Document consistency adjudication method and device based on conflict map and evidence strength

Abstract

The application provides a document consistency adjudication method and device based on conflict graph and evidence strength, and belongs to the technical field of document processing. The method specifically comprises the following steps: generating evidence objects and conclusion objects containing fact primary keys based on a document set and a preset rule set, grouping based on the fact primary keys, and obtaining conflict relationships in combination with conflict detection; constructing a conflict graph, calculating conflict severity scores and evidence strength scores, determining the certainty of the conclusion objects corresponding to each conflict relationship in combination with preset adjudication rules, obtaining corresponding adjudication states, minimizing re-asking of the conflict relationships of the conclusion objects with the adjudication state being to be verified, until a preset re-asking end condition is reached, and outputting a document consistency conclusion according to the final adjudication state. The application realizes the identification and quantification of document conflict information by introducing a conflict graph, conflict severity and evidence strength, and guarantees the accuracy and verification efficiency of document verification in combination with automatic adjudication and directional verification.

Description

Technical Field

[0001] This invention relates to the field of document processing technology, and in particular to a method and apparatus for adjudicating document consistency based on conflict graphs and evidence strength. Background Technology

[0002] In document-intensive scenarios such as material testing report verification, regulatory compliance review, and technical specification review, the corresponding automated verification systems typically employ methods such as rule engine matching, document comparison, or knowledge graph queries to extract candidate conclusions from multiple sources of information, including the report text, appendices, and clause libraries, and compare them with standard requirements to determine compliance. For example, measured values ​​are extracted from the text using preset rule templates and compared with standard clauses, or differences between different paragraphs are identified through text similarity calculations. However, material testing reports and related documents generally suffer from issues such as diverse expressions, differing definitions, inconsistent units, version changes, and incomplete descriptions of applicable conditions. The same testing indicator may use different names, units, or precisions in different paragraphs, appendices, or even different cited clauses. Multiple levels of rules, such as mandatory standards, industry standards, corporate regulations, and contract clauses, may also impose different requirements on the same matter, leading to multiple contradictory candidate conclusions for the same matter from different sources.

[0003] When faced with multiple sources of conclusions, most existing document verification methods output the judgment result of each candidate conclusion independently according to their respective extraction paths, treating each conclusion as an isolated judgment item. When multiple conclusions appear for the same matter, one conclusion is selected as the final output according to the preset source priority or extraction order, and the remaining conclusions are directly discarded. Alternatively, differences are marked by numerical comparison, so that in the subsequent conflict handling stage, the original documents are manually reviewed to check the reasons for the differences, or the entire document extraction and judgment process is re-executed in the hope of obtaining a consistent result.

[0004] However, this verification method, which relies on prioritizing output results, is prone to losing conflict information during simple selection, potentially leading to omissions in the final verification results. Even when conflict marking is done through numerical comparison, it only achieves binary difference comparison and cannot identify the specific type and degree of impact of the conflict. Manual verification or a complete rerun is still required, resulting in low efficiency and accuracy in conflict identification, which in turn leads to low accuracy and efficiency in document verification results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing document verification methods, which are unable to effectively identify and quantify conflicts in multi-source conclusions, resulting in low accuracy and efficiency of document verification results. This invention provides a document consistency adjudication method and apparatus based on conflict graphs and evidence strength. By aligning factual primary keys and constructing conflict graphs, a conflict association structure is established between multi-source conclusions. Based on accurate identification of conclusion conflicts, the invention quantifies the assessment of conclusion conflicts through conflict severity and evidence strength. Furthermore, combined with deterministic adjudication based on fixed rules and a minimum re-questioning mechanism, it achieves automated adjudication and targeted verification of document consistency, ensuring the accuracy and efficiency of document verification.

[0006] The objective of this invention is achieved through the following technical solution: Document consistency adjudication methods based on conflict maps and strength of evidence include: The document set and the preset rule set are processed by document unit segmentation and field normalization to generate evidence objects and conclusion objects containing fact primary keys; The conclusion objects are grouped based on the fact primary key, and conflict detection is performed on each group of conclusion objects to obtain conflict relationships; A conflict graph is constructed based on the evidence object and the conclusion object, and the conflict severity score of each conflict relationship and the evidence strength score of each conclusion object are calculated through the conflict graph. Based on the preset adjudication rules, and combined with the corresponding conflict severity score and evidence strength score, a deterministic adjudication is made on the conclusion object corresponding to each conflict relationship, and the corresponding adjudication status is obtained; Identify the conclusion objects whose adjudication status is pending verification, minimize the corresponding conflict relationships and re-question until the preset re-question termination condition is met, and output the final document consistency conclusion based on the final adjudication status.

[0007] By segmenting documents into units and normalizing fields, unstructured information in the original documents is transformed into computable evidence objects and conclusion objects carrying factual primary keys. Conclusion objects are then grouped according to their factual primary keys, and conflict detection is performed within each group. Contradictions originally scattered across different sources are extracted as conflict relationships, achieving accurate alignment of multi-source conclusions and accurate identification of conflicts. Based on this, a corresponding conflict graph is constructed, establishing the association structure between conclusions and between conclusions and evidence. The conflict severity score for each conflict relationship and the evidence strength score for each conclusion object are calculated based on the conflict graph, transforming the severity of conflict and the sufficiency of evidence into comparable quantitative indicators, thus quantifying conflict. Then, according to preset adjudication rules, a deterministic adjudication is performed on each conflict relationship based on the conflict severity and evidence strength scores, automatically outputting the corresponding adjudication status. This avoids relying on simple priority selection or random results, which could lead to information omissions and ensure the accuracy of the verification results. Furthermore, for conclusion objects whose adjudication is pending verification, only the corresponding conflict relationship is subject to a minimum re-examination for qualitative verification, avoiding a complete rerun and improving the efficiency of document verification.

