Methods and Systems for Detecting Bidirectional Consistency of Heterogeneous Policy Documents

By using hierarchical constraint graphs and bidirectional hierarchical recall methods, combined with multi-dimensional information calculation, the problems of large computational load and incorrect matching in the consistency detection of heterogeneous policy documents are solved, and efficient and reliable policy document consistency analysis is achieved.

CN123068614APending Publication Date: 2026-11-10QUANZHOU INST OF EQUIP MFG +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611576336.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively handle heterogeneous document structures, policy constraint scopes, and two-way coverage relationships in policy document consistency analysis, resulting in high computational costs, frequent mismatches, and an inability to identify changes in the strength and duration of textual obligations.

Method used

A hierarchical constraint graph construction and bidirectional hierarchical recall method are adopted, combined with six-dimensional correlation scores and comprehensive confidence calculation, to identify conflicts and omissions among policy documents. Through structural path filtering, sparse feature recall and semantic vector recall, highly relevant clause pairs are screened out, and the confidence of conflict events is calculated using multi-dimensional information.

Benefits of technology

It improves the accuracy and reliability of policy document consistency detection, reduces computational resource consumption, reduces false matches, provides a traceable chain of evidence, and adapts to different document formats and parsing quality differences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068614A_ABST
    Figure CN123068614A_ABST
Patent Text Reader

Abstract

This application provides a method and system for bidirectional consistency detection of heterogeneous policy documents. The method includes: acquiring the heterogeneous policy documents to be compared and parsing them to obtain a hierarchical constraint graph; dividing the hierarchical constraint graph into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes; performing hierarchical recall and pairing filtering along a first direction from the first constraint unit set (as the query side) to the second constraint unit set (as the candidate side), and along a second direction from the second constraint unit set (as the query side) to the first constraint unit set (as the candidate side) to obtain a list of clause pairs; calculating the six-dimensional association score of each clause pair in the list of clause pairs based on the node information of the structural nodes, constraint element nodes, and subordinate nodes of the constraint units, and filtering out each target clause pair; and determining the comprehensive confidence level of conflict events for each target clause pair by combining the node information associated with each target clause pair and the preset constraint rules.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method and system for detecting bidirectional consistency of heterogeneous policy documents. Background Technology

[0002] Policy consistency analysis typically requires establishing correspondences between higher-level policies and local policies, similar policies in different regions, or different versions of the same policy. Existing computer processing solutions often employ text segmentation, vector encoding, similarity retrieval, and fixed-threshold classification, classifying clauses with high similarity as consistent and those with low similarity as different or newly added. While these solutions can assist in finding clauses with similar wording, their output is easily affected by file format, clause structure, and text length.

[0003] First, the structure of chapters, sections, articles, clauses, and items in different policy documents is inconsistent across Word formats, PDF layouts, scanned images, and webpage tags. Spread-out clauses, headers and footers, numbering errors due to optical character recognition, and non-standard numbering can cause errors in clause boundaries, leading to distortion in subsequent vector representations. Second, a one-way "local clause retrieval of superior clauses" can only indicate which superior clause a local clause might correspond to; it cannot identify requirements already stipulated in superior policies but not covered or only partially covered by local policies. Third, high semantic similarity does not equate to consistent constraints. For example, changing "should be completed within 30 days" to "can be completed within 60 days" results in highly similar texts, but a substantial change in the intensity and timeframe of the obligation.

[0004] Furthermore, fixed thresholds cannot adapt to changes in clause length, policy subject matter, document quality, and OCR confidence; performing pairwise vector comparisons and deep constraint analysis on all clauses also results in high computational cost and redundant calculations in a general-purpose CPU environment. Therefore, a computer implementation scheme is needed that can combine heterogeneous document structures, policy constraint scope, bidirectional coverage relationships, and confidence processing to improve the decidability of clause correspondences, reduce false matches, and control CPU resource consumption. Summary of the Invention

[0005] This application provides a method for detecting bidirectional consistency of heterogeneous policy documents, the method comprising: Obtain the heterogeneous policy documents to be compared and parse them to obtain the hierarchical constraint graph; Based on the document nodes corresponding to the heterogeneous policy documents, the hierarchical constraint diagram is divided into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes; Hierarchical recall and pairing filtering are performed along a first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along a second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side to obtain a list of clause pairs. Based on the node information of the structural nodes, constraint element nodes, and subordinate nodes of the constraint unit, the six-dimensional association score of each clause pair in the clause pair list is calculated, and the target clause pairs that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score are selected. By combining the node information and preset constraint rules associated with each target clause, the overall confidence level of conflict events for each target clause is determined. By comparing the overall confidence level with the preset review threshold, a structured conclusion is output that represents the constraint conflict of heterogeneous policy documents or those that need to be manually reviewed.

[0006] In some embodiments, the auxiliary nodes include effectiveness time nodes and original text location nodes, and the hierarchical recall includes structural path filtering, sparse feature recall and semantic vector recall. Structural path filtering refers to eliminating candidate constraint units whose hierarchical context is obviously irrelevant based on the structural paths carried by the structural nodes of the query-side constraint unit and the candidate-side constraint unit. Sparse feature recall refers to filtering candidate side constraint units retained after structural path filtering based on the constraint elements representing policy themes and keywords carried by the constraint element nodes of the query side constraint units and the validity time type constraint information carried by the validity time nodes, and removing candidate side constraint units with mismatched themes, no overlap of keywords, or invalid validity time. Semantic vector recall refers to retrieving the semantic vectors bound to the structural nodes of the query-side constraint units and the candidate-side constraint units retained through sparse feature recall, performing quantization processing, using the quantized semantic vectors to perform similarity retrieval to obtain semantic similarity, and sorting them in descending order of similarity, selecting the top... Each candidate constraint unit forms a recall candidate unit set, wherein the semantic vector is generated by encoding based on the node information carried by its respective structural node, as well as the associated constraint element node, effectiveness time node, and original text location node.

[0007] In some embodiments, pair filtering yields a list of terms pairs, including: For the first direction, each first constraint unit is used as a query-side constraint unit, and each second candidate unit obtained by hierarchical recall of the first constraint unit is paired with the first constraint unit to form a first clause pair list; For the second direction, each second constraint unit is used as a query-side constraint unit, and each first candidate unit obtained by hierarchical recall of the second constraint unit is paired with the second constraint unit to form a second clause pair list; The clause pairs in the second clause pair list are normalized in the order of first constraint unit first and second constraint unit last, and then merged with the first clause pair list to remove duplicates, resulting in the clause pair list.

[0008] In some embodiments, the selection of target clause pairs from the target candidate side constraint unit whose six-dimensional correlation score is the highest includes: The six-dimensional correlation scores of each clause pair are summarized to form a set of candidate scores for the query-side constraint units of each clause pair; Select the target candidate side constraint unit corresponding to the highest six-dimensional correlation score from the candidate score set; If the highest six-dimensional correlation score is greater than or equal to the matching threshold, then a first set of directed covering edges or a second set of directed covering edges pointing from the query-side constraint unit to the target candidate-side constraint unit is established in the first or second direction corresponding to the query-side constraint unit. A bidirectional coverage matrix is ​​constructed based on the first and second directed coverage edge sets of each query-side constraint unit. Each matrix element represents the bidirectional coverage of the clause pair formed by each query-side constraint unit and the target candidate-side constraint unit. If the highest six-dimensional correlation score is less than the matching threshold, the query-side constraint unit will be assigned to either the first uncovered set or the second uncovered set. The matching status of each clause pair is determined based on the bidirectional coverage, the first uncovered set, the second uncovered set, and the retrieved constraint coverage auxiliary results. Select target clause pairs from the two-way coverage matrix that match the two-way high coverage and one-way partial coverage status.

[0009] In some embodiments, the subordinate node also includes the original text location node, and the six-dimensional association score includes the hierarchical path consistency score, constraint element matching score, parsing confidence score, effectiveness time matching score, numerical relationship consistency score, and semantic similarity score. Among them, the hierarchical path consistency score is calculated based on the structural path carried by the structural nodes of the clause pair; the constraint element matching score is calculated based on the constraint element content, the start and end range of the original text characters, and the scope pointer carried by the constraint element nodes of the clause pair. The analytical confidence score is calculated based on the analytical confidence score carried by the constraint element nodes of the clause pair; The effectiveness time matching score is calculated based on the effectiveness time constraint information carried by the effectiveness time nodes of the clause pair; The numerical relationship consistency score is calculated based on the numerical restriction constraint elements in the constraint element nodes of the clauses, after unit normalization of the numerical range and the results of comparison operator operations. The semantic similarity score is calculated based on the semantic vectors bound to the structural nodes.

[0010] In some embodiments, determining the matching threshold includes: Statistically analyze the score distribution characteristics of the candidate score set for each clause pair's query-side constraint unit; The length of the clause is calculated based on the page coordinates carried by the original text location node of the query-side constraint unit and then normalized to obtain the length correction amount. The parsing quality correction is obtained by taking the complement of the parsing confidence carried by the constraint element node of the query-side constraint unit. The matching threshold of the query-side constraint unit is calculated based on the score distribution characteristics, length correction amount, and parsing quality correction amount.

