An artificial intelligence-based bidding and mining document intelligent data processing method

Abstract

The application belongs to the technical field of intelligent document processing, and discloses a kind of intelligent data processing method of bidding and procurement file based on artificial intelligence, including obtaining multi-source multi-version file to carry out layout rectification and OCR, obtain page fragment object set;Clause record is extracted according to preset schema by using controlled decoding language model, and clause record set is formed;The same name or synonymous clause is aligned across versions, and a time sequence knowledge graph is constructed;Cross-document consistency rules are compiled into executable rule set that can be evaluated on the time sequence knowledge graph using executable bidding and procurement constraint language;The scope of each rule is evaluated to generate conflict candidates;Calculate the minimum hit set, get the minimum conflict explanation set and locate the root cause clause instance, generate the minimum repair suggestion according to the normalized results of the violated rules' operators and nodes.The application realizes the automatic determination of clause range uniformity and the latest effective state, accurate positioning of conflicts and minimum cost repair.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent document processing, and particularly relates to an intelligent data processing method for bidding and procurement files based on artificial intelligence. BACKGROUND

[0002] In a business scenario such as bidding or procurement projects, a large amount of heterogeneous data such as bidding announcements, bidding files, procurement requirement books, clarification files, answer files, supplementary files, correction announcements, tender files and attachments, evaluation reports, bidding notices, contract texts and contract change files, and acceptance reports will be generated. Such files have various sources and frequent version updates. The current industry generally processes the files by using layout rectification and optical character recognition (OCR) technology to obtain text information, and then uses keyword retrieval, regular templates, dictionary matching or simple classifiers to extract elements, and then uses manual correction, deduplication and similarity comparison to complete archiving. Cross-document consistency is usually checked item by item according to a table list, and conflict positioning depends on experience to search for prompt words such as “correction” and “supplement”. The units, currencies and dates in the clauses are manually converted and recorded. The final result is presented in a static report or an explanatory document, and there is a lack of structured expression in the time dimension and reusable constraint expression methods.

[0003] The main shortcomings of the prior art are as follows: it is difficult to cross-version align multiple-source and multiple-version files and reliably determine the “latest effective” state; cross-document consistency is usually manually checked after the fact, and there is a lack of rules that can be evaluated in the time dimension; the units, currencies and dates in the clauses of the files are prone to implicit inconsistencies; when there is an information conflict across documents, the prior art can only present the conflict, cannot trace the root cause of the conflict, and cannot provide an executable repair path. At the same time, the evidence is scattered in different page numbers and layout coordinates of multiple-source and multiple-version files, and there is a lack of unified evidence pointers and chain associations, which makes it difficult to audit and locate the source fragments in time. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an intelligent data processing method for bidding and procurement files based on artificial intelligence to solve the problems in the prior art that multiple-source and multiple-version bidding and procurement files are difficult to cross-version align and determine the latest effective state, and cross-document consistency cannot be automatically evaluated in a unified scale and time dimension.

[0006] To solve the above technical problems, the present application provides the following technical solutions.

[0007] The present application provides an intelligent data processing method for bidding and procurement files based on artificial intelligence, comprising the following steps:

[0008] S1, acquire multi-source and multi-version files pointing to the same procurement project, perform layout rectification and OCR on the multi-source and multi-version files to obtain a page segment object set, and establish a version sequence;

[0009] S2, based on the page segment object set, extract clause records according to a preset Schema using a controlled decoding language model to form a clause record set;

[0010] S3, based on the clause record set and the version sequence, align the same name or synonymous clauses across versions to construct a time sequence knowledge graph;

[0011] S4, compile cross-document consistency rules into an executable rule set that can be evaluated on the time sequence knowledge graph using an executable procurement constraint language;

[0012] S5, based on the executable rule set, evaluate the scope of each rule on the time sequence knowledge graph and generate conflict candidates, and after aggregation, obtain a conflict candidate set;

[0013] S6, based on the conflict candidate set, calculate the minimum hit set that makes all conflict candidates satisfy, obtain the minimum conflict explanation set and locate the root cause clause instance, and generate the minimum repair suggestion according to the normalized results of the violated rules and nodes;

[0014] S7, output an auditable verification result report according to the root cause clause instance and the minimum repair suggestion.

[0015] Further, in S1, the multi-source and multi-version files include tender announcements, tender documents or procurement requirement documents, clarification documents or answer documents and supplementary documents, correction announcements, tender documents and tender document attachments, bid evaluation reports, bid-winning notices, contract texts and contract change documents, acceptance reports, prequalification documents, bid opening records, technical specification sheets, and drawings and price lists.

[0016] Further, performing layout rectification and OCR on the multi-source and multi-version files to obtain a page segment object set and establish a version sequence, includes:

[0017] Performing layout rectification on multi-source and multi-version files to form geometrically normalized page images and retaining file identifiers and version timestamps;

[0018] Performing OCR on the geometrically normalized page images to obtain recognition results containing text content and layout coordinates and forming page segment candidates;

[0019] Based on the page segment candidates, file identifiers, version timestamps, page numbers, and layout coordinates, constructing a page segment object set;

[0020] Based on the version timestamps, establishing a version sequence for the multi-source and multi-version files and recording the coverage relationship.

[0021] Further, S2 is specifically:

[0022] Segmenting the content based on the page fragment object set to obtain a text segment set consistent with the preset Schema field mapping;

[0023] Constructing a controlled decoding language model prompt based on the preset Schema and limiting the output field and output format, combining the text segment set with the prompt to form a controlled decoding input;

[0024] Generating a clause record from the controlled decoding language model based on the controlled decoding input, and attaching an evidence pointer pointing to the page fragment object set;

[0025] Normalizing the clause record based on the information of units, currencies and dates and summarizing to form a clause record set.

[0026] Further, S3 is specifically:

[0027] Determine the cross-version alignment relationship of the same name or synonymous clauses based on the clause record set and the version sequence;

[0028] Generate a clause instance node for each version according to the cross-version alignment relationship;

[0029] Mark the covering edge, revision edge, reference edge and abolition edge between the clause instance nodes according to the version sequence;

[0030] Generate a coverage chain and a reference chain along the coverage edge and the reference edge according to the version sequence order, and select the clause instance of the version sequence later and not revised or abolished as the latest valid node set for each clause identification in combination with the marking of the revision edge and the abolition edge;

[0031] Merge the clause instance node, covering edge, revision edge, reference edge, abolition edge, coverage chain, reference chain and latest valid node set to form a time sequence knowledge graph.

[0032] Further, S4 is specifically:

[0033] Parse the cross-document consistency rule source according to the syntax of the executable procurement constraint language using the controlled decoding language model and generate a rule candidate set;

[0034] Align the fields and unify the terms of the rule candidate set, the clause record set and the latest valid node set in the time sequence knowledge graph to form a cross-document consistency rule;

[0035] Determine the scope of each cross-document consistency rule in the time sequence knowledge graph and point to the latest valid node set;

[0036] Reading the normalized result in the clause record set corresponding to the node in the scope, configuring the unit, currency and date tolerance strategy and completing the compilation to obtain an executable rule set.