[0008] Furthermore, the process of grouping conclusion objects based on fact primary keys and performing conflict detection on each group of conclusion objects to obtain conflict relationships includes: Using the same fact key as the matching rule, all conclusion objects are traversed, and conclusion objects with the same fact key are grouped into the same fact key group; Based on preset conflict rules, conclusion objects in the same fact key group are compared pairwise. All fact key groups are traversed to identify conclusion object pairs that meet the preset conflict rules and are in the same fact key group. Based on the corresponding preset conflict rules, determine the conflict type of each conclusion object pair, and extract the corresponding conflict driving field and conflict difference. Based on the corresponding conflict type, conflict-driving field, and conflict difference, the corresponding conflict relationships are formed for each conclusion object pair.

[0009] Furthermore, the construction of the conflict graph based on the evidence object and the conclusion object includes: Based on the field content of each conclusion object, and combined with the preset support relationship judgment rules, the evidence objects are matched to obtain the associated evidence objects of each conclusion object. Using the conclusion object as the conclusion node and the evidence object as the evidence node, conflict edges are established for conflicting conclusion object pairs grouped under the same fact primary key, and support edges are established for each conclusion object and its associated evidence object. A conflict graph is constructed based on the conclusion nodes, evidence nodes, and their corresponding conflict edges and support edges.

[0010] Furthermore, the calculation of the conflict severity score for each conflict relationship and the evidence strength score for each conclusion object through the conflict map includes: Based on preset conflict index dimensions, the association information of each conflict relationship is extracted from the conflict map. The preset conflict index dimensions include at least the scope of conflict impact, conflict difference value, conflict spread scope, and evidence missing degree. Based on the preset weights of each conflict indicator dimension, the correlation information of each conflict indicator dimension is weighted and summed to obtain the conflict severity score of each conflict relationship. Extract evidence information from all evidence objects associated with each conclusion object from the conflict graph. The evidence information includes at least the completeness of the evidence location field, the degree to which the evidence content covers the conclusion field, the level of evidence source, the consistency of evidence, and the degree of matching of evidence version. Based on preset rules, the evidence information corresponding to each conclusion object is scored item by item, and the scores are weighted and summed according to the preset weights of the evidence items to obtain the evidence strength score of each conclusion object.

[0011] Furthermore, the step of making a deterministic ruling on the conclusion object corresponding to each conflict relationship based on preset ruling rules, combined with the corresponding conflict severity score and evidence strength score, and obtaining the corresponding ruling status includes: Based on the preset adjudication rules, and in combination with the corresponding conflict severity score and evidence strength score, the conclusion object pairs corresponding to each conflict relationship are verified according to the preset verification priority order, which is in the following order: hard rule verification, minimum evidence set verification, evidence strength score comparison verification, and source level and version verification. If the hard rule validation fails, set the ruling status of the corresponding conclusion object to the rejection status; When passing the hard rule verification, the next verification is executed according to the preset verification priority order; After one of the verifications, if the conclusion object pair does not meet the corresponding equal judgment conditions, select the conclusion object to be retained from the conclusion object pair according to the corresponding verification results, set the adjudication status of the selected conclusion object to the retained status, and set the adjudication status of the other conclusion object in the conclusion object pair to the rejection status. If the conclusion object pair meets the corresponding equal judgment condition, the next verification will continue until all verifications are completed. If the conclusion object pair meets the equal judgment condition corresponding to each verification, the adjudication status of the corresponding conclusion object will be set to the pending verification status.

[0012] Furthermore, the process involves identifying the conclusion object whose adjudication status is pending verification, minimizing the corresponding conflict relationships through re-questioning until a preset re-questioning termination condition is met, and outputting the final document consistency conclusion based on the final adjudication status, including: Identify the conclusion objects whose adjudication status is pending verification and determine the minimum conflict subgraph; Extract the minimum problem set used to distinguish the conclusion object from the judgment result based on the minimum conflict subgraph; Perform minimal query verification based on the minimum question set, and obtain the corresponding verification feedback information; Update the corresponding evidence strength score and conflict severity score based on the verification feedback information, and re-execute the preset adjudication rules to update the adjudication status; When the minimum conflict subgraph verification is completed and the minimum problem set is fully confirmed, or when the preset re-question termination condition is met, the final document consistency conclusion is generated and output based on all conclusion objects with the decision status set to "reserved".

[0013] Furthermore, the process of identifying the conclusion object whose adjudication status is pending verification and determining the minimum conflict subgraph includes: Filter out conflict relationships with a pending adjudication status and construct the corresponding set of unresolved conflict edges; Extract the endpoint conclusion object nodes of all conflicting edges in the set of unresolved conflicting edges to form a set of conclusion nodes; Select evidence nodes that are connected to the conclusion node set, filter target evidence nodes based on the conflict-driven field association relationship and the comparison judgment factors of the conclusion objects at both ends, and form a corresponding evidence node set; Determine the set of supporting edges between the set of conclusion nodes and the set of evidence nodes. Using the set of conclusion nodes and the set of evidence nodes as the node set, and the set of unresolved conflict edges and the set of supporting edges as the edge set, form a minimum conflict subgraph.