[0011] In some embodiments, the overall confidence level of conflict events for each pair of target clauses is determined by combining the associated node information and preset constraint rules, including: Based on the preset constraint rules, the constraint relationship operation is performed by combining the constraint element content, original character range and scope pointer carried by the constraint element node of each target clause pair, and the constraint conflict event set of each target clause pair is identified and output. For each constraint conflict event in the constraint conflict event set, the completeness of the evidence source of the original text location node, the multi-candidate ambiguity of the target clause pair, and the extraction confidence of the constraint element are determined by combining the node information of the target clause pair related to the constraint conflict event. The comprehensive confidence of the conflict event is calculated by combining the six-dimensional association score of the relevant target clause pair and the certainty of the retrieved preset constraint rules. Among them, the certainty of the constraint rule is the preset rule score corresponding to the conflict type; the extraction confidence of the constraint element is obtained by aggregating the analytical confidence of the constraint element nodes participating in the conflict determination; the completeness of the evidence source of the constraint element is determined based on the completeness of the source file identifier, page number, and page coordinates carried by the original text positioning node associated with the constraint element participating in the determination; the multi-candidate ambiguity of the target clause pair is determined based on the difference in the six-dimensional association scores between the best and second-best candidates in the candidate score set corresponding to the query-side constraint unit.

[0012] In some embodiments, the heterogeneous policy documents include a first policy document and a second policy document, and the method further includes: If the version of the first policy document is updated, the updated first and second policy documents will be parsed into a new level constraint graph; Calculate the content fingerprint of each node in the new hierarchical constraint graph and the hierarchical constraint graph respectively; Compare the content fingerprints of each node to determine the set of newly added nodes, deleted nodes, and changed nodes; For nodes whose fingerprints have not changed, the cached node semantic vectors, the topological relationships of the nodes in the hierarchical constraint graph, and the recall candidate unit set corresponding to the nodes are reused. Based on the set of changed nodes, determine the set of affected constraint units. The set of affected constraint units includes the constraint units that have changed, as well as the constraint units whose cached recall candidate unit set contains the constraint units with changed nodes. For nodes that have not undergone fingerprint changes and do not belong to the set of affected constraint units, the cached node semantic vector, the topological relationship of the node in the hierarchical constraint graph, and the recall candidate unit set corresponding to the node are reused. The hierarchical recall, pairing screening, six-dimensional association score calculation, bidirectional coverage relationship detection, conflict event comprehensive confidence calculation, and structured conclusion output steps are re-executed only for the set of changed nodes and the set of affected constraint units.

[0013] In some embodiments, the overall confidence level is compared with a preset review threshold to output a structured conclusion characterizing the constraint conflict of heterogeneous policy documents or those requiring manual review, including: If the overall confidence level is higher than the preset review threshold, a structured conclusion representing the constraint conflict between the first set of policy documents and the second set of policy documents is output.

[0014] If the overall confidence level is lower than or equal to the preset review threshold, a structured conclusion representing the need for manual review will be output.

[0015] This application also provides a system for detecting bidirectional consistency of heterogeneous policy documents, the system comprising: The hierarchical constraint graph construction module is used to obtain heterogeneous policy documents to be compared and parse them to obtain hierarchical constraint graphs. The constraint graph partitioning module is used to partition the hierarchical constraint graph into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes based on the file nodes corresponding to the heterogeneous policy documents. The bidirectional hierarchical recall module is used to perform hierarchical recall and pair filtering to obtain a list of clause pairs by pointing along a first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along a second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side. The score and target pair filtering module is used to calculate the six-dimensional association score of each clause pair in the clause pair list based on the node information of the structural nodes, constraint element nodes and subordinate nodes of the constraint unit, and to filter out each target clause pair that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score. The confidence calculation module is used to determine the overall confidence of conflict events for each target clause pair by combining the associated node information and preset constraint rules. The structured conclusion output module is used to compare the overall confidence with the preset review threshold and output a structured conclusion that represents the constraint conflict of heterogeneous policy documents or is pending manual review.

[0016] Understandably, this application provides a method and system for bidirectional consistency detection of heterogeneous policy documents. First, it parses and constructs a hierarchical constraint graph from the heterogeneous policy documents, dividing them into a set of constraint units and a set of subordinate nodes. This method is compatible with formats such as Word, PDF, and scanned documents, and incorporates the document's hierarchical structure, constraint elements, effective time, and original text location information into a unified graph structure representation. This avoids the problem of clause boundary confusion caused by different document layouts and OCR recognition defects, providing a structured foundation for subsequent comparisons. Second, it employs bidirectional hierarchical recall and pairing filtering in the first and second directions to obtain a list of clause pairs. Unlike traditional unidirectional retrieval, this method can simultaneously detect policy omissions, partial coverage, and newly added requirements. The hierarchical recall includes structural path filtering, sparse features, and semantic vector recall, significantly compressing the size of clause pairs requiring in-depth computation. This limits in-depth analysis to a limited number of candidates, reduces CPU computational overhead, and avoids the performance loss caused by full pairwise comparisons. Furthermore, by utilizing six-dimensional correlation scores to integrate hierarchical paths, constraint elements, resolution confidence, effectiveness time, numerical relationships, and semantic multi-dimensional information, it no longer relies solely on textual semantic similarity. It can identify situations where texts have similar wording but are substantially inconsistent in terms of obligations, values, and timeliness. By combining candidate score distribution and other information to select the optimal target clause pair, it can distinguish between two-way high coverage, one-way partial coverage, and no coverage matching states, reducing false matches caused by fixed thresholds. Finally, it integrates multi-type node information to calculate the comprehensive confidence of conflict events, correlates with original text location, extracts confidence and candidate ambiguity, and then compares it with the review threshold to output structured conclusions. For scenarios with poor resolution quality and incomplete evidence, it outputs manual review prompts, forming a traceable evidence chain, reducing erroneous judgments caused by OCR resolution errors, and improving the reliability and interpretability of policy conflict detection. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 is a schematic diagram of the overall process of the heterogeneous policy document bidirectional consistency detection method provided in the embodiments of this application; Figure 2 is a schematic diagram of the hierarchical constraint diagram provided in an embodiment of this application; Figure 3 is a schematic diagram of bidirectional coverage relationship detection provided in an embodiment of this application; Figure 4 is a schematic diagram of the calculation of the comprehensive confidence level of conflict events provided in the embodiments of this application; Figure 5 is a schematic diagram of CPU environment hierarchical recall and incremental update processing provided in an embodiment of this application; Figure 6 is a functional structure diagram of the heterogeneous policy document bidirectional consistency detection system provided in the embodiments of this application.

[0019] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] The terms “first”, “second”, etc. used in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated.

[0022] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0023] Please combine Figure 1 , Figure 1 A flowchart illustrating the heterogeneous policy document bidirectional consistency detection method and system provided in this application may include the following steps: Step S110: Obtain the heterogeneous policy documents to be compared and parse them to obtain the hierarchical constraint diagram.

[0024] Specifically, such as Figure 2 As shown, the hierarchical constraint layer contains file nodes, structure nodes, constraint feature nodes, validity time nodes, and original text location nodes. Nodes are connected by contain, constraint, exception, and association edges, and each node carries node information.

[0025] Step S120: Based on the document nodes corresponding to the heterogeneous policy documents, divide the hierarchical constraint diagram into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes. Specifically, first, the articles, clauses, and items of the policy documents are divided into independently comparable clause units, and then a constraint unit is constructed for each clause unit that enters the comparison. A constraint unit consists of the structural node corresponding to the clause unit and one or more constraint element nodes extracted from it; if the same clause unit contains multiple constraint elements, it is still counted as one constraint unit. Structural nodes such as pure titles that do not carry constraint content are only used to construct hierarchical paths and are not included in the number of constraint units. If an original article is divided into multiple independently comparable clauses or items, corresponding constraint units are formed for each.

[0026] Step S130: Perform hierarchical recall and pairing filtering along the first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along the second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side to obtain a list of clause pairs.

[0027] It should be noted that document nodes, effective time nodes, and original text location nodes are subordinate nodes associated with constraint units and are not counted separately in the total number of units in the first constraint unit set or the total number of units in the second constraint unit set.

[0028] In one implementation, the auxiliary nodes include effectiveness time nodes and original text location nodes, and the hierarchical recall includes structural path filtering, sparse feature recall, and semantic vector recall. The structural path filtering refers to eliminating candidate side constraint units whose hierarchical context is obviously irrelevant, based on the structural paths carried by the structural nodes of the query side constraint unit and the candidate side constraint unit. The sparse feature recall refers to filtering the candidate side constraint units retained after structural path filtering based on the constraint elements representing policy themes and keywords carried by the constraint element nodes of the query side constraint units and the validity time type constraint information carried by the validity time nodes, and removing candidate side constraint units that do not match the theme, have no overlap in keywords, or have expired validity time. The semantic vector recall refers to retrieving the semantic vectors bound to the structural nodes of the query-side constraint unit and the candidate-side constraint units retained through sparse feature recall, performing quantization processing, using the quantized semantic vectors to perform similarity retrieval to obtain semantic similarity, sorting them in descending order of similarity, and selecting the top... The candidate side constraint units form a recall candidate unit set, wherein the semantic vector is generated by encoding based on the node information carried by the respective structural nodes and the associated constraint element nodes, effectiveness time nodes and original text positioning nodes.