[0037] Further, the executable rule set comprises a MUST_EQUAL operator, an IN operator, a RANGE operator, a MUTEX operator, an IMPLIES operator and an EXCEPT_IF operator.

[0038] Further, S5 is specifically:

[0039] Based on the executable rule set, the scope of each rule in the time sequence knowledge graph is selected, and each rule is evaluated to obtain a consistency state of each rule;

[0040] Based on the rule whose consistency state is not satisfied, a conflict candidate is generated;

[0041] All conflict candidates are converged to obtain a conflict candidate set.

[0042] Further, S6 is specifically:

[0043] Based on the conflict candidate set, a corresponding relationship between the conflict and the clause instance node is established, and the minimum hit set that makes all conflict candidates resume to be satisfied is calculated;

[0044] Based on the minimum hit set, the minimum conflict explanation set is determined, and the root cause clause instance is located on the time sequence knowledge graph;

[0045] Based on the operator of the violated rule and the normalized result of the clause instance node, a minimum repair suggestion is generated.

[0046] Further, S7 is specifically:

[0047] Based on the root cause clause instance and the minimum repair suggestion, an evidence pointer in the evidence page fragment object set is converged and an evidence chain is assembled;

[0048] Based on the evidence chain and the minimum repair suggestion, an auditable verification result report is generated.

[0049] The method has the advantages that the method compiles cross-document consistency rules into an executable rule set that can be evaluated on a time sequence knowledge graph by using an executable procurement constraint language, realizes accurate scope evaluation of rules on the latest valid node set, and is used for automatically determining compliance and exceptions; the minimum hit set is calculated based on the conflict candidate set, the minimum conflict explanation set is obtained and the root cause clause instance is located, the minimum repair suggestion is generated, which is used for guiding value alignment and compliance unification, and finally the conflict is eliminated in a closed loop with the minimum change and the traceable verification result is output. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0051] Fig. 1 Flowchart of the intelligent data processing method for the artificial intelligence-based bidding and procurement document;

[0052] Fig. 2 Flowchart of the processing of page rectification and OCR to the page segment object set and version sequence;

[0053] Fig. 3 Flowchart of the controlled decoding and extracting clause records;

[0054] Fig. 4 Flowchart of the construction of the time sequence knowledge graph. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0056] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in other ways different from those described herein without departing from the scope of the present application, and those skilled in the art can make similar extensions without departing from the scope of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0057] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0058] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides an intelligent data processing method for artificial intelligence-based bidding and procurement documents, comprising the following steps:

[0059] S1, acquiring multi-source multi-version documents pointing to the same bidding project, rectifying the layout of the multi-source multi-version documents and performing OCR to obtain a page segment object set, and establishing a version sequence.

[0060] S1.1, from the designated bidding announcement catalog to grab the bidding announcement and extract the project name, bidding number, purchaser name, administrative division code and announcement date, generate the main anchor document; With the project name, bidding number, purchaser name, administrative division code and announcement date of the main anchor document as the retrieval condition, all the files related to the bidding and procurement in the government procurement and public resource transaction public page, mail inbox archiving and internal sharing directory within the time window are batched and the source path is recorded; Calculate the content fingerprint and layout fingerprint of the batched files to remove duplicate files and retain the source path, capture time and abstract text of each file to form a candidate file set; Extract the project name, bidding number, purchaser name, administrative division code and file date from the candidate file set one by one and perform the same name standardization processing to form the alignable field pair with the corresponding fields of the main anchor document.

[0061] S1.2, based on the candidate file set and the main anchor document, calculate the comprehensive ranking, select the optimal time sequence chain with the minimum path cost according to the comprehensive ranking and date distance ratio cumulative value, and arrange in ascending order of file date to generate a multi-source multi-version file list pointing to the same bidding and procurement project.

[0062] Further, by cleaning the project name in the candidate file set and the main anchor document (removing parentheses and extra spaces), and calculating the number of steps of Levenshtein edit distance between the two strings (allowing replacement, insertion, deletion and counting the minimum number of steps); Take the larger value of the length of the two strings as the standardization base, divide the number of steps of the edit distance by the base to get a proportion, and take "one minus the proportion" as the project name similarity; Calculate the purchaser name similarity using the same calculation method;

[0063] After removing punctuation and spaces from the string, compare them for equality, calculate the bidding number consistency, and mark 1 for the same and 0 for the different; Map the administrative division codes in the candidate file set and the main anchor document to equal-length bit strings and compare them bit by bit to get the difference ratio and convert it to a similarity score accordingly, the closer the similarity score is to the maximum value, the closer it is, and the closer the administrative division is; Generate a fixed-length overlapping substring set from the denoised text of the main anchor document and the denoised text of the candidate file, calculate the intersection ratio of the two sets to get the full-text content overlap, the closer the overlap score is to the maximum value, the more similar the content is; Convert the candidate file date and the main anchor document announcement date to the same time scale in days and calculate their difference, then standardize the difference according to the fixed time scale to get the date distance ratio, the smaller the date distance ratio, the more adjacent the time is;

[0064] The candidate files are respectively ranked on six indexes and summarized as comprehensive rankings, and the earlier the comprehensive ranking is, the closer the overall is to the main anchor document; if the rankings are tied, the tie is broken in the order of higher source path authority first, containing more coverage prompt words first, and earlier crawling time first; a directed acyclic relationship is established for the candidate files in chronological order, and the main anchor document is taken as the starting point to enumerate the date increasing path, the path cost is calculated for each path, the path cost is composed of the cumulative comprehensive ranking of the files in the path and the cumulative date distance ratio between adjacent files, and the adjacent files are required to have non-zero full-text content overlap, and if not, the suboptimal adjacent file is used; the path with the minimum path cost is selected as the optimal time sequence chain from all feasible paths, and when the optimal time sequence chains are tied, the path containing the file with higher source path authority is given priority, and then the path containing more concentrated coverage prompt words is given priority, and when the tie still exists, the path with later end date is selected; the files in the optimal time sequence chain are output in ascending order of file date as a multi-source multi-version file list pointing to the same bidding project.

[0065] S1.3, deskewing the multi-source multi-version file to form a geometrically normalized page image;

[0066] Specifically, the file identifier and version timestamp of the page image to be processed in the multi-source multi-version file are read and temporarily stored as metadata; after the page image is converted into a grayscale image, denoising processing and contrast-limited histogram equalization are performed, and a binary image is obtained through adaptive threshold segmentation; edge detection is performed on the binary image, and small broken edges are closed and thinned to obtain an edge set with good continuity; probability Hough straight line detection is performed on the edge set, and short straight lines and repeated collinear straight lines are filtered out, and then adjacent straight lines with similar angles are merged to obtain a straight line set representing the page boundary; a plurality of groups of corner points are generated by intersecting each two straight lines in the straight line set, and candidate quadrilaterals are screened out according to the area and aspect ratio of the closed polygon surrounded by the corner points, and a quadrilateral covering the largest text area is selected preferentially and the coordinates of its four vertices are recorded, and if no valid quadrilateral is detected, the largest rectangle outside the connected domain is extracted as a substitute boundary;