[0014] Furthermore, the extraction of the minimum problem set for distinguishing the conclusion object pair judgment result based on the minimum conflict subgraph includes: Extract the conflict type and conflict driving field corresponding to each unresolved conflict edge from the unresolved conflict edges corresponding to the minimum conflict subgraph; Based on the extracted conflict type and conflict driving fields, a corresponding verification task template is generated; Perform a merging process on the generated verification task templates to obtain the minimum set of questions used to distinguish the conclusion objects from the judgment results.

[0015] Furthermore, the step of performing document unit segmentation and field normalization on the document set and the preset rule set to generate evidence objects and conclusion objects containing fact primary keys includes: The document collection is divided into corresponding document units according to data type, and the fields of each document unit are normalized according to a set of preset rules to generate evidence objects; For each document unit, candidate conclusions are extracted, and fact keys are generated by combining them with preset tuple rules to form a conclusion object containing the fact keys.

[0016] A document consistency adjudication apparatus based on conflict graphs and evidence strength, used to execute any of the above-mentioned document consistency adjudication methods, includes: The document processing module is used to perform document unit segmentation and field normalization on the document collection and the preset rule collection, generate evidence objects and conclusion objects containing fact keys, group the conclusion objects according to the fact keys, and perform conflict detection on each group of conclusion objects to obtain conflict relationships. The conflict graph generation module is used to construct conflict graphs based on evidence objects and conclusion objects; The document deterministic adjudication module is used to calculate the conflict severity score and the evidence strength score of each conclusion object through the conflict graph. Based on the preset adjudication rules, it performs a deterministic adjudication on the conclusion object corresponding to each conflict relationship, obtains the corresponding adjudication status, and outputs the final document consistency conclusion based on the adjudication status. The minimize verification module is used to minimize the corresponding conflict relationship when a conclusion object with a judgment status pending verification is identified, until the preset re-questioning termination condition is met.

[0017] The beneficial effects of this invention are: By segmenting documents into units and normalizing fields, unstructured information in the original documents is transformed into computable evidence objects and conclusion objects carrying factual primary keys. Conclusion objects are then grouped according to their factual primary keys, and conflict detection is performed within each group. Contradictions originally scattered across different sources are extracted as conflict relationships, achieving accurate alignment of multi-source conclusions and accurate identification of conflicts. Based on this, a corresponding conflict graph is constructed, establishing the association structure between conclusions and between conclusions and evidence. The conflict severity score for each conflict relationship and the evidence strength score for each conclusion object are calculated based on the conflict graph, transforming the severity of conflict and the sufficiency of evidence into comparable quantitative indicators, thus quantifying conflict. Then, according to preset adjudication rules, a deterministic adjudication is performed on each conflict relationship based on the conflict severity and evidence strength scores, automatically outputting the corresponding adjudication status. This avoids relying on simple priority selection or random results, which could lead to information omissions and ensure the accuracy of the verification results. Furthermore, for conclusion objects whose adjudication is pending verification, only the corresponding conflict relationship is subject to a minimum re-examination for qualitative verification, avoiding a complete rerun and improving the efficiency of document verification. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a process of the present invention; Figure 2 This is a schematic diagram of a deterministic decision-making process according to an embodiment of the present invention; Figure 3This is a schematic diagram of a document consistency adjudication device according to an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Example: A document consistency adjudication method based on conflict graphs and evidence strength, such as... Figure 1 As shown, it includes: The document set and the preset rule set are processed by document unit segmentation and field normalization to generate evidence objects and conclusion objects containing fact primary keys; The conclusion objects are grouped based on the fact primary key, and conflict detection is performed on each group of conclusion objects to obtain conflict relationships; A conflict graph is constructed based on the evidence object and the conclusion object, and the conflict severity score of each conflict relationship and the evidence strength score of each conclusion object are calculated through the conflict graph. Based on the preset adjudication rules, and combined with the corresponding conflict severity score and evidence strength score, a deterministic adjudication is made on the conclusion object corresponding to each conflict relationship, and the corresponding adjudication status is obtained; Identify the conclusion objects whose adjudication status is pending verification, minimize the corresponding conflict relationships and re-question until the preset re-question termination condition is met, and output the final document consistency conclusion based on the final adjudication status.

[0021] By performing document unit segmentation and field normalization on the document set and the preset rule set, the unstructured original documents are transformed into computable evidence objects and conclusion objects carrying fact keys. Then, the conclusion objects are grouped according to the fact keys and conflict detection is performed to explicitly extract contradictions from different sources into structured conflict relationships, avoid the loss of conflict information in simple selection, and achieve accurate identification of conflict situations.

[0022] Furthermore, by constructing a conflict graph containing conclusion nodes and evidence nodes, the association structure between conclusions and between conclusions and evidence is established. Based on this, conflict severity scores and evidence strength scores are calculated, transforming the originally qualitatively ambiguous degree of conflict and sufficiency of evidence into comparable quantitative indicators. This is then combined with fixed, preset adjudication rules to make definitive decisions, automatically outputting the adjudication status as retained, rejected, or pending verification. This avoids the accuracy loss caused by relying on simple priority selection or random results, ensuring the accuracy of the verification results.

[0023] Furthermore, for conclusions that are pending verification, only the corresponding conflict relationships are subject to minimal re-questioning for targeted verification, avoiding overall reruns or redundant queries and improving verification efficiency.

[0024] This fixed-rule adjudication method also ensures the reproducibility of the adjudication process. Even if the same document set is run multiple times, the same conclusion can be obtained. The adjudication result does not change with the operating environment or processing order. Moreover, every rule triggered and every score comparison during the adjudication process can be traced and verified. This ensures the accuracy and efficiency of the verification results, as well as the stability and auditability of the verification results.