[0029] Specifically, semantic vectors are generated from a 384-dimensional Chinese sentence vector model and run on the CPU using the ONNX Runtime; quantization can be performed using INT8 quantization; similarity retrieval can use the FAISS CPU index, HNSW index, or other approximate nearest neighbor structures, selecting the top results after sorting by similarity in descending order. The above model and index are only for implementation selection and do not constitute a limitation on the scope of protection.

[0030] Understandably, the hierarchical recall process sequentially filters candidates through structural path filtering, sparse feature recall, and semantic vector recall. First, it eliminates context-irrelevant units at each level; then it filters out objects that do not match in terms of topic, keywords, or validity period; finally, it relies on semantic vectors for similarity retrieval, narrowing the candidate range layer by layer. This approach eliminates a large number of obviously irrelevant objects, reducing the computational overhead of subsequent six-dimensional scores, while retaining highly relevant candidates, balancing retrieval efficiency and recall quality, and adapting to the comparison needs of heterogeneous policy documents in CPU environments.

[0031] In one implementation, the pairing filter yields a list of terms pairs, including: For the first direction, each first constraint unit is used as a query-side constraint unit, and each second candidate unit obtained by hierarchical recall of the first constraint unit is paired with the first constraint unit to form a first clause pair list; For the second direction, each second constraint unit is used as a query-side constraint unit, and each first candidate unit obtained by hierarchical recall of the second constraint unit is paired with the second constraint unit to form a second clause pair list; The clause pairs in the second clause pair list are normalized in the order of first constraint unit first and second constraint unit last, and then merged with the first clause pair list to remove duplicates, resulting in the clause pair list.

[0032] The terms in the list satisfy the following formula: , This indicates the number of terms in the terms list. This represents the total number of elements in the first constraint element set. This represents the total number of elements in the second constraint element set. This indicates the number of candidate constraint units corresponding to a single query-side constraint unit. , Not greater than 、 The smaller of the two.

[0033] Understandably, this technique is based on a candidate unit set obtained through bidirectional recall, which is used to construct clause pairs for both query directions, and then merged and deduplicated to obtain the final clause pair list. → , → Two-way pairing overcomes the limitations of traditional one-way retrieval. It can capture the matching relationship between a first policy document and a second policy document, and also identify the coverage and addition of the second policy document to the first, avoiding the risk scenarios of "a higher-level regulation that is not implemented at the lower level" or "a lower-level newly added constraint" that might be missed in one-way comparisons. Pairing is performed only within the candidate set output by the hierarchical recall, rather than combining all constraint units in pairs. The constraints significantly reduce the number of clause pairs requiring in-depth analysis, avoiding the computational surge caused by full comparison and adapting to CPU operating environments. Merging and deduplication processes eliminate duplicate clause pairs generated by bidirectional pairing, preventing repeated six-dimensional association score calculations and conflict determinations for the same set of clause pairs, thus reducing redundant computation. Simultaneously, it fully preserves bidirectional matching information, providing a complete input foundation for subsequent construction of a bidirectional coverage matrix and identification of bidirectional high coverage, unidirectional partial coverage, and uncovered states, ensuring the completeness and reliability of subsequent consistency determination results.

[0034] Step S140: Based on the node information of the structural nodes, constraint element nodes and subordinate nodes of the constraint unit, calculate the six-dimensional association score of each clause pair in the clause pair list, and filter out each target clause pair that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score.

[0035] In one implementation, the six-dimensional association score includes a hierarchical path consistency score, a constraint element matching score, a resolution confidence score, an effectiveness time matching score, a numerical relationship consistency score, and a semantic similarity score. Among them, the hierarchical path consistency score is calculated based on the structural path carried by the structural nodes of the clause pair; the constraint element matching score is calculated based on the constraint element content, the start and end range of the original text characters, and the scope pointer carried by the constraint element nodes of the clause pair. The resolution confidence score is calculated based on the resolution confidence score carried by the constraint element nodes of the clause pair.

[0036] The validity time matching score is calculated based on the validity time constraint information carried by the validity time nodes of the clause pair.

[0037] The consistency score for numerical relationships is calculated based on the unit-normalized numerical range of numerical restriction elements in the constraint element nodes according to the clause, as well as the results of comparison operators. For example, "not exceeding 30 days" is represented as an upper limit constraint after unit normalization. ≤30 (days), "at least 60%" indicates a lower limit constraint. ≥0.60; "one month" can be converted to the corresponding time unit (such as 30 days) under explicit conversion rules before participating in interval comparisons. Intersection and inclusion operations are performed on intervals with similar constraints to identify relaxed timeframes, raised thresholds, or incompatible numerical conditions.

[0038] The semantic similarity score is calculated based on the semantic vectors bound to the structural nodes.

[0039] For example, the formula for calculating the six-dimensional correlation score is as follows: (1) , .

[0040] in, For query-side constraint units With candidate side constraint unit The total score of the six-dimensional correlation of the constituent clauses; The hierarchical path consistency score is calculated based on the structural path carried by the structural nodes of the clause pair. The constraint element matching score is calculated based on the constraint element content, the start and end range of the original text characters, and the scope pointer carried by the constraint element node of the clause pair. To determine the analytical confidence score, it is calculated based on the analytical confidence score carried by the constraint element nodes of the clause pair; The score for matching the effective time is calculated based on the effective time constraint information carried by the effective time nodes of the clauses. The numerical consistency score is calculated based on the numerical restriction constraint elements in the constraint element nodes of the clauses, after unit normalization of the numerical range and the results of comparison operator operations. The semantic similarity score is calculated based on the semantic vector bound to the structural node. This semantic vector is generated by encoding the node information carried by the structural node and the associated constraint element node, validity time node, and original text location node. , , , , , Assuming non-negative weights, it is assumed that the sum of all weights equals 1.

[0041] Understandably, this six-dimensional correlation score integrates hierarchical path consistency score, constraint element matching score, parsing confidence score, validity time matching score, numerical relationship consistency score, and semantic similarity score. It comprehensively evaluates the matching degree of clause pairs from multiple perspectives, including document hierarchical structure, constraint element content, parsing quality, timeliness constraints, numerical conditions, and textual semantics, overcoming the shortcomings of relying solely on semantic similarity. It can identify situations where the text is similar but there are substantial differences in hierarchy, constraint elements, validity time, and numerical conditions, while also taking into account the interference caused by OCR parsing defects in heterogeneous documents, thanks to the parsing confidence score. This multi-dimensional weighted fusion improves the accuracy of constraint unit matching, providing a reliable quantitative basis for subsequent selection of target clause pairs and construction of a two-way coverage matrix.

[0042] In one implementation, the step of filtering out the target clause pairs that point to the target candidate side constraint unit with the highest six-dimensional correlation score from the query-side constraint unit includes: a1: Summarize the six-dimensional association scores of each clause pair to form a candidate score set for the query-side constraint unit of each clause pair. The expression for the candidate score set can be... ,in, Represents query-side constraint unit The set of associated scores for all corresponding recall candidate units; This represents the query side constraint unit after tiered recall. The corresponding set of recall candidate units.

[0043] a2: Select the target candidate constraint unit corresponding to the highest six-dimensional correlation score from the candidate score set. This can be understood as selecting the candidate with the highest six-dimensional correlation score, representing the optimal overall fit, locking in the clause pair with the most comparison value, reducing invalid calculations, and supporting the construction of a two-way coverage matrix.

[0044] a3: If the highest six-dimensional correlation score is greater than or equal to the matching threshold, then a first set of directed covering edges or a second set of directed covering edges pointing from the query-side constraint unit to the target candidate-side constraint unit is established in the first or second direction corresponding to the query-side constraint unit.

[0045] Specifically, determining the matching threshold includes: statistically analyzing the score distribution characteristics (median, median difference) of the candidate score set of each clause pair's query-side constraint unit; calculating the clause length based on the page coordinates carried by the original text location node of the query-side constraint unit and normalizing it to obtain a length correction amount; then, obtaining a parsing quality correction amount by taking the complement of the parsing confidence carried by the constraint element node of the query-side constraint unit; finally, calculating the matching threshold of the query-side constraint unit based on the score distribution characteristics, length correction amount, and parsing quality correction amount. This approach can overcome the limitations of fixed thresholds, adapt to differences in score distribution, clause length, and parsing quality, reduce mismatches caused by heterogeneous policy documents, and improve the robustness of matching decisions.

[0046] For example, the formula for calculating the matching threshold is: .