[0067] A one-to-one correspondence is established between the boundary vertex and the four corners of the target rectangle, a homographic mapping is estimated, and perspective correction is performed on the page image to align the page boundary to a rectangular shape, while cutting off the excess black edges and keeping the aspect ratio unchanged; the main direction of the text line is calculated on the perspective-corrected image, which can be found by locating the peak value direction of the projection histogram in the horizontal direction and the vertical direction, or by locating the text line direction angle through the frequency energy peak value, and then the page image is rotated to make the text line horizontal after the residual rotation angle is obtained;

[0068] The rotated corrected image is subjected to size and resolution unification processing, including scaling the image to a unified target size, unified resolution, unified page margin and content centering, and then performing black edge secondary cutting to eliminate the blank area generated by rotation, to obtain a unified image. Geometric quality verification is performed on the unified image, and the verification includes indicators such as four-corner right-angle, page boundary closure and text line direction. If any indicator does not meet the preset passing condition, the process reverts to the quadrilateral candidate generation step, and the next quadrilateral candidate in the order of the current candidate is selected according to the descending order of the candidate area (and the descending order of the perimeter to break the tie), and the perspective correction and rotation correction are re-executed until it passes or it is determined that the page is completed with the whole page rectangle as the boundary for correction;

[0069] The file identification and version timestamp are written into the metadata field of the accompanying record and the image header, and the parameter summary and quality verification result summary of this perspective correction and rotation correction are recorded, to generate a geometric normalized page image.

[0070] S1.4, performing OCR on the geometric normalized page image to obtain recognition results containing text content and layout coordinates and form page segment candidates;

[0071] The geometric normalized page image is subjected to grayscale, denoising, contrast enhancement and adaptive binarization to obtain a binary image suitable for text region detection. Extreme value stable region detection is performed on the binary image to generate a set of connected text candidate regions, and the bounding box of each candidate region is extracted;

[0072] The connected text candidate regions are subjected to overlap merging and near neighbor merging, the main direction angle of each candidate region is estimated based on the long side direction of the region, and the area of the extremely small or elongated abnormal noise region is filtered. The candidate regions are density clustered according to the criteria of similar main direction angle and vertical proximity to obtain a set of text lines, and the minimum enclosing quadrilateral of each text line is fitted as the text line boundary. The vertical direction projection is calculated inside each text line, the adjacent projection valley values are taken as the segmentation points to obtain word sub-blocks, and the recognition segment image is cut out in the geometric normalized page image coordinate system. The recognition segment image is sent to the OCR (Optical Character Recognition) engine, which uses connection time sequence classification decoding to output character sequences, character-by-character confidence sequences and character-by-character boundary coordinates. The character-by-character boundary coordinates are mapped back to the layout coordinate system of the geometric normalized page image, adjacent characters in the same text line are concatenated into line-level text in reading order, and initial segment entries containing line-level text and layout coordinates are generated. For each initial segment entry, multiple indicators such as direction consistency, recognition confidence aggregation, geometric compactness, language readability and stroke stability are calculated, and each indicator is judged with a fixed passing condition. The initial segment entries that meet all the passing conditions enter the subsequent merging process, and the initial segment entries that do not meet the passing conditions are directly rejected;

[0073] The initial fragment entries are sorted in page reading order, and adjacent and small-interval initial fragment entries in the same text line are merged to generate page fragment candidates containing text content, layout coordinates and confidence, and the page fragment candidates are supplemented with file identifier, version timestamp and page number.

[0074] S1.5, based on the page fragment candidate, the file identifier, the version timestamp, the page number and the layout coordinates, constructing a page fragment object set;

[0075] The page fragment candidates and corresponding file identifier, version timestamp and page number are grouped by page number and enter the same page processing; the page fragment candidates in the same page are uniformly encoded, uniformly formatted and uniformly converted to the coordinate system of the geometric normalized page image; the adjacent candidate pairs are indexed by the upper left corner coordinates in the same page to provide a search entry for subsequent merging and deduplication; the adjacent candidate pairs are judged for text deduplication, and when the text content is completely consistent and the layout coordinates are highly overlapped, only the candidate with the newer source time is retained to form a deduplicated candidate set;

[0076] The deduplicated candidate set is sorted in reading order according to the criteria of consistent text direction and close baseline to obtain an in-line sequence; the in-line sequence is merged in-line according to the adjacent interval and alignment relationship to generate a line-level fragment, and the text content and layout coordinates are updated after merging; the line-level fragments are merged into paragraphs according to the line spacing, starting position alignment and indentation relationship of adjacent lines to generate a paragraph-level fragment, and the paragraph bounding box is updated after merging;

[0077] The row and column alignment relationship in the table region is recognized, and the line-level fragments with boundary alignment and continuous content in the same row or column are merged into cells to generate a table cell-level fragment, and the row and column positioning information of the cells is recorded after merging;

[0078] The file identifier, version timestamp and page number are written for each page fragment object, and the fragment text, layout coordinates bounding box, reading order position of the text in the page and merging source record are also written;

[0079] Consistency check is performed on the page fragment objects, including whether the layout coordinates fall within the page range, whether the bounding box is closed, whether the text is empty and whether the merging source is traceable, and the objects that do not pass the check are rejected or rolled back to the previous merging stage for regeneration;

[0080] The page fragment objects that pass the consistency check are sorted by page number and reading order to form a page fragment object list for the page. The page fragment object lists of all pages are summarized to generate a page fragment object set.

[0081] S1.6, establish version sequence and record coverage relationship for multi-source multi-version files based on version timestamp;

[0082] Parse the version timestamp into a unified time scale and correct common format differences or missing fields. Unresolvable items are first marked for review and temporarily placed at the end of the sequence. Sort the initial time sequence from early to late according to the version timestamp. When multiple files appear at the same time, disperse them according to the fixed authority order of the source path. Determine the file category according to the fixed keyword set of the title first line, source column, and first paragraph of the text, and record the remaining judgment clues in the object for future use.

[0083] In the time sequence, traverse the files one by one, extract the standardized fields of project name, tender number, purchaser name, administrative division code, and file date, and first establish a reference index with earlier files in time. Connect objects with consistent tender numbers and high project name similarity, followed by objects with high project name similarity and high purchaser name similarity. If neither condition is met, only establish a time adjacency.

[0084] Scan the text to identify coverage prompt words, and determine the coverage type and covered object based on the file category and reference index: correct and modify as change coverage, supplement as supplementary coverage, abolish as invalid coverage, replace as replacement coverage, and continue to be effective as continuation coverage. When the coverage object can be located to the clause number or paragraph title, record the coverage range description as clause level, and when it cannot be located, record the coverage range description as full-text level. Extract the page number and location of the sentence where the coverage prompt word is located as evidence excerpt location.

[0085] Establish a version node for each file and write in the file identifier, version timestamp, source path, file category, and in-edge and out-edge count of coverage relationship. String the version nodes into a version sequence in chronological order and record the previous and subsequent identifiers in the node.

[0086] Perform consistency check on the version sequence. When a coverage loop is found, keep the earlier time point and remove the recent reference. When multiple coverage conflicts of the same type are found, keep the one with higher source path authority and record the conflict summary. When a continuation coverage after being abolished is found, use the invalid coverage as a reference and remove the subsequent continuation. Output the version sequence sorted by time and coverage relationship records, and keep the file identifier and version timestamp in each version node.