[0025] In this embodiment, the document consistency adjudication is a process in which, during the document verification process, for the same matter, such as whether one detection indicator is qualified or whether one standard clause is applicable, multiple candidate conclusions are extracted from multiple sources such as main text paragraphs, appendices, clause libraries, and database records. Then, it is identified whether there are contradictions among these conclusions, and the most credible conclusion is selected as the output based on unified rules and evidence quality.

[0026] The documents from which candidate conclusions are extracted include various types of content such as main text and appendices. These documents exhibit heterogeneous structures, disorganized formats, and are predominantly unstructured text. They commonly suffer from inconsistencies and non-standardization in the expression of indicator names, definitions, units, precision, and data location, making it impossible to directly extract effective information and generate standardized conclusions. Directly processing the raw unstructured data would lead to missing evidence and inconsistent conclusions due to chaotic data formats and non-standard fields, making it difficult to create aligned and verifiable processing objects. Therefore, this approach first performs document unit segmentation and field normalization on the candidate conclusion extraction sources, including the document set and the preset rule set, to transform the heterogeneous unstructured documents into a standardized format. This process then generates evidence objects containing document location and attribute information, as well as conclusion objects carrying a unified factual primary key.

[0027] Specifically, the step of performing document unit segmentation and field normalization on the document set and the preset rule set to generate evidence objects and conclusion objects containing fact primary keys includes: The document collection is divided into corresponding document units according to data type, and the fields of each document unit are normalized according to a set of preset rules to generate evidence objects; For each document unit, candidate conclusions are extracted, and fact keys are generated by combining them with preset tuple rules to form a conclusion object containing the fact keys.

[0028] First, the document set for which candidate conclusions need to be extracted is received. Then, the document set is identified by data type according to the corresponding file extension and internal format, and different data forms such as text paragraphs, tables, titles, signature pages, pictures and attachments are distinguished, and then the corresponding document units are obtained.

[0029] For the identified text paragraphs, natural paragraphs are segmented according to the corresponding line breaks and indentation boundaries, and the page number, paragraph number, and starting character offset of each paragraph in the document are recorded.

[0030] For the identified table, the row and column coordinates and original text of each cell are extracted, while the structural information of merged cells is preserved.

[0031] For the identified headings, the hierarchical relationship is identified by font size, bolding style, or numbering pattern. The headings are treated as independent units, and the range of their subordinate paragraphs is recorded.

[0032] For the identified signature pages, keyword matching is used to locate the signature area and extract the text and image information.

[0033] For the identified image units, the image location information is first obtained, and then the text in the image is converted into a processable text block through the optical character recognition (OCR) engine.

[0034] For the identified attachment units, they are processed recursively, following the same document unit segmentation method.

[0035] Furthermore, after the segmentation is completed, each segmented document unit is assigned a unique number, and its source type, document to which it belongs, page number, and coordinate position are recorded.

[0036] Then, based on a preset set of rules, each document unit is normalized to generate an evidence object. The preset set of rules includes a category dictionary, an indicator dictionary, a unit conversion table, a clause identification template, and a conclusion sentence template.

[0037] During field normalization, the text of each document unit is first cleaned, removing redundant spaces and line breaks, and standardizing punctuation. Then, a maximum positive matching is performed using a category dictionary and an indicator dictionary to extract the material category and testing indicator name from the text. If multiple candidates are matched, the finest granularity is selected based on dictionary priority. Next, regular expressions are used to extract numerical values ​​and their associated units, and these values ​​are uniformly converted to standard units and corresponding standard values ​​according to a unit conversion table. Simultaneously, a clause recognition template is used to extract the standard number and clause number.

[0038] Based on this, the fields are integrated according to the field normalization results of each document unit to form corresponding evidence objects. Each evidence object includes at least the source type, page number, location, original text fragment, abstract fingerprint, version number, timestamp, and source level.

[0039] The original text fragment records the corresponding cleaned text, the digest fingerprint is the hash value calculated for that text, and the version number is the normalized standard version format. Meanwhile, the evidence object also retains its original location information, including page number, paragraph number or table coordinates, start and end offsets, source type, and document timestamp, forming a locatable and verifiable evidence anchor.

[0040] Based on this, the system further extracts candidate conclusions from each document unit to form a conclusion object that includes the fact key.

[0041] When extracting candidate conclusions, for statements directly containing judgment terms, the states of qualified, unqualified, satisfied, and unsatisfied are extracted through conclusion sentence template matching. For units that only contain measured values ​​without providing specific conclusions, the corresponding conclusions are derived based on the judgment rules in the preset rule template library and the preset standard clause library. For clause units, the applicability of the clause is determined based on the sample category and testing items in the report, generating an applicable or inapplicable conclusion. Then, combined with the normalized fields, each obtained candidate conclusion is encapsulated into a conclusion object. The conclusion object includes at least a unique number, source document type, numerical value, judgment status, applicable conditions, clause reference, a list of evidence object numbers on which it is based, initial confidence level, and fact key.

[0042] The fact key is generated by normalizing fields and using preset tuple rules. These preset tuple rules define the fact key as being constructed by sequentially concatenating fields such as material category, sample or object, test item name, judgment attribute, unit caliber, version number, and applicable scope. Based on these preset tuple rules, corresponding field values ​​are extracted from the normalized fields to form the corresponding fact key string.