[0047] in, For the first The adaptive matching threshold corresponding to each query-side constraint unit. For query-side constraint units The corresponding candidate score set is the set of six-dimensional correlation scores between the query unit and each recall candidate constraint unit. For the candidate score set the median; For the candidate score set The absolute median difference, ; For query-side constraint units The length correction is calculated by normalizing the page coordinates of the original text's positioning node to obtain the clause length. For query-side constraint units The parsing quality correction is obtained by taking the complement of the parsing confidence of the constraint element nodes; , , The preset weighting coefficients are determined from historical labeled samples; This is a truncation function; This limits the threshold to a safe range; This is the lower limit allowed by the threshold. This represents the upper limit allowed by the threshold. This refers to the original calculated threshold that has not been clamped at the upper or lower limits.

[0048] a4: Construct a bidirectional coverage matrix based on the first and second directed coverage edge sets of each query-side constraint unit. Each matrix element represents the bidirectional coverage of the clause pair formed by each query-side constraint unit and the target candidate-side constraint unit.

[0049] a5: If the highest six-dimensional correlation score is less than the matching threshold, the query-side constraint unit will be assigned to either the first uncovered set or the second uncovered set.

[0050] For example, the clauses in the first policy document 3. If the six-dimensional correlation scores of all recall candidates in the second policy document are lower than their adaptive matching threshold, then... 3 is assigned to the first uncovered set. If The resolution confidence level of 3 is high, and the system outputs "Second policy not covered". 3. Requires a structured conclusion; if If the confidence level of the parsing is low (e.g., there is an OCR recognition error), then the output will be "Suspected to be uncovered, manual review of the parsed text is required" to avoid forming a high-risk conclusion directly due to parsing errors.

[0051] a6: Determine the matching status of each clause pair based on the bidirectional coverage, the first uncovered set, the second uncovered set, and the retrieved constraint coverage auxiliary results.

[0052] a7: Filter out target clause pairs from the bidirectional coverage matrix that have a matching status of bidirectional high coverage and unidirectional partial coverage.

[0053] For example, the first policy clause 2. The second policy clause states that "the applicant should submit both a letter of commitment and an audit report." 2. "The applicant should submit a letter of commitment regarding creditworthiness." (In the first direction) → middle, 2 and 2. Establish directed coverage edges, but the coverage rate of constraint elements (audit reports not covered) is lower than the full coverage threshold; in the second direction... → middle, 2 can be 2. Full Coverage. The system combines the bidirectional coverage matrix with the constraint coverage auxiliary results, marks the clause pair as unidirectional partial coverage, and outputs the "Audit Report" as an uncovered constraint element.

[0054] like Figure 3 As shown in Figure 3, this application provides a schematic diagram of bidirectional coverage relationship detection, which is executed respectively. → , → The bidirectional hierarchical recall method combines six-dimensional correlation scores and adaptive matching thresholds to construct a bidirectional coverage matrix, identify matching states such as bidirectional high coverage, unidirectional partial coverage, no coverage, and new additions, and outputs a queue for manual review.

[0055] For example, the mathematical expressions for steps a3 to a7 are as follows: , , , .

[0056] in, For query-side constraint units The index of the candidate unit with the highest six-dimensional correlation score within the candidate set; For query-side constraint units The corresponding set of recall candidate units; Represents the first direction point to The set of directed covering edges, when the six-dimensional association score corresponding to the optimal candidate is greater than or equal to the adaptive matching threshold, then a directed covering edge is established from the query-side constraint unit to the optimal candidate-side constraint unit. This represents the first uncovered set. When the recall candidate unit set is empty, or the maximum six-dimensional association score in the candidate set is less than the adaptive matching threshold, the query-side constraint unit is included in the first uncovered set. For query unit with the best candidate unit The six-dimensional correlation score between them, the second direction point to Using the same computational logic, the set of directed covering edges can be obtained. and the second uncovered set The bidirectional covering matrix M is composed of... , Constructed together, matrix elements Record the bidirectional coverage status of the terms and conditions, with status types including bidirectional high coverage, unidirectional partial coverage, and no coverage. Indicating in the candidate unit set Find the subscript in the middle. This makes the query-side constraint unit With candidate side constraint unit Six-dimensional correlation score between When the maximum value is reached, the object corresponding to that index is the optimal candidate side constraint unit. .

[0057] For example, the calculation process for the constraint coverage auxiliary result is as follows: .

[0058] In the formula, To constrain feature nodes The corresponding constraint coverage rate is used to assist in determining the partial coverage status; This is the set of candidate units to be recalled after hierarchical recall of the current query constraint unit. The corresponding set of recall candidate units is denoted as With candidate score set Maintain consistency To recall any candidate unit in the candidate unit set, For indicator functions, Indicates candidate unit After executing constraint relationship calculations by invoking preset constraint rules, the constraint element nodes are satisfied. The constraint semantics carried by the constraint elements; if it is determined that the constraint element node satisfies the constraint semantics. If the indicator function outputs 1, the constraint element node is not satisfied. The output is 0; the constraint satisfies the same set of preset constraint rules used when identifying constraint conflict events.

[0059] Step S150: Combine the node information associated with each target clause with the preset constraint rules to determine the overall confidence level of the conflict events for each target clause pair.

[0060] In one implementation, firstly, according to preset constraint rules, constraint relationship operations are performed by combining the constraint element content carried by the constraint element node of each target clause pair, the original text character range, and the scope pointer, and the constraint conflict event set of each target clause pair is identified and output.

[0061] For example, in constraint relation operations, the obligation modality is parsed into polarity and intensity levels: affirmative obligations are divided into levels such as "must / shall," "need / should," "may," and "encourage" according to intensity; prohibitive obligations retain their negative polarity separately (such as "must not") and are not compared with permissive obligations only by word intensity. When the superior clause is a mandatory affirmative obligation and the clause to be compared becomes permissive or advocacy, it is identified as an event of weakened obligation intensity.

[0062] Understandably, it relies on preset constraint rules, combining the content of constraint elements, the original text character range, and the scope pointer to perform constraint relationship operations, not limited to surface text comparison. It can accurately locate substantive conflicts such as obligation strength, value, and exceptions, while anchoring the original text position, clarifying the scope of the constraint, and outputting a complete set of conflict events.

[0063] Next, for each constraint conflict event in the constraint conflict event set, the completeness of the evidence source of the original text location node, the multi-candidate ambiguity of the target clause pair, and the extraction confidence of the constraint element are determined by combining the node information associated with the target clause pair related to the constraint conflict event. The comprehensive confidence of the conflict event is calculated by combining the six-dimensional association score of the relevant target clause pair and the determinism of the retrieved preset constraint rules.

[0064] It is understandable that calculating the overall confidence level of a conflict event by integrating information from multiple dimensions does not directly establish a conflict. The completeness of the original source information is assessed through the completeness of the evidence sources; the degree of confusion in candidate matching is reflected by the ambiguity of multiple candidates; the confidence level of constraint element extraction is used to measure the quality of element parsing; and the certainty of the clauses' six-dimensional correlation scores and constraint rules are integrated. This effectively avoids misjudgments caused by document parsing defects, matching ambiguities, and missing evidence, and distinguishes between definitive conflicts and questionable situations.

[0065] Wherein, the determinism of the constraint rule is the preset rule score corresponding to the conflict type; the extraction confidence of the constraint element is obtained by aggregating the analytical confidence of the constraint element nodes participating in the conflict determination; the completeness of the evidence source of the constraint element is determined based on the completeness of the source file identifier, page number, and page coordinates carried by the original text positioning node associated with the constraint element participating in the determination; the multi-candidate ambiguity of the target clause pair is determined based on the difference in the six-dimensional association scores between the best and second-best candidates in the candidate score set corresponding to the query-side constraint unit.

[0066] For example, the calculation process for the overall confidence level of a conflict event is as follows: , .

[0067] in, To constrain the overall confidence level of conflict events, when When, output the structured conclusion of constraint conflict, when At that time, output the structured conclusions that require manual review; The six-dimensional correlation score corresponding to the current target clause is the output result of formula (1); To ensure the determinism of the constraint rules, a preset rule score corresponding to the conflict type is taken; The confidence level is extracted for the constraint elements, which is obtained by aggregating the analytical confidence levels carried by the constraint element nodes participating in the conflict determination. The completeness of the evidence source is calculated based on the completeness of the source file identifier, page number, and page coordinates carried by the original text location nodes associated with the constraint elements involved in the determination; To address multiple candidate ambiguity, the candidate score set corresponding to the query-side constraint unit is used. Calculate the difference in six-dimensional correlation scores between the best and second-best candidates; to Preset non-negative weighting coefficients; The Sigmoid normalization function outputs values ​​in the range of... ; Input for the Sigmoid function; For the natural constant An exponential function with base 0; To preset the review threshold, the overall confidence level of the conflict events is compared with the threshold to output the final structured conclusion.

[0068] like Figure 4 This application provides a schematic diagram for calculating the comprehensive confidence level of conflict events. Based on the six-dimensional correlation score, the comprehensive confidence level of conflict events is calculated by integrating rule certainty, element extraction confidence, evidence completeness, and multi-candidate ambiguity, and the conflict event is distinguished as either a constraint conflict or one awaiting manual review.