[0087] S2, based on the page fragment object set, use a controlled decoding language model to extract clause records according to the pre-set Schema, forming a clause record set.

[0088] S2.1, based on the page fragment object set, perform content segmentation to obtain a text segment set consistent with the pre-set Schema field mapping;

[0089] Specifically, text normalization is performed on the page fragment object set, and the normalization content includes unified character encoding, unified full-width and half-width, unified punctuation form, merging of redundant spaces, and unified line break expression.

[0090] A field anchor point word list is generated from the preset Schema, which includes preset Schema field names and enumerable aliases, common title writing methods, and colon writing methods. A deterministic matching order and priority are generated for each word list entry.

[0091] Field anchor point matching is performed on each page fragment object in the page fragment object set. When a match is found, the anchor point position, the page fragment object identifier occupied by the anchor point, and the anchor point level type are recorded. The level type is determined based on whether it is independent, whether it contains a colon, and whether it is a table column header.

[0092] Starting from the anchor point position, the page fragment object set is scanned from front to back. When a new anchor point of the same level type is encountered or the end of the page is encountered, the scanning stops. The continuous page fragment objects between the start and stop positions are concatenated in reading order to form a candidate segmented text, and the evidence coordinate sequence of the candidate segmentation is generated.

[0093] Column header recognition is performed on page fragment objects belonging to table regions. Column header recognition is based on the one-to-one correspondence between the phrases in the first row or the significantly bolded row and the field anchor point word list. After the column header is mapped to the field in the preset Schema, the candidate segmented text is generated by merging the cell texts in the same column by row, and the row and column positioning of the candidate segmentation and the evidence coordinate sequence are recorded.

[0094] Boundary clipping is performed on the candidate segmented text. The clipping rules are to remove duplicate anchor point phrases at the start and end positions, unify the in-line connectors and meaningless separators, and retain the structural markers within the segment (such as clause numbers and subheadings). After clipping, the segmented text and the corresponding field mapping draft are obtained.

[0095] When multiple candidate segmented texts appear in the same field, conflict resolution is performed. Conflict resolution is performed in the following deterministic order: selecting candidate segmented texts containing override prompt words such as "corrected as", "modified as", and "as such"; if there are no override prompt words, selecting the candidate segmented text closest to the field anchor point; if there are still parallel conditions, selecting the position on the later page or in the later reading order.

[0096] Each final determined segmented text is written into the output along with the corresponding field, file identifier, version timestamp, page number, evidence coordinate sequence, segmentation identifier, and reading order, forming a text segment set.

[0097] S2.2, based on the preset Schema, construct a language model prompt for controlled decoding and limit the output field and output format, combine the text segment set with the prompt to form a controlled decoding input.

[0098] Generate a field list from the preset Schema, which lists the field name, value type, whether it is mandatory, the number of occurrences allowed, value constraint table or format expression, and whether it needs an evidence pointer, and keeps the field name consistent with the preset Schema;

[0099] Formulate output format constraints according to the field list, including the structure of the top-level clause record set, the fixed key name order of the clause record, the mandatory key name always appearing and allowing to be empty, the key name not appearing in the field list, the nesting level being limited to the level allowed by the field list, and the evidence pointer must contain the page number, layout coordinates, and source page fragment object identifier;

[0100] Generate a field value template for each type of value, keep only Arabic numerals and decimal points for numerical values and remove unit symbols, unify date types into hyphenated year-month-day format, split range types into start and end key names, limit enumeration types to table entries, separate currency types and values and place them in corresponding key names, and select units from a unified table and place them in the unit key name;

[0101] Prepare a prompt skeleton, which includes role constraints for extraction tools, task objectives for mapping a set of text segments into structured text, and step-by-step requirements for field mapping rules and field value templates, evidence rules require each field to be accompanied by a page number and layout coordinates, conflict resolution rules specify that when the same field appears in multiple places, the segment containing the coverage prompt word is preferred, otherwise the segment with the later page number and reading order is selected, and output of content not in the field list is prohibited;

[0102] Arrange the set of text segments in ascending order of page number and reading order, and add a prompt to the end of each segment, with a line of label explaining the field candidate information, page number and layout coordinates, and segment identifier before the text content, and when the text is too long, preferentially retain the hit field anchor point or the segment containing the coverage prompt word according to the fixed strategy;

[0103] Generate structure checking instructions, which require the language model under controlled decoding to internally check whether the mandatory key names are complete, whether the key names are all from the field list, and whether the values meet the field value template before generating structured text, and correct any discrepancies within a single generation and output the structured text, without allowing explanatory text;

[0104] Set controlled decoding constraints, use deterministic decoding, turn off random sampling, limit output to only include structured text and prohibit free text supplementation, and when the structure check fails, retry internally for a fixed number of times within a single call and use the last compliant output as the final output;

[0105] The context header information, field list, output format constraint, field value template, prompt skeleton, and the sorted text segment set are spliced in a fixed order to form a complete prompt text, forming a controlled decoding input.

[0106] It should also be noted that the language model of controlled decoding refers to a general autoregressive language model that is constrained by a pre-set Schema in the generation phase and only outputs structured results that meet the field list and output format constraints. It is derived from a general language model by fixing the pre-set Schema-related constraints in the prompt and setting controlled decoding configurations such as deterministic decoding, without additional training.

[0107] S2.3, generating a clause record from the controlled decoding input by the language model of controlled decoding, and appending an evidence pointer pointing to the page segment object set.

[0108] Receive the controlled decoding input and the text segment set index and the page segment object set index, send the controlled decoding input directly to the language model of controlled decoding, and set deterministic decoding and fixed output structure, only allow to generate structured text;

[0109] Get the structured text generated by the language model of controlled decoding at one time, parse the structured text into a clause record candidate set, and check whether the key name, mandatory item and level are compliant according to the pre-set Schema. If not compliant, format correction according to the field list and output format constraint in a single call, and then parse the clause record candidate set again;

[0110] Extract the clause identifier, constraint type, and value or range from the clause record candidate set, and read the segmentation label information in the controlled decoding input. If the clause record candidate contains a source segment identifier, locate the corresponding segment in the text segment set index with the segment identifier, and find the corresponding page segment object identifier, page number and layout coordinates in the page segment object set index to generate an evidence pointer and append it to the clause record candidate;

[0111] When the clause record candidate does not carry the source segment identifier, two-stage positioning is performed in the text segment set: first, perform exact substring matching to preferentially find a segment in the segmented text that is completely consistent with the clause record value or range; if not hit, perform normalized approximate matching, which includes normalizing case, unifying full-width and half-width, removing unnecessary white space, and unifying punctuation, then perform inclusion matching in the segmented text, and after hitting, find the page segment object identifier, page number and layout coordinates in the page segment object set according to the hit segment to generate an evidence pointer and append it to the clause record candidate;

[0112] When the same clause record is located in multiple segments, the evidence pointer is selected according to the deterministic priority, and the priority order is that the segment containing the coverage prompt word is prior to the segment not containing the coverage prompt word, the segment with the later page number is prior to the segment with the earlier page number, the segment with the later reading order is prior to the segment with the earlier reading order, and finally the first evidence pointer is reserved and the remaining evidence pointers are recorded as supplementary pointers;

[0113] The structural consistency of the clause record candidate to which the evidence pointer is attached is checked, and the checking content includes whether the key names are all from the preset Schema, whether the mandatory keys exist, whether the range type appears simultaneously with the start and end, whether the evidence pointer contains the file identifier, the version timestamp, the page number and the layout coordinates, and the page segment object identifier, and the clause record candidate that passes the checking is converted into a clause record.