[0043] By generating evidence objects and conclusion objects, the unstructured information of the original document can be transformed into two interconnected structured sets: evidence objects and conclusion objects. The evidence objects carry the normalized evidence fragments and their location information, while the conclusion objects carry the normalized conclusion content and its factual primary key. This ensures the efficiency of conflict identification and guarantees the reliability and accuracy of the identification results.

[0044] Because conclusions generated from multi-source documents are scattered and expressed in diverse ways, conclusions on the same issue may appear in different locations such as the main text, appendices, and clauses, and may differ in terms of name, scope, units, and precision. Lacking a unified alignment identifier, it is impossible to directly determine which conclusions correspond to the same issue. Performing indiscriminate pairwise comparisons of all conclusions would generate numerous irrelevant comparisons and misjudgments, making it impossible to accurately locate the true conflict. Therefore, we first introduce a fact key to group each conclusion object, grouping conclusions on the same issue into the same group. Then, conflict detection is performed within each group. This ensures that conflict judgments are only performed on conclusions on the same issue, eliminating interference from irrelevant conclusions, ensuring the accuracy of conflict detection, reducing the scope of invalid comparisons, and improving detection efficiency.

[0045] Specifically, the process of grouping conclusion objects based on fact primary keys and performing conflict detection on each group of conclusion objects to obtain conflict relationships includes: Using the same fact key as the matching rule, all conclusion objects are traversed, and conclusion objects with the same fact key are grouped into the same fact key group; Based on preset conflict rules, conclusion objects in the same fact key group are compared pairwise. All fact key groups are traversed to identify conclusion object pairs that meet the preset conflict rules and are in the same fact key group. Based on the corresponding preset conflict rules, determine the conflict type of each conclusion object pair, and extract the corresponding conflict driving field and conflict difference. Based on the corresponding conflict type, conflict-driving field, and conflict difference, the corresponding conflict relationships are formed for each conclusion object pair.

[0046] Iterate through all conclusion objects, using the same fact primary key as the matching rule. Determine whether the fact primary keys are the same by comparing each field, and then group conclusion objects with the same fact primary key into the same group.

[0047] Each group corresponds to a unique fact item, and a group may contain two or more conclusion objects from different sources. Groups containing only a single conclusion object do not have conflicts and require no processing. Groups containing multiple conclusion objects undergo further conflict detection.

[0048] For each fact key group that needs to be processed, all conclusion objects within the corresponding fact key group are compared pairwise, that is, every two different conclusion objects are combined to form a conclusion pair. For each conclusion pair, the preset conflict rules are called sequentially according to the preset rule priority order for judgment.

[0049] The preset conflict rules include at least numerical conflict rules, unit conflict rules, range conflict rules, reference conflict rules, and version conflict rules.

[0050] The determination condition for the numerical conflict rule is that the absolute difference or relative difference of the numerical fields of two conclusion objects after unit normalization exceeds a preset threshold. The threshold can be dynamically configured according to the indicator type.

[0051] The determination criteria for the unit conflict rule are that the unit fields of the two conclusion objects are different, and it is determined that the conversion cannot be performed after calling the unit conversion table, or the difference in values ​​after conversion still exceeds the threshold.

[0052] The determination condition for the scope conflict rule is that the applicable condition fields of the two conclusion objects exist and are determined to be incompatible according to the preset condition logic determination table. For example, one conclusion object requires that it be true only under specific conditions, while the other conclusion object gives a conclusion that is unconditionally applicable.

[0053] The criteria for determining the citation conflict rule are that the evidence sets on which the two conclusion objects are based lack evidence anchors, the hash values ​​of the evidence fragments do not match, or the evidence content does not cover the conditional clause.

[0054] The determination condition for the version conflict rule is that the version fields of the two conclusion objects are different and do not meet the preset version field compatibility rules.

[0055] The specific judgment conditions of the conflict rules can be set according to actual needs, and the judgment order can also be adjusted according to actual needs.

[0056] Once the condition for a rule is met, a conflict is considered to exist between the conclusions. At this point, the conflict type is determined based on the type of rule that was triggered. For example, if a numerical conflict rule is triggered, the conflict type is numerical conflict; if a range conflict rule is triggered, the conflict type is range conflict.

[0057] Simultaneously, the fields causing the conflict are extracted from the two conclusion objects as conflict-driving fields. For example, for numerical conflicts, the values ​​and units of the two numerical fields are extracted; for unit conflicts, the two unit strings are extracted; and for range conflicts, the descriptive text of the two applicable conditions is extracted. Furthermore, the difference between the two is calculated as the conflict difference. For numerical conflicts, the difference can be expressed as an absolute difference or a relative percentage; for unit conflicts, the difference can be expressed as a conversion factor between the two units or a statement that they cannot be converted; and for range conflicts, the difference can be expressed as the description of the two conditions.

[0058] For a pair of conclusions that simultaneously satisfy multiple conflict rules, multiple independent conflict relationship objects can be generated, or the results of multiple conflict rules can be merged into a single conflict relationship object of a comprehensive conflict type.

[0059] Finally, the conflict type, conflict-driving field, conflict difference, and the identification information of the two conclusion objects in the conclusion pair are assembled into a structured conflict relationship. Each conflict relationship contains at least a unique identifier, the fact primary key group identifier to which it belongs, a conflict type enumeration value, the numbers of the two conclusion objects, the name of the conflict-driving field and its specific value in the conclusion object, and a descriptive text of the difference.