[0069] Step S160: Compare the overall confidence level with the preset review threshold, and output a structured conclusion that represents the constraint conflict of heterogeneous policy documents or that needs to be manually reviewed.

[0070] Specifically, if the overall confidence level is higher than the preset review threshold, a structured conclusion representing the constraint conflict between the first policy document set and the second policy document set is output; if the overall confidence level is lower than or equal to the preset review threshold, a structured conclusion representing the need for manual review is output.

[0071] In one application scenario, let's define the first policy clause. The first clause states that "the project's competent authority should complete the review within 30 days," and the second policy clause... 1. The project's competent authority can complete the review within 60 days. The two statements have high semantic vector similarity and their subjects, behaviors, and objects can be aligned, but the constraint extraction results show... The obligation mode of 1 is mandatory, with a maximum time limit of 30 days. The obligation modality of 1 is permission, with a maximum time limit of 60 days. After constraint relation calculation, the system outputs two constraint conflict events: "weakened obligation strength" and "relaxed processing time limit." Since their combined confidence level is higher than the review threshold, a structured conclusion representing the constraint conflict is output, and the chain of evidence includes... 1 and 1. The original text, source file identifier, page number, structure path, relevant constraint fields, and judgment rules.

[0072] Understandably, the above technical solution involves, firstly, parsing heterogeneous policy documents to construct a hierarchical constraint graph, dividing it into a set of constraint units and a set of subordinate nodes. This approach is compatible with formats such as Word, PDF, and scanned documents, unifying the document hierarchy, constraint elements, effective time, and original text location information into a graph structure. This avoids the problem of inconsistent clause boundaries caused by different document layouts and OCR recognition defects, providing a structured foundation for subsequent comparisons. Secondly, a bidirectional hierarchical recall and pairing filtering method is used to obtain a list of clause pairs. Unlike traditional unidirectional retrieval, this method can simultaneously detect policy omissions, partial coverage, and newly added requirements. The hierarchical recall method includes structural path filtering, sparse features, and semantic vector recall, significantly compressing the size of clause pairs requiring in-depth computation. This limits deep analysis to a limited number of candidates, reduces CPU computational overhead, and avoids the performance loss caused by full pairwise comparisons. Furthermore, by utilizing six-dimensional correlation scores to integrate hierarchical paths, constraint elements, resolution confidence, effectiveness time, numerical relationships, and semantic multi-dimensional information, it no longer relies solely on textual semantic similarity. It can identify situations where texts have similar wording but are substantially inconsistent in terms of obligations, values, and timeliness. By combining candidate score distribution and other information to select the optimal target clause pair, it can distinguish between two-way high coverage, one-way partial coverage, and no coverage matching states, reducing false matches caused by fixed thresholds. Finally, it integrates multi-type node information to calculate the comprehensive confidence of conflict events, correlates with original text location, extracts confidence and candidate ambiguity, and then compares it with the review threshold to output structured conclusions. For scenarios with poor resolution quality and incomplete evidence, it outputs manual review prompts, forming a traceable evidence chain, reducing erroneous judgments caused by OCR resolution errors, and improving the reliability and interpretability of policy conflict detection.

[0073] In some embodiments, the heterogeneous policy documents include a first policy document and a second policy document, and the method further includes: Step S210: If the version of the first policy document is updated, then the updated first and second policy documents are parsed into a new hierarchical constraint diagram; Step S220: Calculate the content fingerprint of each node in the new hierarchical constraint graph and the hierarchical constraint graph respectively; Step S230: Compare the content fingerprint of each node to determine the set of newly added nodes, deleted nodes, and changed nodes.

[0074] Specifically, the set of changing nodes The expression is , Represents the set of newly added nodes. This indicates the deletion of a set of nodes. Represents the set of changed nodes. This represents the set union operation.

[0075] Step S240: For target nodes that have not undergone fingerprint changes, reuse the cached target node semantic vector, the topological relationship of the target node in the hierarchical constraint graph, and the recall candidate unit set corresponding to the target node.

[0076] Step S250: Determine the set of affected constraint units based on the set of changed nodes. The set of affected constraint units includes constraint units that have changed, and constraint units whose cached recall candidate unit set contains the changed nodes.

[0077] Step S260: For nodes that have not undergone fingerprint changes and do not belong to the set of affected constraint units, reuse the cached node semantic vector, the topological relationship of the node in the hierarchical constraint graph, and the recall candidate unit set corresponding to the node; only for the set of changed nodes and the set of affected constraint units, re-execute the hierarchical recall, pairing screening, six-dimensional association score calculation, bidirectional coverage relationship detection, conflict event comprehensive confidence calculation, and structured conclusion output steps.

[0078] Among them, bidirectional coverage relationship detection refers to steps a1 to a7 in the above embodiments.

[0079] Understandably, this embodiment introduces an incremental update mechanism based on node content fingerprints for policy document version updates, avoiding the need to repeatedly perform the complete detection process on all constraint nodes after policy revision. When the first policy document is updated, the system constructs a new hierarchical constraint graph. By comparing the content fingerprints of each node in the old and new hierarchical constraint graphs, it identifies the set of added, deleted, and modified nodes. For nodes whose fingerprints have not changed, the system directly reuses cached semantic vectors, graph topology relationships, and recall candidate unit sets, without repeating parsing, encoding, and recall operations. Only the steps of re-hierarchical recall, pairing and filtering, six-dimensional association score calculation, bidirectional coverage relationship detection, conflict event comprehensive confidence calculation, and structured conclusion output are performed on the changed node set. Unchanged nodes directly call existing cached data to participate in the calculation. Unlike the full-processing mode that re-compares all clauses after a document update, this incremental processing method significantly reduces repetitive vector encoding, hierarchical recall, constraint operations, and other computational tasks, significantly reducing CPU computing power consumption and time overhead, and adapting to business scenarios with frequent policy revisions and iterations.

[0080] Meanwhile, cache reuse does not disrupt the topological logic of the original hierarchical constraint graph or the intermediate results of historical comparisons, ensuring that the comparison conclusions of the unmodified parts remain valid, and only performing consistency checks on the changed parts. This ensures the accuracy of conflict detection results after version updates, while also taking into account system processing performance, and retains the complete historical comparison context, providing support for the difference analysis of policy version iterations, and effectively solving the problem of low efficiency of full comparison in scenarios with frequent policy updates.

[0081] Please refer to Figure 5, which is a schematic diagram of CPU environment hierarchical recall and incremental update processing provided in the embodiment of this application; hierarchical recall filters candidates step by step to limit the number of deep analysis terms; when the version is updated, the changed nodes are identified by the node content fingerprint, the unchanged nodes reuse the cache, and only the changed nodes are re-performed for consistency detection, reducing computing power overhead.

[0082] In some embodiments, step S110: obtaining the heterogeneous policy documents to be compared and parsing them to obtain a hierarchical constraint graph includes the following steps: Step S111: Obtain the first and second policy documents, which are heterogeneous policy documents to be compared, and parse them into a set of text units. Each text unit carries a unique text unit identifier, normalized text content, original page number, page coordinates, page style attributes, source document identifier, and text unit parsing confidence.

[0083] Specifically, heterogeneous policy documents include, but are not limited to, Word, PDF, scanned images, and HTML files. For Word files, paragraph, table, style, and numbering attributes are read; for PDF files, character coordinates, text blocks, page numbers, and font attributes are read; for scanned images, layout analysis and OCR are performed while retaining character or line-level recognition confidence; for HTML files, headings, paragraphs, list, and table tags are parsed; the parsing results are uniformly represented as text units. U .

[0084] Text unit U The field structure is defined as follows: In the formula, A unique identifier for a text unit; The standardized text content; Page numbers are from the original text. For page coordinates; For layout style attributes; For source file identification; Analyze the confidence level for each text unit.

[0085] Step S112: Generate a text stream from the set of text units based on the original page number and page coordinates. The text stream is an ordered sequence of text units arranged in the reading order.

[0086] Step S113: Traverse the gaps between adjacent text units in the text stream. When there are boundary clue features in the preceding and / or following text units associated with the gap, create a clause boundary candidate in the gap and calculate the boundary score of the boundary candidate based on the boundary clue features. Otherwise, do not create a boundary candidate, forming a clause boundary candidate set. The boundary clue features are obtained by comparing the text unit with a predetermined boundary triggering condition.

[0087] Specifically, the boundary clue features include: numbering patterns (such as "Chapter ×", "Article ×", "I", "(I)", "1.", etc.), layout attributes (font size, bolding, paragraph spacing, indentation, line breaks), table of contents items, and semantic continuity of adjacent text. When a text unit hits any of the above clues, a clause boundary candidate is created in that gap.

[0088] Step S114: Calculate the comprehensive boundary score of each boundary candidate based on the boundary score and preset weights to form a boundary candidate set.

[0089] Step S115: Under the premise of satisfying the preset text level constraints, dynamic programming or constraint search is used to solve for the clause boundary sequence with the largest total boundary comprehensive score in the full boundary candidate set.