[0114] S2.4, based on the unit, currency and date information, the clause record is normalized and summarized to form a clause record set.

[0115] The clause record candidate set, clause identifier, constraint type, value or range and evidence pointer are received, and the unit of measurement mapping table, currency mapping table, date expression mode set and exchange rate table are loaded; the original value is first normalized, the character encoding, full and half width, punctuation and blank are unified, and the Chinese numerals and quantifiers are converted into continuous numerals; the target unit is determined in the unit of measurement mapping table and the numerical conversion is completed, the start and end of the interval are processed respectively and the end point inclusion relationship is clarified, and when the unit is missing, the same field is referenced the last time and the source is recorded; the currency code is determined in the currency mapping table, and if necessary, it is converted to a unified currency according to the version timestamp from the exchange rate table and the original currency and the original amount are retained, and when the interval or step rule is contained, it is normalized and stored in order respectively; according to the date expression mode set, the complete date, year and month, year, relative date and period are converted to a unified format or absolute period and the resolution and anchor source are marked; after completion, the structural consistency check and conflict folding are performed, the field name, mandatory item, interval end point and evidence pointer integrity are checked, the non-passing ones are marked for rechecking and excluded from the summary, the passing ones are sorted according to the clause identifier and reading order, the records containing the coverage prompt word are preferentially reserved, followed by the selection according to the page number and reading order, and finally the clause record set containing the original string and the normalized result is output while keeping the evidence pointer unchanged.

[0116] S3, based on the clause record set and the version sequence, the same name or synonymous clause is aligned across versions, and a time sequence knowledge graph is constructed.

[0117] S3.1, based on the clause record set and the version sequence, the cross-version alignment relationship of the same name or synonymous clause is determined.

[0118] Further, generate field mapping list according to preset Schema field name, field anchor word list and synonym list, and perform synonym merging on clause identification. Group clause records in each version node according to field mapping list, forming intra-version clause grouping list. Compare adjacent nodes along the version sequence two by two, perform title phrase matching, paragraph number matching and evidence pointer page number proximity matching on the same name or synonymous field grouping, and record the matching score source. When multiple candidate corresponding occur, determine priority according to priority order, with clause records containing coverage cues first, page numbers later, and reading order later.

[0119] For the determined corresponding pair, establish cross-version alignment relationship, write prequel version pointer and subsequent version pointer, and attach evidence location summary of matching basis. Establish placeholder entry for missing fields, mark missing type as not appearing or being abolished, and maintain link continuity with adjacent aligned clause records. Perform consistency check on the full sequence, resolve one-to-many or many-to-one contradictions, retain the highest priority corresponding pair and generate conflict summary for audit, output cross-version alignment relationship list and alignment graph index.

[0120] S3.2, generate clause instance nodes for each version according to cross-version alignment relationship;

[0121] Specifically, according to the cross-version alignment relationship, bind the clause records of each field in each version to the corresponding version node and generate clause instance nodes, write clause identification, version timestamp, file identification, file category and reading order in the node; complete the normalized results into node attributes, including numerical values or intervals, units, currencies and amounts, dates or periods and resolutions, inclusion or exclusion semantics of interval endpoints; backfill source page number, layout coordinates, page segment object identification and evidence text summary in the page segment object set according to the evidence pointer; when multiple candidates occur in the same version, determine the main instance according to the fixed order of containing coverage cues first, page numbers later, and reading order later, and record the rest as backup instances in the node; when the version is missing or abolished, generate a placeholder instance and mark the missing type and invalid reason; calculate the unique key (composed of clause identification and version timestamp) for all instances and write the prequel and subsequent pointers to maintain link continuity, finally output the clause instance nodes arranged in version order.

[0122] S3.3, label coverage edges, revision edges, reference edges and abolition edges between clause instance nodes according to version sequence;

[0123] Specifically, according to the version sequence, the clause instance nodes of the same clause identifier are traversed from early to late, the coverage prompt words and coverage relationship records in the source file are read for each subsequent node, and the affected predecessor node is located: when "corrected as" or "modified as" appears, a revision edge is marked between the predecessor node and the current node; when "supplemented as follows", "for reference", or "replace" appears, a coverage edge is marked, and the replace semantics also mark an invalid edge for the replaced node; when the text explicitly references other clause numbers or file numbers, the target node is located according to the number or file number in the version sequence and a reference edge is marked; when "invalidated" or "abandoned" appears, an invalid edge is marked from the current node to the indicated node.

[0124] Further, for each edge, record the evidence page number, layout coordinates, hit phrase, coverage range (clause level or full text level), and effective version timestamp interval; when multiple types of edges of the same pair of nodes are hit at the same time, keep only one label in the order of invalidation first, revision second, coverage third, and reference fourth, and the rest as annotations.

[0125] S3.4, generate coverage chains and reference chains along coverage edges and reference edges in the order of the version sequence, and select the latest valid node set for each clause identifier by combining the revision edges and invalidation edges.

[0126] Further, from the earliest node, proceed in the order of coverage edge first and reference edge second, generate coverage chains and reference chains, and record the node order, edge source page number, and range description in the chains; when the same node appears with parallel successors, choose one path in the order of containing coverage prompt words first, page number later, and reading order later, and solidify it as the main path; exclude the nodes pointed to by the invalidation edges and their all successors, and fill in the jump relationship on the chain; mark the nodes affected by the revision edges as invalid and connect the revised nodes to the end of the coverage chain, if there is no coverage or reference reachable, then treat the node as a single-node coverage chain; find the first node that is not revised and not invalidated from the end of each cleaned chain, and record it as the latest valid node for the clause identifier, and finally, all the latest valid nodes of the clause identifiers are combined into the latest valid node set.

[0127] S3.5, merge the clause instance nodes, coverage edges, revision edges, reference edges, invalidation edges, coverage chains, reference chains, and the latest valid node set to form a time sequence knowledge graph.

[0128] Further, initialize the graph structure and fix the namespace and field dictionary; write the clause instance nodes one by one in the version sequence, and the node attributes fall into the clause identifier, version timestamp, file identifier, normalized result, evidence pointer and reading order, while generating a unique key and solidifying the previous pointer and subsequent pointer; according to the records of the covering edge, the revision edge, the reference edge and the abolition edge, a directed edge is established in batches, and the edge attribute is written into the evidence page number, the layout coordinate and the coverage range, and the single label is reserved in the order of abolition priority to revision, revision priority to coverage, coverage priority to reference; the coverage chain and the reference chain are converted into a searchable link index and the chain level order and chain length summary are written in the head and tail nodes; the latest valid node set is marked with an effective label and an effective interval and a dependent association is established; a multi-dimensional index is generated according to the clause identifier, the version timestamp and the file category, and finally it is stored as a time sequence knowledge graph.