[0060] After generating evidence objects, conclusion objects, and conflict relationships, these three types of data objects are stored in different sets. Although the conflict relationships between conclusion objects are marked by conflict relationship objects, the conflict relationship objects themselves do not establish a direct connection with the evidence objects associated with the conclusion objects. When making a conflict verification decision, multiple random lookup operations across sets are required. When the number of conclusion objects, evidence objects, and conflict relationships is large, this distributed storage and multiple lookup method will lead to an increase in processing steps, a longer data access path, and lower decision efficiency.

[0061] Therefore, a conflict graph is further established to explicitly connect the three independent data sets through nodes and edges, so that the support relationship between the conclusion node and the evidence node, and the conflict relationship between the conclusion nodes can be expressed in the same graph structure. This integrates discrete information into an associative graph model and improves the processing efficiency of subsequent conflict verification and adjudication.

[0062] Specifically, the construction of the conflict graph based on evidence objects and conclusion objects includes: Based on the field content of each conclusion object, and combined with the preset support relationship judgment rules, the evidence objects are matched to obtain the associated evidence objects of each conclusion object. Using the conclusion object as the conclusion node and the evidence object as the evidence node, conflict edges are established for conflicting conclusion object pairs grouped under the same fact primary key, and support edges are established for each conclusion object and its associated evidence object. A conflict graph is constructed based on the conclusion nodes, evidence nodes, and their corresponding conflict edges and support edges.

[0063] For each conclusion object, its list of evidence object numbers is read, and matching and verification are performed according to preset support relationship determination rules. The support relationship determination rules are used to determine whether an evidence object has a support relationship with the corresponding conclusion object. The determination conditions include whether the document unit of the evidence source and the conclusion source belongs to the same report, whether the evidence fragment contains the numerical value or conditional statement corresponding to the conclusion, and whether the timestamp or version of the evidence matches the conclusion, etc.

[0064] Identify the evidence objects that meet the corresponding judgment conditions of the supporting relationship determination rules, and use the identified evidence objects as the related evidence objects of the conclusion objects.

[0065] Each conclusion object is defined as a conclusion node, and each piece of evidence is defined as a piece of evidence node. For each pair of conclusion objects marked as conflicting by a conflict relationship, a conflict edge is established between their corresponding two conclusion nodes. For each conclusion node and each associated piece of evidence node determined by the support relationship determination rule, a support edge is established between them. All conclusion nodes and evidence nodes constitute the node set of the conflict graph, and all conflict edges and support edges constitute the edge set of the conflict graph.

[0066] Based on the set of nodes and edges, a conflict graph is constructed. In the constructed conflict graph, conclusion nodes are connected to evidence nodes that provide evidence to support them through supporting edges, and evidence nodes are connected in reverse to the conclusion nodes they support through supporting edges. Conclusion nodes that belong to the same fact key group and have conflicts are directly connected through conflict edges.

[0067] However, constructing a conflict map only qualitatively identifies the conflict relationships between conclusions, failing to distinguish the degree of conflict impact and the strength of evidence supporting the conclusions, making it difficult to achieve objective and stable consistent rulings. Therefore, based on the conflict map, further calculations of conflict severity scores and evidence strength scores transform qualitative conflict relationships into comparable quantitative values, quantifying the level of conflict impact and the strength of evidence support. This provides objective data for subsequent definitive rulings, thereby improving the accuracy and stability of document consistency rulings.

[0068] Specifically, the calculation of the conflict severity score for each conflict relationship and the evidence strength score for each conclusion object through the conflict map includes: Based on preset conflict index dimensions, the association information of each conflict relationship is extracted from the conflict map. The preset conflict index dimensions include at least the scope of conflict impact, conflict difference value, conflict spread scope, and evidence missing degree. Based on the preset weights of each conflict indicator dimension, the correlation information of each conflict indicator dimension is weighted and summed to obtain the conflict severity score of each conflict relationship. Extract evidence information from all evidence objects associated with each conclusion object from the conflict graph. The evidence information includes at least the completeness of the evidence location field, the degree to which the evidence content covers the conclusion field, the level of evidence source, the consistency of evidence, and the degree of matching of evidence version. Based on preset rules, the evidence information corresponding to each conclusion object is scored item by item, and the scores are weighted and summed according to the preset weights of the evidence items to obtain the evidence strength score of each conclusion object.

[0069] For the conflict severity score, first traverse each conflict edge in the conflict graph, and extract the associated information of the preset conflict index dimensions based on the two conclusion nodes connected by the conflict edge and their associated evidence nodes.

[0070] The dimensions of conflict indicators should include at least the scope of conflict impact, conflict difference value, scope of conflict spread, and degree of evidence absence.

[0071] The scope of conflict impact is determined by the importance of the conclusions involved in the conflict relationship corresponding to the conflict edge. It is obtained by reading the clause level field in the conclusion node. The higher the clause level, the greater the scope of conflict impact.

[0072] The conflict difference value is the degree of difference between two conclusion nodes in terms of value or state. For numerical conflicts, the absolute or relative difference of the values ​​of the two conclusion nodes after unit normalization is calculated and mapped to a score between zero and one. For non-numerical conflicts, the corresponding preset difference value score is extracted according to the conflict type.

[0073] The scope of conflict diffusion refers to whether the conclusion node connected by the conflict edge is also connected to other conclusion nodes through supporting edges or conflict edges. The number of other conflict edges that can be accessed from the current conflict edge is counted by graph traversal algorithm. The more of these other conflict edges, the higher the diffusion scope score.

[0074] The evidence missingness is the proportion of evidence nodes associated with the conclusion nodes of both sides of the conflict that lack anchor point location information, conditional sentences, or version information. The number of evidence nodes that do not meet the completeness requirements in the evidence set of each conclusion node is counted, and the average value is taken and normalized to obtain the score.