[0090] Specifically, all boundary candidates are sorted according to their gap positions in the text stream; the text stream position and the current text level are used as the dynamic programming state; the boundary candidates at each gap position are traversed, and when a boundary candidate is selected, it is checked whether the level jump after selecting the candidate meets the preset text level constraints; if the level jump is valid, the cumulative boundary comprehensive score in the corresponding state is updated and the boundary selection path is recorded; if the level jump is invalid, the selection branch is discarded; after the traversal is completed, the boundary selection path with the largest cumulative boundary comprehensive score is selected to obtain the optimal boundary sequence.

[0091] Understandably, by combining numbering, layout, semantics, and OCR quality to generate boundary candidates, and by using dynamic programming with hierarchical constraints to solve for the optimal boundary sequence, problems such as page breaks and corrupted numbering can be handled, thus improving the accuracy of clause hierarchy recovery.

[0092] Step S116: Generate the structural path corresponding to each clause based on the clause boundary sequence, and aggregate the text unit-level parsing confidence of each text unit that makes up the clause along the structural path to obtain the clause-level parsing confidence.

[0093] Specifically, for example, after a clause is restored by the boundary sequence, the structural path is "Chapter 2 / Article 5 / Paragraph 2 / Item (3)". This structural path is associated level by level along the document node → chapter node → section node → clause node → item node.

[0094] For example, the mathematical expressions for steps S113 to S116 are as follows: , .

[0095] in, Candidates for the boundary When dynamically programming to solve for the optimal boundary sequence, the sum of the boundary comprehensive scores of the selected boundary candidates is maximized under the premise of satisfying the preset text level constraints. The score for the numbering pattern clue is obtained by comparing the text unit with the predetermined numbering pattern. The score for layout clues is determined based on layout features such as indentation, font size, and line breaks. This is a score for the legality of the numbering sequence, used to verify whether there are skipped or disordered numbering sequences in the clause numbering sequence; The context continuity score represents the degree of semantic continuity between text units before and after the boundary candidate. The OCR recognition quality score is set, and the score will be reduced when there is a case of OCR recognition of a corrupted number. For example, when the OCR recognizes "Article 10" as "Article 10th", the system combines the policy number dictionary, adjacent article numbers and edit distance to generate error correction candidates, and reduces the OCR recognition quality score of the boundary candidate accordingly. At the same time, the parsing confidence of the text unit is marked as low confidence. Assuming non-negative weights, the sum of all weights equals 1.

[0096] Dynamic programming execution logic: Using the text flow position and the current text level as the dynamic programming state, iterates through all boundary candidates, checks whether the level jump is legal, retains the boundary selection path with the largest cumulative boundary comprehensive score as the optimal clause boundary sequence, generates the clause structure path based on the optimal clause boundary sequence, and aggregates the text unit-level parsing confidence to obtain the clause-level parsing confidence.

[0097] Understandably, by aggregating text unit-level parsing confidence along the hierarchical structure path, clause-level confidence is obtained, local OCR defects are quantified into clause credibility indicators, and parsing quality is transmitted downstream, providing a basis for triggering manual review of low-quality samples.

[0098] Step S117: Extract multiple types of constraint elements from each text unit of each clause, record the start and end character range and scope pointer of each type of constraint element in the heterogeneous policy document, and use the clause-level resolution confidence as the resolution confidence of each type of constraint element to form a set of constraint elements. Each type of constraint element carries the start and end character range, scope pointer and resolution confidence.

[0099] Specifically, for example, the clause "The county-level finance department shall complete the review of the subsidy funds within 30 days, except where the materials are incomplete" is extracted and represented as: Subject = County-level finance department, Action = Complete review, Object = Subsidy funds, Obligation modality = Should, Polarity = Affirmative, Numerical limit = Time limit ≤ 30 days, Exception = Incomplete materials, Structural path = Corresponding clause path; Each constraint element records its start and end character range in the original text, and the exception element also records that the scope pointer points to "Complete review of subsidy funds".

[0100] It is understandable that it extracts constraint elements such as subject, behavior, obligation modality, value, and exception, while recording the original character range and scope pointer, preserving the constraint position and modification relationship, and supporting numerical calculation, obligation strength comparison and exception range comparison.

[0101] Step S118: Generate document nodes for the first and second policy documents respectively, generate structure nodes for each text unit entering the comparison, and generate corresponding constraint element nodes for the constraint elements extracted from the original text of the text unit. Generate effective time nodes based on the information such as the implementation date and the start and end dates of the validity period clearly recorded in the original text of the policy document. If the corresponding information is not recorded in the original text of the policy document, do not generate the effective time node. Generate original text positioning nodes based on the document identifier, page number and page coordinates of the text unit in the original text, and summarize the node set.

[0102] Step S119: Construct containment edges between file nodes and their subordinate structural nodes; construct constraint, applicability, exception, and basis edges between structural nodes and their corresponding constraint element nodes; construct association edges between file nodes or structural nodes and validity time nodes; establish association edges between file nodes, structural nodes, and constraint element nodes and the original text location nodes respectively; summarize the edge set; and construct a hierarchical constraint graph based on the node set and edge set.

[0103] Specifically, such as Figure 2 The hierarchical constraint diagram is formally denoted as: in It is the complete set of nodes in the hierarchical constraint diagram, including file nodes, structural nodes and constraint element nodes that form constraint units, effectiveness time nodes that contain subordinate nodes that form constraint units, and original text location nodes, etc. The set of edges is the entire set; the set of nodes is partitioned according to the file nodes corresponding to the heterogeneous policy documents to obtain the first set of constraint units. Second constraint unit set in, This represents the total number of constraint elements within the first set of constraint elements. This represents the total number of constraint units within the second constraint unit set.

[0104] The following three specific scenarios illustrate the complete execution process of the above method. Let the first set of constraint units be... Derived from higher-level policy documents, the second set of constraint units It originates from local policy documents.

[0105] Scenario 1: Weakening of Obligation Intensity and Extension of Time Limit. Let... Medium constraint element 1. The corresponding clause states that "the project's competent authority shall complete the review within 30 days." Medium constraint element 1. The corresponding clause states, "The project's competent authority may complete the review within 60 days." Following a tiered recall process, 1 and 1. Candidate clause pairs were formed, and their semantic similarity score and hierarchical path consistency score were both high in the six-dimensional association scores, indicating matching of subject, behavior, and object constraint elements; however, the constraint element extraction results showed that... The obligation modality of 1 is mandatory (shall), and the numerical limit is a maximum time limit of ≤30 days. The obligation modality of 1 is permitted (may), and the numerical limitation is a time limit ≤ 60 days. After the system performs constraint relationship calculations, it identifies two constraint conflict events: "weakening of obligation strength" and "relaxation of processing time limit." The combined confidence level of both events is higher than the preset review threshold; therefore, a structured conclusion representing the constraint conflict is output, and the evidence chain includes... 1 and 1. The original text, source file identifier, page number, structural path, relevant constraint element fields, unit normalization results, and judgment rules.

[0106] Scenario 2: One-way partial coverage. Let... Medium constraint element 2. The corresponding clause states that "the applicant should submit both a letter of commitment and an audit report." Medium constraint element 2. The corresponding clause states, "The applicant shall submit a letter of commitment regarding creditworthiness." In the first direction... → middle, The optimal candidate for 2 is 2. Its six-dimensional correlation score is greater than or equal to The adaptive matching threshold of 2 is thus established. 2 points 2. Directed covering edges; however, the constraint coverage auxiliary results show that the constraint element "Audit Report" was not covered. 2. Coverage, with constraint coverage rate below the full coverage threshold. In the second direction. → middle, The optimal candidate for 2 is 2, and All constraints of 2 were 2. Coverage, Establishment 2 points 2. Directed covering edge. The system combines the bidirectional covering matrix and the constraint coverage auxiliary result to determine the matching status of the clause pair as unidirectional partial coverage, and outputs "Audit Report" as an uncovered constraint element along with the structured conclusion.

[0107] Scenario 3: Coverage Requirements and Low-Confidence Verification. Let... Medium constraint element 3 In All recall candidates had a six-dimensional correlation score lower than An adaptive matching threshold of 3, then 3 is assigned to the first uncovered set. If The system outputs "Second policy document not covered" when the resolution confidence level of 3 is high (e.g., it originates from a Word document, and the numbering and layout are complete). 3. Requires a structured conclusion; if If the resolution confidence of 3 is low (e.g., it originates from a scanned image and contains OCR recognition errors), then the overall confidence of its conflict events is lower than or equal to the preset review threshold. The system outputs a structured conclusion of "awaiting manual review" and marks the field that caused the low confidence as the resolution confidence, thereby avoiding the direct formation of high-risk conclusions due to OCR errors.