[0129] S4, compile the cross-document consistency rules into executable rule sets that can be evaluated on the time sequence knowledge graph using executable recruitment constraint language.

[0130] S4.1, parse the cross-document consistency rule source according to the syntax of the executable recruitment constraint language using the controlled decoding language model and generate a rule candidate set.

[0131] The cross-document consistency rule source is received and merged into a parseable text according to the text type and page number order, and a syntax summary, field dictionary, operator list and fixed output template are prepared based on the executable recruitment constraint language. The source text is divided into rule candidate paragraphs according to the paragraph title, clause number and sentence anchor point; the semantic labels such as mandatory, prohibited, contained, range, mutual exclusion, implication, exception, reference and their context windows are extracted from each candidate rule paragraph, together with the source file identifier, version timestamp, page number and layout coordinate, and written into the prompt tail; the complete prompt is sent to the controlled decoding language model, enabling deterministic decoding and fixed key name hierarchy, prohibiting free text supplement, and outputting structured rule units by the controlled decoding language model, including rule name, action target, scope clue, operator, comparison object, exception condition and evidence pointer; each output is executed for syntax checking, checking whether the key name is complete, the operator is in the list, the field exists in the dictionary, the value type and range are matched, and the unqualified items are corrected once according to the template in a single call; the rules that pass the checking are aggregated into a rule candidate set.

[0132] S4.2, align the fields and unify the terms of the rule candidate set, the clause record set and the latest valid node set in the time sequence knowledge graph to form the cross-document consistency rules.

[0133] According to the rule candidate set, the clause record set and the latest valid node set in the time sequence knowledge graph, a standard field mapping table is established based on the preset Schema field list and the synonym table; the rule candidates are traversed one by one, the field names in the rule items are replaced with the standard field names, and the normalized results in the clause record set are unified in unit, currency and date caliber; according to the action target and field anchor point of the rule item, the corresponding clause instance node set is parsed in the latest valid node set and the evidence pointer is backfilled; when the rule item involves multiple fields, it is checked whether there is available value for each field on the corresponding node, and if missing, a missing mark is recorded and the adjacent previous node is located as a substitute value source; the rule candidates with the same source, target, operator and caliber are de-duplicated and merged, the source with later version timestamp is retained and the evidence location is summarized; finally, the cross-document consistency rule is generated, which contains rule identification, standard field name, value caliber, action target node set, applicable scope, source file identification, version timestamp and evidence pointer.

[0134] S4.3, determine the scope of each cross-document consistency rule in the time sequence knowledge graph and point to the latest valid node set.

[0135] The target field, comparison object, action target, applicable scope and exception condition of each cross-document consistency rule are parsed; the candidate clause instance node set is located according to the target field in the time sequence knowledge graph, and is expanded to all reachable nodes along the coverage chain and reference chain; the valid range is limited by the intersection of the latest valid node set and the candidate, and the revised or abolished nodes are removed according to the revision edge and the abolition edge; filtering is performed according to the applicable scope, including file category, clause number section, page range and version time window; clause instance nodes that meet the exceptions are removed from the valid range according to the exception condition; the scope of each cross-document consistency rule is generated, recording the pointed clause instance node, source file identification, version timestamp and evidence pointer; when the valid range is empty and the cross-document consistency rule contains a reference field, the previous valid version is backtracked along the reference chain once again.

[0136] S4.4, read the normalized results in the clause record set corresponding to the node in the scope, configure the tolerance strategy of unit, currency and date, and complete the compilation to obtain the executable rule set.

[0137] The normalized results in the clause record set are located by node within the scope, and the values or intervals, units, currencies, dates or periods, and resolutions and endpoint meanings are read; based on the read results, unit tolerances (small deviations after uniformity to target units and rounding strategies), currency tolerances (precision bits and small difference windows after uniformity of currency), and date tolerances (determination of equal and inclusion judgment thresholds and time windows according to daily, monthly, and other resolutions) are configured for the current rule; the comparison objects, action targets, and tolerance items of the rule are assembled into executable statements; the executable statements are statically checked to verify whether the field names, value thresholds, tolerance items, and scopes are complete and consistent, and the record reasons that do not pass are corrected and returned to the scope index; the executable rule set is generated by summarizing the statements that pass the check. The executable rule set includes MUST_EQUAL operators, IN operators, RANGE operators, MUTEX operators, IMPLIES operators, and EXCEPT_IF operators.

[0138] S5, based on the executable rule set, the scope of each rule is evaluated on the time sequence knowledge graph, and conflict candidates are generated, and after convergence, a conflict candidate set is obtained.

[0139] S5.1, based on the executable rule set, the scope of each rule in the time sequence knowledge graph is selected, and each rule is evaluated to obtain the consistency state of each rule.

[0140] The intersection of the rule scope and the corresponding latest valid node set in the time sequence knowledge graph is located, and the revised or abolished nodes are removed; each rule is evaluated by operator:

[0141] MUST_EQUAL: The values that need to be equal in the scope are unified in units and currencies, and then compared after alignment according to the date resolution; if the difference does not exceed the unit tolerance and currency tolerance and the dates are equal or within the date tolerance window, it is recorded as satisfied, otherwise it is recorded as not satisfied, and the difference direction and difference amount are recorded;

[0142] IN: The target value is mapped to a standard token according to the field synonym and enumeration alias, and is searched in the allowed set; a hit is recorded as satisfied, and a miss is recorded as not satisfied, and the miss token and allowed set summary are recorded;

[0143] RANGE: Convert the numerical value or date to an interval or scalar and compare according to the endpoint inclusion semantics; located within the interval or satisfying the inclusion relationship is recorded as satisfied, and lower than the lower limit or higher than the upper limit is recorded as not satisfied, and the out-of-range end and deviation are recorded;

[0144] MUTEX: Cluster by standardized value within the scope, detect whether the same mutual exclusion group appears with an allowed coexistence; no coexistence is recorded as satisfied, and coexistence is recorded as not satisfied, and the conflict member list is listed;

[0145] IMPLIES: first determine the antecedent; evaluate the consequent when the antecedent is satisfied, record as satisfied if the consequent is satisfied, record as satisfied (vacuously) if the antecedent is not satisfied, and keep the antecedent value for traceability;

[0146] EXCEPT_IF: evaluate the exception condition after the operator determines; if the exception hits, change the "not satisfied" to "explainable inconsistency" and mark the exception basis and evidence location; if the exception does not hit, maintain the original conclusion.

[0147] For example, the amount consistency requirement is MUST_EQUAL, the difference between the two file amounts falls within the currency tolerance after the amounts are unified to the same currency, and the date is within the allowable window, which is recorded as satisfied; if the difference exceeds the window but there is an "as is" exception condition that hits, it is recorded as an explainable inconsistency.

[0148] Each determination is written to the conclusion, value range, tolerance details, evidence pointer, and link segment to the temporary results table.