[0075] By combining the pre-set weighting coefficients for each conflict indicator dimension, the scores of each dimension are weighted and summed to obtain the conflict severity score for the corresponding conflict relationship.

[0076] For the evidence strength score, each conclusion node in the conflict graph is traversed, and all associated evidence nodes are accessed along the supporting edges from the corresponding conclusion node. For each accessed evidence node, multiple sub-items are extracted from the evidence information, including the completeness of the evidence location field, the degree to which the evidence content covers the conclusion field, the level of evidence source, the consistency of evidence, and the degree of matching of evidence version.

[0077] The completeness of the evidence location fields refers to whether the page number, segment number or table coordinates, start and end offsets, and summary fingerprint in the evidence node are complete. Each missing item will result in a deduction of the corresponding score.

[0078] The degree to which the evidence content covers the conclusion field is determined by whether the evidence fragment contains the numerical values, units, conditional statements, or clause numbers on which the conclusion depends. This is determined through text matching or rule templates, and the more comprehensive the coverage, the higher the score.

[0079] The evidence source level is specifically mapped to a preset value based on the source type field of the evidence node. The level of the main text and standard clause library is higher than that of the appendix, and the level of the appendix is ​​higher than that of the image recognition result.

[0080] Evidence consistency refers to whether there is a contradiction between the conclusion nodes supported by the corresponding evidence node. Specifically, it involves traversing all conclusion nodes supported by the same evidence node. If the values ​​or states of these conclusion nodes are consistent, a predetermined score is assigned; if there is a contradiction, a predetermined score is deducted.

[0081] Evidence version matching score is determined by whether the version field of the evidence node is consistent with the standard version used in the current verification. Full marks are awarded when they are consistent, and scores are reduced based on the degree of version difference when they are inconsistent.

[0082] After scoring each sub-item according to preset rules, the scores are then combined with the weight coefficients corresponding to each sub-item and summed to obtain the evidence strength score for the corresponding conclusion node.

[0083] By using conflict severity scoring and evidence strength scoring, the determination of a ruling can be transformed from a qualitative comparison based on a fixed priority sequence to a weighted comparison combining quantitative indicators. This quantitative approach improves the stability of the ruling and ensures that every score value in the ruling process can be recorded and traced, thereby enhancing the auditability of the ruling.

[0084] Furthermore, to further improve the stability of the rulings, a fixed rule sequence is adopted, which combines the conflict severity score and the evidence strength score to make a deterministic ruling. This ensures that each ruling is executed in the same order, without introducing random comparisons or dynamic weights that depend on the input order. This ensures that the mapping relationship from the conflict severity score and the evidence strength score to the final retention, rejection, or pending verification status is always determined and is not affected by external factors that change the judgment path.

[0085] Specifically, the process of making a definitive ruling on the conclusion object corresponding to each conflict relationship based on preset ruling rules, combined with the corresponding conflict severity score and evidence strength score, and obtaining the corresponding ruling status includes: Based on the preset adjudication rules, and in combination with the corresponding conflict severity score and evidence strength score, the conclusion object pairs corresponding to each conflict relationship are verified according to the preset verification priority order, which is in the following order: hard rule verification, minimum evidence set verification, evidence strength score comparison verification, and source level and version verification. If the hard rule validation fails, set the ruling status of the corresponding conclusion object to the rejection status; When passing the hard rule verification, the next verification is executed according to the preset verification priority order; After one of the verifications, if the conclusion object pair does not meet the corresponding equal judgment conditions, select the conclusion object to be retained from the conclusion object pair according to the corresponding verification results, set the adjudication status of the selected conclusion object to the retained status, and set the adjudication status of the other conclusion object in the conclusion object pair to the rejection status. If the conclusion object pair meets the corresponding equal judgment condition, the next verification will continue until all verifications are completed. If the conclusion object pair meets the equal judgment condition corresponding to each verification, the adjudication status of the corresponding conclusion object will be set to the pending verification status.

[0086] The preset decision rules include a preset verification priority order and specific judgment conditions for each verification.

[0087] The first priority is hard rule verification, with specific criteria including whether the standard version referenced by the conclusion object is outdated, whether the image sample name is significantly inconsistent with the sample name on the homepage, whether required signatures or fields are missing, and whether units are unconvertible and not explained in the document. If one conclusion object triggers any hard rule, the decision status of that conclusion object is set to rejected, the other conclusion object that did not trigger a hard rule is set to reserved, and subsequent verification of the conflict relationship is terminated. If neither conclusion object triggers any hard rule, the second priority, minimum evidence set verification, proceeds.

[0088] In the minimum evidence set verification, the corresponding preset minimum evidence set requirements are obtained based on the factual primary key group to which the conflict relationship belongs. Different factual items have different minimum evidence set requirements; for example, evidence must include anchor point information such as page numbers and paragraph numbers, must cover key conditional sentences, and must include version numbers. The actual evidence sets of the two conclusion objects are compared with the minimum evidence sets. The comparison results determine whether the equivalence criteria are met; in this case, the equivalence criteria are that the comparison results of the two are consistent. That is, if one conclusion object meets the minimum evidence set requirements while the other does not, the conclusion object that meets the requirements is marked as reserved, and the conclusion object that does not meet the requirements is marked as rejected, and subsequent verification is terminated. If both conclusion objects meet or do not meet the minimum evidence set requirements, the process proceeds to the third priority level, namely, evidence strength score comparison verification.