[0108] like Figure 6 As shown, this application also provides a bidirectional consistency detection system for heterogeneous policy documents, including: The hierarchical constraint graph construction module is used to obtain heterogeneous policy documents to be compared and parse them to obtain hierarchical constraint graphs. The constraint graph partitioning module is used to partition the hierarchical constraint graph into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes based on the file nodes corresponding to the heterogeneous policy documents. The bidirectional hierarchical recall module is used to perform hierarchical recall and pair filtering to obtain a list of clause pairs by pointing along a first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along a second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side. The score and target pair filtering module is used to calculate the six-dimensional association score of each clause pair in the clause pair list based on the node information of the structural nodes, constraint element nodes and subordinate nodes of the constraint unit, and to filter out each target clause pair that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score. The confidence calculation module is used to determine the overall confidence of conflict events for each target clause pair by combining the associated node information and preset constraint rules. The structured conclusion output module is used to compare the overall confidence with the preset review threshold and output a structured conclusion that represents the constraint conflict of heterogeneous policy documents or is pending manual review.

[0109] In some embodiments, the auxiliary nodes include an effectiveness time node and an original text location node, and the bidirectional hierarchical recall module includes a hierarchical recall submodule. The hierarchical recall submodule is used to perform structural path filtering, sparse feature recall, and semantic vector recall. Structural path filtering refers to removing candidate constraint units whose hierarchical context is clearly irrelevant based on the structural paths carried by the structural nodes of the query-side constraint units and candidate-side constraint units. Sparse feature recall refers to filtering the candidate-side constraint units retained after structural path filtering based on the constraint elements representing policy themes and keywords carried by the constraint element nodes of the query-side constraint units and the effectiveness time-related constraint information carried by the effectiveness time nodes, removing candidate-side constraint units with mismatched themes, no keyword overlap, or invalid effectiveness time. Semantic vector recall refers to retrieving the semantic vectors bound to the structural nodes of the query-side constraint units and the candidate-side constraint units retained after sparse feature recall, performing quantization processing, using the quantized semantic vectors to perform similarity retrieval to obtain semantic similarity, sorting them in descending order of similarity, and selecting the top... The candidate side constraint units form a recall candidate unit set, wherein the semantic vector is generated by encoding based on the node information carried by the respective structural nodes and the associated constraint element nodes, effectiveness time nodes and original text positioning nodes.

[0110] In some embodiments, the bidirectional hierarchical recall module further includes a clause pair list generation submodule; the clause pair list generation submodule is used to, in the first direction, treat each first constraint unit as a query-side constraint unit, and pair each second candidate unit obtained by hierarchical recall of the first constraint unit with the first constraint unit to form a first clause pair list; in the second direction, treat each second constraint unit as a query-side constraint unit, and pair each first candidate unit obtained by hierarchical recall of the second constraint unit with the second constraint unit to form a second clause pair list; the clause pairs in the second clause pair list are normalized in the order of first constraint unit first, second constraint unit last, and merged with the first clause pair list to remove duplicates, thus obtaining a clause pair list.

[0111] In some embodiments, the score and target pair filtering module includes a bidirectional coverage matrix construction submodule; The bidirectional coverage matrix construction submodule is used to summarize the six-dimensional association scores of each clause pair to form a candidate score set for the query-side constraint unit of each clause pair; select the target candidate-side constraint unit corresponding to the highest six-dimensional association score from the candidate score set; if the highest six-dimensional association score is greater than or equal to the matching threshold, establish a first or second directed coverage edge set pointing from the query-side constraint unit to the target candidate-side constraint unit in the first or second direction corresponding to the query-side constraint unit; construct a bidirectional coverage matrix based on the first and second directed coverage edge sets of each query-side constraint unit, where each matrix element represents the bidirectional coverage of the clause pair formed by each query-side constraint unit and the target candidate-side constraint unit; if the highest six-dimensional association score is less than the matching threshold, classify the query-side constraint unit into the first or second uncovered set; determine the matching status of each clause pair based on the bidirectional coverage status, the first uncovered set, the second uncovered set, and the retrieved constraint coverage rate auxiliary result; and filter out each target clause pair with a matching status of bidirectional high coverage and unidirectional partial coverage from the bidirectional coverage matrix.

[0112] In some embodiments, the auxiliary node further includes a source text location node, and the score and target pair filtering module further includes a six-dimensional score calculation submodule; the six-dimensional score calculation submodule is used to calculate a six-dimensional association score, which includes a hierarchical path consistency score, a constraint element matching score, a resolution confidence score, an effectiveness time matching score, a numerical relationship consistency score, and a semantic similarity score; wherein, the hierarchical path consistency score is calculated based on the structural path carried by the structural node of the clause pair; the constraint element matching score is calculated based on the constraint element content, source text character start and end range, and scope pointer carried by the constraint element node of the clause pair; the resolution confidence score is calculated based on the resolution confidence carried by the constraint element node of the clause pair; the effectiveness time matching score is calculated based on the effectiveness time type constraint information carried by the effectiveness time node of the clause pair; the numerical relationship consistency score is calculated based on the numerical restriction type constraint elements in the constraint element node of the clause pair, the numerical range after unit normalization, and the comparison operator operation result; and the semantic similarity score is calculated based on the semantic vector bound to the structural node.

[0113] In some embodiments, the score and target pair filtering module further includes a matching threshold determination submodule; the matching threshold determination submodule is used to statistically analyze the score distribution characteristics of the candidate score set of each clause pair's query-side constraint unit; calculate the clause length based on the page coordinates carried by the original text location node of the query-side constraint unit and normalize it to obtain a length correction amount; obtain a parsing quality correction amount by taking the complement of the parsing confidence carried by the constraint element node of the query-side constraint unit; and calculate the matching threshold of the query-side constraint unit based on the score distribution characteristics, length correction amount, and parsing quality correction amount.

[0114] In some embodiments, the confidence calculation module includes a conflict event identification submodule and a comprehensive confidence calculation submodule. The conflict event identification submodule is used to perform constraint relationship operations based on preset constraint rules, combining the constraint element content carried by the constraint element node of each target clause pair, the original text character range, and the scope pointer, to identify and output a set of constraint conflict events for each target clause pair. The comprehensive confidence calculation submodule is used to determine the evidence source completeness of the original text location node, the multi-candidate ambiguity of the target clause pair, and the extraction confidence of the constraint element for each constraint conflict event in the set of constraint conflict events, combining the node information associated with the target clause pair related to the constraint conflict event, and to calculate the comprehensive confidence of the conflict event by combining the six-dimensional association score of the relevant target clause pair and the determinism of the retrieved preset constraint rules. Wherein, the determinism of the constraint rule is the preset rule score corresponding to the conflict type; the extraction confidence of the constraint element is obtained by aggregating the analytical confidence of the constraint element nodes participating in the conflict determination; the completeness of the evidence source of the constraint element is determined based on the completeness of the source file identifier, page number, and page coordinates carried by the original text positioning node associated with the constraint element participating in the determination; the multi-candidate ambiguity of the target clause pair is determined based on the difference in the six-dimensional association scores between the best and second-best candidates in the candidate score set corresponding to the query-side constraint unit.

[0115] In some embodiments, an incremental update module is further included; the heterogeneous policy documents include a first policy document and a second policy document, and the incremental update module is used to parse the updated first policy document and the second policy document into a new hierarchical constraint graph when the version of the first policy document is updated; calculate the content fingerprint of each node in the new hierarchical constraint graph and the hierarchical constraint graph respectively; compare the content fingerprint of each node node by node to determine the set of newly added nodes, deleted nodes, and changed nodes; for nodes whose fingerprints have not changed, reuse the cached node semantic vector, the topological relationship of the node in the hierarchical constraint graph, and the recall candidate unit set corresponding to the node; and according to the changed nodes... The set of affected constraint units is determined, which includes constraint units that have changed, and constraint units whose cached recall candidate unit set contains changed nodes. For nodes that have not changed their fingerprints and do not belong to the affected constraint unit set, the cached node semantic vector, the node's topological relationship in the hierarchical constraint graph, and the node's corresponding recall candidate unit set are reused. The hierarchical recall, pairing screening, six-dimensional association score calculation, bidirectional coverage relationship detection, conflict event comprehensive confidence calculation, and structured conclusion output steps are re-executed only for the changed node set and the affected constraint unit set.

[0116] In some embodiments, the structured conclusion output module includes a review threshold comparison submodule; the review threshold comparison submodule is used to output a structured conclusion representing the constraint conflict between the first policy document set and the second policy document set if the overall confidence level is higher than a preset review threshold; and to output a structured conclusion representing the need for manual review if the overall confidence level is lower than or equal to the preset review threshold.

[0117] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting bidirectional consistency of heterogeneous policy documents, characterized in that, The method includes: Obtain the heterogeneous policy documents to be compared and parse them to obtain the hierarchical constraint graph; Based on the document nodes corresponding to the heterogeneous policy documents, the hierarchical constraint diagram is divided into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes; Hierarchical recall and pairing filtering are performed along a first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along a second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side to obtain a list of clause pairs. Based on the node information of the structural nodes, constraint element nodes, and subordinate nodes of the constraint unit, the six-dimensional association score of each clause pair in the clause pair list is calculated, and the target clause pairs that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score are selected. By combining the node information and preset constraint rules associated with each target clause, the overall confidence level of conflict events for each target clause is determined. By comparing the overall confidence level with the preset review threshold, a structured conclusion is output that represents the constraint conflict of heterogeneous policy documents or those that need to be manually reviewed.