[0149] S5.2, based on the rule that the consistency state is not satisfied, generate conflict candidates.

[0150] From the temporary results table, select records with a conclusion of "not satisfied", and generate conflict candidates according to rule identification and scope; conflict candidates include operator type, violated field or field pair, involved clause instance node list, corresponding coverage chain and reference chain segment, normalized value and list expression, tolerance range, difference direction or out-of-bound end, evidence pointer and mismatch description; if it is a cross-document constraint, the nodes and values of each source file are listed in the same candidate; if the same record is rewritten by EXCEPT_IF as "explainable inconsistency", mark the exception hit and exception basis in the candidate to facilitate subsequent hit set calculation and identify the exempted impact.

[0151] S5.3, aggregate all conflict candidates to get the conflict candidate set.

[0152] Merge all conflict candidates by rule identification, operator type, clause identification, field name, and version time window as keys; merge multiple inconsistencies caused by the same reason into a single representative item, and incorporate the remaining evidence locations into the evidence list of the representative item; when the same pair of sources appears repeatedly, determine the representative item according to the inclusion coverage prompt word first, page number last, and reading order last; calculate the impact range label (clause level or full-text level) and disposition priority label (correction first, replacement first, etc.) for each candidate, and complete the audit information (source file identification, version timestamp, and evidence summary); output the conflict candidate set sorted by grouping index and priority.

[0153] S6. Based on the conflict candidate set, calculate the minimum hit set that makes all conflict candidates resume to be satisfied, obtain the minimum conflict explanation set and locate the root cause clause instance, and generate the minimum repair suggestion according to the normalized results of the violated rule operators and nodes.

[0154] S6.1. Establish the correspondence between the conflict and the clause instance node based on the conflict candidate set, and calculate the minimum hit set that makes all conflict candidates resume to be satisfied.

[0155] The correspondence is established between the conflict candidate and the changeable object, and the changeable object is limited to the clause instance node and field pair in the scope that can be changed by one atomic change. The type of atomic change is determined by the operator, such as value alignment, unit caliber alignment, currency unification, date endpoint caliber alignment, removing an item from a mutual exclusion group, supplementing a consequent or withdrawing an antecedent for implication, and supplementing a trigger condition for exception. The number of conflict candidates that can be resumed to be satisfied by each changeable object is registered in the correspondence, forming a coverage mapping from changeable object to conflict candidate.

[0156] The minimum hit set is obtained by deterministic step-by-step deepening backtracking search. Specifically, the expansion order is determined from high to low according to the number of conflict candidates covered by the changeable object, and the order of the same coverage object is determined according to the fixed rule of containing coverage prompt word first, version timestamp later, and reading order later. From the object with the largest number of hits, the upper limit of the number of hits is increased, backtracking is performed, pruning occurs when the coverage subset is completely covered by the selected object, and backtracking occurs when the remaining conflict candidate cannot be covered. The first solution that covers all conflict candidates is the minimum hit set.

[0157] For example, when the amounts of multiple files are inconsistent and there are mutual exclusion options superimposed, the changeable objects are the value alignment of the amount field and the removal of the mutual exclusion clause, respectively. Each of the two types of objects covers a number of conflict candidates. The backtracking search first tries to cover the amount alignment object with larger coverage, and then supplements the mutual exclusion removal object if there are remaining conflicts, until all conflicts are covered.

[0158] S6.2. Determine the minimum conflict explanation set based on the minimum hit set and locate the root cause clause instance on the temporal knowledge graph.

[0159] Further, based on the minimum hit set, all conflict candidates covered by the hit changeable objects are collected, the inconsistencies that can be directly explained by the exception condition are removed, and the remaining set is taken as the candidate explanation set; in the candidate explanation set, the conflict candidates that have a subordinate relationship with the same changeable object are checked in turn, and multiple records that are simultaneously eliminated by the same atomic change are folded into an explanation unit, which contains the rule identifier, operator type, field name, changeable object identifier, and evidence pointer; the minimum conflict explanation set is obtained in units of explanation units, and in the time sequence knowledge graph, the unique clause instance node is located according to the changeable object, and the source node that causes the current value to enter the latest valid state is tracked in reverse combination with the coverage chain and the reference chain, if the source node is located in the correction announcement or the replacement range, the source node is recorded as the root cause clause instance, otherwise the current clause instance node is recorded as the root cause clause instance; for each root cause clause instance, the file identifier, version timestamp, file category, normalized result, and evidence location are recorded, and a one-to-one association is established with the corresponding explanation unit, and the minimum conflict explanation set and the root cause clause instance index are output.

[0160] S6.3, based on the operator of the violated rule and the normalized result of the clause instance node, a minimum repair suggestion is generated.

[0161] Specifically, the minimum conflict explanation set and the root cause clause instance index are read one by one, deterministic repair action templates are formulated according to the operator type and filled with normalized results and evidence pointers: for equality constraints, suggestions for aligning to reference node values are given, and reference nodes prefer to select the latest valid node containing coverage cues and with a later version timestamp; for set inclusion constraints, suggestions for mapping current values to allowed set standard tokens or replacing them with the most consistent item in the set are given; for range constraints, suggestions for moving values to the nearest boundary or adjusting the endpoint range to achieve inclusion are given, and the endpoint inclusion or exclusion semantics are explicitly stated; for mutual exclusion constraints, suggestions for revoking lower priority options in the conflict group or retaining the latest valid option with coverage cues are given; for implication constraints, suggestions for completing the consequent field or withdrawing the antecedent expression are given, and both prefer to select the smaller range that can meet the action; for exception conditions, suggestions for supplementing exception trigger conditions or explicitly stating exception clauses in new versions are given; each repair suggestion carries the source file identifier, version timestamp, page number and layout coordinates, and gives the affected clause range label (clause level or full-text level) and the suggestion landing point (coverage chain end or reference chain target); the group of all repair actions that can simultaneously cover the corresponding objects of the minimum hit set is taken as the minimum repair suggestion

[0162] S7, according to the root cause clause instance and the minimum repair suggestion, an auditable verification result report is output.

[0163] S7.1, based on the root clause instance and the minimum repair suggestion, gather the evidence pointers in the page fragment object set and assemble the evidence chain.

[0164] Further, read the recorded evidence pointers in the root clause instance, generate the first batch of evidence entries accordingly; trace back from early to late along the coverage chain and the reference chain to the starting node that affects the root clause instance value, extract the evidence pointer and the coverage cue position of the same field at each passed node, and merge the corresponding original excerpts, normalized ranges, and change statements into new evidence entries; perform deduplication and sorting on all evidence entries, and retain the representative entries according to the fixed rule of "containing coverage cues first, version timestamp later, reading order later"; the rest are incorporated into the reference list of the representative entries; label the role of each evidence according to the minimum repair suggestion: "reference evidence" if consistent with the target value of the suggestion, "to-be-repaired evidence" if corresponding to the to-be-repaired value, and "exception evidence" if triggering or justifying the exception condition; generate the evidence chain in chronological order and link type, and establish the forward and backward pointers and cross-references; check the integrity of the evidence chain (the starting point can explain the source of the current value, the ending point points to the root clause instance, and the reference evidence is reachable), if there is a gap, rearrange after filling in the missing evidence according to the segment identifier in the page fragment object set; output the evidence chain with sequential numbering and traceable index.

