset of an object's text within its parent object or native document structure. i Indicates the anchor state of evidence; text content hash H text =Hash(N(c i )), where N(c i This represents the text after performing Unicode normalization, whitespace normalization, and punctuation normalization on the object text; the structure path hash H struct =Hash(T i ||X i ||R i ); Geometric location hash H geo =Hash(P i ||B i ||Q i Image region-aware hashing H img =pHash(Image(P i B i Page object Merkle root M page H of multiple objects on the same page text H struct and H geo It is calculated from the bottom up.

5. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The evidence anchor status in S2 specifically includes: unverified, located, semantically passed, numerically passed, structurally drifted, coordinate invalid, text hash invalid, image hash invalid, suspected OCR error, awaiting manual review, repaired, and discarded; The evidence anchor state is updated after each retrieval, generation, verification, and repair, and the state transition time, triggering task, failure vector, and repair action are recorded; multiple composite evidence anchors are allowed to form an evidence anchor group G. k ={A a A b In a table spanning multiple pages, multiple cells, formulas and their variable explanation paragraphs, charts and their captions each form an evidence anchor group, and the evidence anchor groups are set to a uniform state.

6. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The retrieval in S3 specifically involves calculating the comprehensive retrieval score R(q,o) of the document object according to the following formula. i And select candidate document objects based on the comprehensive retrieval score: ， Among them, S sem Indicates the semantic similarity between a user's question and the object's text or object summary; S struct This indicates the degree of match between the chapters, clauses, tables, or attachments in a user question and the object structure path; S type Indicates the degree of matching between the user task type and the object type; S geo S represents the page proximity between the target object and the matched object; hist This represents the historical credibility of the object being successfully verified in similar tasks; α, β, γ, δ, and η are non-negative weights, and α+β+γ+δ+η=1.

7. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The minimum sufficient context package in S3 specifically includes candidate object text, object type, page number, structural path, adjacent objects, table row and column relationships, formula variable explanations, and figure annotations, with the total number of tokens not exceeding a preset upper limit; For tasks requiring cross-object understanding, the neighborhood objects are expanded based on containment edges, reading order edges, cross-page continuation edges, image and text reference edges, and formula variable edges to form a local evidence subgraph.

8. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The generation of atomic facts and the binding of evidence anchors in S4 specifically involve: requiring the large model to output atomic facts through constraint hints, structured output patterns, or function call interfaces, with each atomic fact f... j The corresponding structured record includes at least the fact text, fact type, normalized value, unit, confidence level, candidate evidence anchor list, and reason for generation. Binding rules are executed according to fact type: monetary facts are bound to table cells or paragraph evidence anchors containing currency, amount field names, or price list positions; date facts are bound to paragraphs, table cells, or annotation evidence anchors containing date text and contextual roles; subject facts are bound to contract headers, signature areas, business name fields, or attachment subject descriptions; clause facts are bound to clause numbers, title paths, and adjacent paragraph evidence anchors; conclusion facts are bound to evidence anchor groups consisting of data cells, formula objects, and conclusion paragraphs; time-limited facts are bound to paragraphs or table cells containing time-limited values, time unit fields, or performance period clauses; and ratio facts are bound to paragraphs, table cells, or formula objects containing percentage, rate, ratio fields, or calculation formula positions. When multiple candidate evidence anchors conflict, support is calculated based on anchor status, object type matching degree, version identifier, hash integrity, and business priority, and low-support evidence anchors are marked as conflict candidates.

9. The trusted extraction and verification method based on document object composite evidence anchors according to claim 1, characterized in that, The joint verification in S5 specifically involves calculating the overall consistency score C(f) according to the following formula. j A i ): ， Among them, C sem This indicates the semantic consistency between the factual text and the original text; C num Indicates the degree of matching after normalization of numerical values, units, dates, proportions, or formula calculation results; C type Indicates the degree of matching between the fact type and the document object type; C struct Indicates the degree of consistency between the structure path and the native object path; C geo Indicates whether the anchor coordinates of the evidence are locatable on the current page and fall within the corresponding object area; C hash H represents text H struct H geo H img and M page Combinatorial integrity; λ1 to λ6 are weight parameters, and λ1+λ2+λ3+λ4+λ5+λ6=1; when C(f j A i If the atomic fact is greater than or equal to the confidence threshold τ and there is no serious failure dimension, it is determined that the atomic fact is supported by the evidence anchor; otherwise, it proceeds to the failure cause determination. The failure vector E j Each component takes the value of 0, 1, or 2, where 0 indicates pass, 1 indicates slight abnormality, and 2 indicates severe abnormality.

10. The trusted extraction and verification method based on document object composite evidence anchors according to claim 9, characterized in that, The failure cause determination in S5 and the failure repair in S6 are specifically as follows: If H text Changes and H geo If unchanged, it is judged as suspected content modification; if H geo Changes and H text If consistent, it is judged as a suspected layout rearrangement; if H img Change and H text Changes are identified as suspected OCR or changes in scan quality; if M page Changes are identified as page-level version changes; if the object pointed to by the structural path does not exist, it is identified as structural drift; accordingly, when semantic consistency is insufficient but coordinates, structure, and hash are valid, a neighborhood expansion search is performed on the same chapter, the same table row and column, adjacent figure captions, or formula variable interpretation objects; when numerical consistency is insufficient, the original text object is reread and numerical normalization, unit conversion, table cell positioning correction, or formula recalculation is performed; when object types do not match, paragraph routing is switched to table cell routing, formula routing, or annotation routing; when structural drift occurs, the original object path, heading level, adjacent object hash, and reading order are used for repositioning. When coordinates become invalid or hashes are inconsistent, a message is displayed indicating a document version change, content modification, or reformatting, and the relevant evidence anchors are regenerated or discarded. If an OCR error or layout parsing error is determined, the system returns to the original page image area based on the coordinate frame in the evidence anchor, and performs cropping, enlargement, binarization, noise reduction, rotation correction, or multi-OCR engine re-recognition on that area. When the re-recognition result is consistent with the context structure, the document object text and evidence anchor hash are updated; otherwise, a manual review task is output. An audit log is generated for each failed repair process, including the user task, candidate evidence anchors, failure score, failure reason, repair action, object text before and after repair, evidence anchor after repair, and final credibility status.

Images