2. The method according to claim 1, characterized in that, The auxiliary nodes include effectiveness time nodes and original text location nodes, and the hierarchical recall includes structural path filtering, sparse feature recall and semantic vector recall. The structural path filtering refers to eliminating candidate side constraint units whose hierarchical context is obviously irrelevant, based on the structural paths carried by the structural nodes of the query side constraint unit and the candidate side constraint unit. The sparse feature recall refers to filtering the candidate side constraint units retained after structural path filtering based on the constraint elements representing policy themes and keywords carried by the constraint element nodes of the query side constraint units and the validity time type constraint information carried by the validity time nodes, and removing candidate side constraint units that do not match the theme, have no overlap in keywords, or have expired validity time. The semantic vector recall refers to retrieving the semantic vectors bound to the structural nodes of the query-side constraint unit and the candidate-side constraint units retained through sparse feature recall, performing quantization processing, using the quantized semantic vectors to perform similarity retrieval to obtain semantic similarity, sorting them in descending order of similarity, and selecting the top... The candidate side constraint units form a recall candidate unit set, wherein the semantic vector is generated by encoding based on the node information carried by the respective structural nodes and the associated constraint element nodes, effectiveness time nodes and original text positioning nodes.

3. The method according to claim 1, characterized in that, The pairing filter yields a list of terms pairs, including: For the first direction, each first constraint unit is used as a query-side constraint unit, and each second candidate unit obtained by hierarchical recall of the first constraint unit is paired with the first constraint unit to form a first clause pair list; For the second direction, each second constraint unit is used as a query-side constraint unit, and each first candidate unit obtained by hierarchical recall of the second constraint unit is paired with the second constraint unit to form a second clause pair list; The clause pairs in the second clause pair list are normalized in the order of first constraint unit first and second constraint unit last, and then merged with the first clause pair list to remove duplicates, resulting in the clause pair list.

4. The method according to claim 1, characterized in that, The target clause pairs that filter out the target candidate side constraint units that point to the target candidate side constraint unit with the highest six-dimensional correlation score include: The six-dimensional correlation scores of each clause pair are summarized to form a set of candidate scores for the query-side constraint units of each clause pair. Select the target candidate side constraint unit corresponding to the highest six-dimensional correlation score from the candidate score set; If the highest six-dimensional correlation score is greater than or equal to the matching threshold, then a first set of directed covering edges or a second set of directed covering edges pointing from the query-side constraint unit to the target candidate-side constraint unit is established in the first or second direction corresponding to the query-side constraint unit. A bidirectional coverage matrix is ​​constructed based on the first and second directed coverage edge sets of each query-side constraint unit. Each matrix element represents the bidirectional coverage of the clause pair formed by each query-side constraint unit and the target candidate-side constraint unit. If the highest six-dimensional correlation score is less than the matching threshold, the query-side constraint unit will be assigned to either the first uncovered set or the second uncovered set. The matching status of each clause pair is determined based on the bidirectional coverage, the first uncovered set, the second uncovered set, and the retrieved constraint coverage auxiliary results. Select target clause pairs from the two-way coverage matrix that match the two-way high coverage and one-way partial coverage status.

5. The method according to claim 4, characterized in that, The auxiliary nodes also include original text location nodes, and the six-dimensional association scores include hierarchical path consistency score, constraint element matching score, parsing confidence score, effectiveness time matching score, numerical relationship consistency score, and semantic similarity score. Among them, the hierarchical path consistency score is calculated based on the structural path carried by the structural nodes of the clause pair; the constraint element matching score is calculated based on the constraint element content, the start and end range of the original text characters, and the scope pointer carried by the constraint element nodes of the clause pair. The analytical confidence score is calculated based on the analytical confidence score carried by the constraint element nodes of the clause pair; The effectiveness time matching score is calculated based on the effectiveness time constraint information carried by the effectiveness time nodes of the clause pair; The numerical relationship consistency score is calculated based on the numerical restriction constraint elements in the constraint element nodes of the clauses, after unit normalization of the numerical range and the results of comparison operator operations. The semantic similarity score is calculated based on the semantic vectors bound to the structural nodes.

6. The method according to claim 4, characterized in that, The determination of the matching threshold includes: Statistically analyze the score distribution characteristics of the candidate score set for each clause pair's query-side constraint unit; The length of the clause is calculated based on the page coordinates carried by the original text location node of the query-side constraint unit and normalized to obtain the length correction amount. The parsing quality correction is obtained by taking the complement of the parsing confidence carried by the constraint element node of the query-side constraint unit. The matching threshold of the query-side constraint unit is calculated based on the score distribution characteristics, length correction amount, and parsing quality correction amount.

7. The method according to claim 1, characterized in that, The determination of the overall confidence level of conflict events for each pair of target clauses, by combining the associated node information and preset constraint rules, includes: Based on the preset constraint rules, the constraint relationship operation is performed by combining the constraint element content, original character range and scope pointer carried by the constraint element node of each target clause pair, and the constraint conflict event set of each target clause pair is identified and output. For each constraint conflict event in the constraint conflict event set, the evidence source completeness of the original text location node, the multi-candidate ambiguity of the target clause pair, and the extraction confidence of the constraint element are determined by combining the node information associated with the target clause pair related to the constraint conflict event. The comprehensive confidence of the conflict event is calculated by combining the six-dimensional association score of the relevant target clause pair and the determinism of the retrieved preset constraint rules. Wherein, the determinism of the constraint rule is the preset rule score corresponding to the conflict type; the extraction confidence of the constraint element is obtained by aggregating the analytical confidence of the constraint element nodes participating in the conflict determination; the completeness of the evidence source of the constraint element is determined based on the completeness of the source file identifier, page number, and page coordinates carried by the original text positioning node associated with the constraint element participating in the determination; the multi-candidate ambiguity of the target clause pair is determined based on the difference in the six-dimensional association scores between the best and second-best candidates in the candidate score set corresponding to the query-side constraint unit.

8. The method according to claim 1, characterized in that, The heterogeneous policy documents include a first policy document and a second policy document, and the method further includes: If the version of the first policy document is updated, the updated first and second policy documents will be parsed into a new level constraint graph; Calculate the content fingerprint of each node in the new hierarchical constraint graph and the hierarchical constraint graph respectively; Compare the content fingerprints of each node to determine the set of newly added nodes, deleted nodes, and changed nodes. For nodes whose fingerprints have not changed, the cached node semantic vectors, the topological relationships of the nodes in the hierarchical constraint graph, and the corresponding recall candidate unit set are reused. Based on the set of changed nodes, a set of affected constraint units is determined, which includes constraint units that have changed, and constraint units whose cached recall candidate unit set contains the changed nodes. For nodes that have not undergone fingerprint changes and do not belong to the set of affected constraint units, the cached node semantic vector, the topological relationship of the node in the hierarchical constraint graph, and the recall candidate unit set corresponding to the node are reused. The hierarchical recall, pairing screening, six-dimensional association score calculation, bidirectional coverage relationship detection, conflict event comprehensive confidence calculation, and structured conclusion output steps are re-executed only for the set of changed nodes and the set of affected constraint units.

9. The method according to claim 1, characterized in that, The process of comparing the comprehensive confidence level with a preset review threshold and outputting structured conclusions that characterize the constraint conflicts of heterogeneous policy documents or require manual review includes: If the overall confidence level is higher than the preset review threshold, a structured conclusion representing the constraint conflict between the first set of policy documents and the second set of policy documents will be output. If the overall confidence level is lower than or equal to the preset review threshold, a structured conclusion representing the need for manual review will be output.

10. A bidirectional consistency detection system for heterogeneous policy documents, characterized in that, The system includes: The hierarchical constraint graph construction module is used to obtain heterogeneous policy documents to be compared and parse them to obtain hierarchical constraint graphs. The constraint graph partitioning module is used to partition the hierarchical constraint graph into a first constraint unit set, a second constraint unit set, and a set of their subordinate nodes based on the file nodes corresponding to the heterogeneous policy documents. The bidirectional hierarchical recall module is used to perform hierarchical recall and pair filtering to obtain a list of clause pairs by pointing along a first direction from the first set of constraint units as the query side to the second set of constraint units as the candidate side, and along a second direction from the second set of constraint units as the query side to the first set of constraint units as the candidate side. The score and target pair filtering module is used to calculate the six-dimensional association score of each clause pair in the clause pair list based on the node information of the structural nodes, constraint element nodes and subordinate nodes of the constraint unit, and to filter out each target clause pair that the query-side constraint unit points to the target candidate side constraint unit with the highest six-dimensional association score. The confidence calculation module is used to determine the overall confidence level of conflict events for each target clause pair by combining the associated node information and preset constraint rules. The structured conclusion output module is used to compare the overall confidence level with the preset review threshold and output structured conclusions that represent the constraint conflicts of heterogeneous policy documents or those that need to be manually reviewed.