[0165] S7.2, based on the evidence chain and the minimum repair suggestion, generate an auditable verification result report.

[0166] According to the evidence chain and the minimum repair suggestion, assemble the verification result report in a fixed format: write the project name, tender number, purchaser name, and generation time in the cover section; list the violated rule list (rule identifier, operator, scope, tolerance summary) and consistency status statistics in the overview section; display the evidence chain one by one in the fact section: list the file identifier, version timestamp, file category, page number and layout coordinates, field name, original text and normalized result for each node, and mark the link label and coverage range description between adjacent nodes, while giving the "to-be-repaired evidence → reference evidence" contrast relationship; give a mismatch explanation for each rule that does not meet the rule in the analysis section, including the difference direction, the difference or the out-of-bound end, whether the exception is hit, and using the evidence chain node number as the index; fall into the minimum repair suggestion in the disposal section, and clearly suggest the action (value alignment, unit range unification, currency unification, date endpoint adjustment, mutual exclusion item withdrawal, supplement of implied antecedent or establishment of exception clause), suggestion landing point (end of coverage chain or target of reference chain), expected evaluation result and impact range (clause level or full-text level); list the page fragment object identifier list and fingerprint summary, evidence chain integrity check result, and mapping table and tolerance strategy version summary in the audit section; finally generate the report number and output both machine-readable structure and readable text, both of which are embedded with evidence chain node number and page coordinates.

[0167] In summary, the present application compiles cross-document consistency rules into executable rule sets that can be evaluated on a temporal knowledge graph by using an executable rule language, realizes accurate scope evaluation of rules on the latest valid node set, and is used for automatically determining consistency and exceptions; calculates the minimum hit set based on the conflict candidate set, obtains the minimum conflict explanation set and locates the root cause clause instance, generates the minimum repair suggestion, and is used for guiding value alignment and caliber unification, finally eliminates the conflict with the minimum change and outputs the traceable verification result.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligent data processing of procurement documents based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain multi-source, multi-version files pointing to the same procurement project, perform layout correction and OCR on the multi-source, multi-version files to obtain a set of page fragment objects, and establish a version sequence; S2. Based on the set of page fragment objects, extract the clause records according to the preset schema using a language model with controlled decoding to form a set of clause records; S3. Based on the set of clause records and version sequence, cross-version alignment is performed on clauses with the same name or synonyms to construct a temporal knowledge graph; Specifically: Cross-version alignment of clauses with the same name or synonyms is determined based on the set of clause records and the version sequence. Generate a clause instance node for each version based on cross-version alignment; Label the clause instance nodes with overlay edges, revision edges, reference edges, and repeal edges according to the version sequence; Generate covering chains and referencing chains along the covering edges and referencing edges in the order of version sequence, and select the latest valid node set for each clause identifier by combining the markings of the revision edges and repeal edges, and the clause instances that are later in the version sequence and have not been revised or repealed. The clause instance nodes, covering edges, revision edges, referencing edges, repeal edges, covering chains, referencing chains, and the latest valid node set are merged to form a temporal knowledge graph; S4. Use an executable procurement constraint language to compile cross-document consistency rules into an executable rule set that can be evaluated on a temporal knowledge graph; S5. Based on the executable rule set, evaluate the scope of each rule on the time-series knowledge graph and generate conflict candidates. After aggregation, obtain the conflict candidate set. S6. Based on the conflict candidate set, calculate the minimum hit set that restores all conflict candidates to the required state, obtain the minimum conflict explanation set and locate the root cause clause instance, and generate the minimum repair suggestion based on the normalization result of the operator and node of the violated rule. S7. Based on the root cause clause examples and minimum repair recommendations, output an auditable verification result report.

2. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, In S1, the multi-source, multi-version documents include tender notices, tender documents or procurement requirements, clarification documents or Q&A documents and addenda, correction notices, tender documents and tender document attachments, bid evaluation reports, notices of award, contract texts and contract amendment documents, acceptance reports, prequalification documents, bid opening records, technical specifications, and drawings and price lists.

3. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 2, characterized in that, The multi-source, multi-version files are subjected to layout correction and OCR to obtain a set of page fragment objects, and a version sequence is established, including: Perform layout correction on multi-source, multi-version files to generate a geometrically normalized page image while retaining file identifiers and version timestamps; Perform OCR on the geometrically normalized page image to obtain recognition results containing text content and page coordinates, and form page fragment candidates; Construct a collection of page fragment objects based on page fragment candidates, file identifiers, version timestamps, page numbers, and page layout coordinates; A version sequence is established for multi-source, multi-version files based on version timestamps, and the overwrite relationship is recorded.

4. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, S2 specifically refers to: Content segmentation is performed based on a collection of page fragment objects to obtain a collection of text segments that are consistent with the mapping of preset schema fields. Based on a preset schema, a language model prompt for controlled decoding is constructed, and the output fields and output format are limited. The set of text segments and the prompt are combined to form a controlled decoding input. The term record is generated by the controlled decoding language model based on the controlled decoding input, and an evidence pointer to a set of page fragment objects is attached. The terms and conditions records are normalized based on the unit, currency, and date information and then aggregated to form a set of terms and conditions records.

5. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, S4 specifically refers to: Using a controlled decoding language model, cross-document consistency rule sources are parsed according to the syntax of an executable procurement constraint language, and a set of rule candidates is generated. By aligning the fields and unifying the terminology of the rule candidate set, the clause record set, and the latest valid node set in the temporal knowledge graph, cross-document consistency rules are formed. In the temporal knowledge graph, determine the scope of each cross-document consistency rule and point it to the latest set of valid nodes; Within the scope, read the normalized results from the set of terms records corresponding to the node, configure the tolerance strategies for units, currencies, and dates, and complete the compilation to obtain an executable rule set.

6. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 5, characterized in that, The executable rule set includes the MUST_EQUAL operator, IN operator, RANGE operator, MUTEX operator, IMPLIES operator, and EXCEPT_IF operator.

7. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, S5 specifically refers to: Based on the executable rule set, the scope of each rule in the time-series knowledge graph is selected, and each rule is evaluated to obtain the consistency state of each rule. Based on the rules that the consistency state is not satisfied, conflict candidates are generated; All conflict candidates are aggregated to obtain the conflict candidate set.

8. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, S6 specifically refers to: Establish the correspondence between conflict and clause instance nodes based on the conflict candidate set, and calculate the minimum hit set that restores all conflict candidates to the desired state. Determine the minimum conflict interpretation set based on the minimum hit set and locate root cause clause instances on the temporal knowledge graph; Based on the normalized results of the operators and clause instance nodes of the violated rules, a minimum repair suggestion is generated.

9. The intelligent data processing method for procurement documents based on artificial intelligence as described in claim 1, characterized in that, S7 specifically refers to: Based on the root cause clause examples and minimum repair recommendations, evidence pointers are gathered from the set of page fragment objects and an evidence chain is assembled. Generate auditable verification results reports based on the chain of evidence and minimum repair recommendations.

Images