[0089] In the evidence strength score comparison and verification, the calculated evidence strength scores of the two conclusion objects are extracted, and the difference between them is calculated. The result of the difference calculation is used to determine whether the equivalence criteria are met. At this point, the equivalence criteria are that the difference between the two scores does not exceed a preset threshold. That is, if the difference exceeds the preset threshold, the conclusion object with the higher evidence strength score is set to a retained state, and the conclusion object with the lower score is set to a rejected state, and subsequent verification is terminated. If the difference does not exceed the preset threshold, the fourth priority level, namely source level and version verification, is initiated.

[0090] In the source level and version verification, the source level and version number of the evidence for the two conclusion objects are first compared. Based on the comparison result, it is determined whether the equivalence criteria are met. In this case, the equivalence criteria are that both the source level and version number are identical. That is, if the source levels of the two conclusion objects are different, the one with the higher level is set to a reserved state, and the one with the lower level is set to a rejected state, and the verification is terminated. If the source levels of the evidence are the same, the version numbers of the two conclusion objects are further compared. The one with the newer version is set to a reserved state, and the one with the older version is set to a rejected state, and the verification is terminated. If both the source level and version number are the same, it is determined that the two conclusion objects cannot be distinguished by the preset adjudication rules.

[0091] If, after full verification, the superiority or inferiority of the two conclusion objects in a conclusion object pair cannot be distinguished by the preset adjudication rules, the adjudication status of both parties will be set to pending verification.

[0092] After a definitive decision is made, a portion of the decisions remain pending verification. The conflicts between these pending decisions are those that cannot be distinguished even after applying all preset priority rules. If no action is taken against these conflicts and the final conclusion is output directly, the factual items involved in these conflicts will lack definitive judgments, compromising the integrity of the document's consistency. Re-running the entire process or indiscriminately manually querying all conflicts would increase processing costs.

[0093] Therefore, qualitative verification is further performed based on the conflict relationships in the pending verification status. Information is supplemented by minimizing the re-questioning of the corresponding conflict relationships. After obtaining verification feedback, the relevant data is updated to re-execute the ruling. This is to supplement the handling of conflicts that cannot be handled by automatic ruling while avoiding a complete rerun, thus ensuring the integrity of the document consistency conclusion.

[0094] Specifically, such as Figure 2 As shown, based on preset adjudication rules, and combined with corresponding conflict severity scores and evidence strength scores, a deterministic adjudication is made for the conclusion objects corresponding to each conflict relationship. Based on obtaining the adjudication status of each conclusion object, conclusion objects with adjudication status pending verification are identified. The corresponding conflict relationships are then subjected to minimized re-questioning until a preset re-questioning termination condition is met. Finally, based on the final adjudication status, the final document consistency conclusion is output, including: Identify the conclusion objects whose adjudication status is pending verification and determine the minimum conflict subgraph; Extract the minimum problem set used to distinguish the conclusion object from the judgment result based on the minimum conflict subgraph; Perform minimal query verification based on the minimum question set, and obtain the corresponding verification feedback information; Update the corresponding evidence strength score and conflict severity score based on the verification feedback information, and re-execute the preset adjudication rules to update the adjudication status; When the minimum conflict subgraph verification is completed and the minimum problem set is fully confirmed, or when the preset re-question termination condition is met, the final document consistency conclusion is generated and output based on all conclusion objects with the decision status set to "reserved".

[0095] If information supplementation or manual inquiry is performed indiscriminately on all objects whose conclusions need to be verified, it will involve a large number of evidence and conclusion points that are not related to the conflict, resulting in an excessively large scope of verification and reduced processing efficiency.

[0096] Therefore, we first filter out all conflict edges to be verified from the conflict graph, extract the endpoint conclusion nodes of these conflict edges, and then filter out the evidence nodes directly related to the conflict-driving field. At the same time, we prune evidence nodes that do not contribute to the conflict determination, and finally form a minimal graph structure containing only the conflicts necessary for verification. By determining the minimal conflict subgraph, we can accurately locate the relevant information that causes the conflict to be unresolved, thereby accurately defining the scope of targeted inquiries and avoiding overall reruns or redundant inquiries.

[0097] The process of identifying the decision status as a conclusion object to be verified and determining the minimum conflict subgraph includes: Filter out conflict relationships with a pending adjudication status and construct the corresponding set of unresolved conflict edges; Extract the endpoint conclusion object nodes of all conflicting edges in the set of unresolved conflicting edges to form a set of conclusion nodes; Select evidence nodes that are connected to the conclusion node set, filter target evidence nodes based on the conflict-driven field association relationship and the comparison judgment factors of the conclusion objects at both ends, and form a corresponding evidence node set; Determine the set of supporting edges between the set of conclusion nodes and the set of evidence nodes. Using the set of conclusion nodes and the set of evidence nodes as the node set, and the set of unresolved conflict edges and the set of supporting edges as the edge set, form a minimum conflict subgraph.

[0098] Iterate through all adjudicated conflict relationships, filter out those with an adjudication status of "awaiting verification," and extract the conflict edges corresponding to each of the filtered conflict relationships to form a set of unresolved conflict edges. Each conflict edge in this set connects two conclusion nodes, and both of these conclusion nodes are currently in an "awaiting verification" state.

[0099] Extract the two endpoints of all conflicting edges in the set of unresolved conflicting edges, i.e., the two conclusion nodes connected by each conflicting edge, and collect all endpoint conclusion nodes into a set of conclusion nodes. Each conclusion node in this set is at least an endpoint of an unresolved conflicting edge, indicating that the conclusion object participated in one or more conflicts that cannot be automatically adjudicated.

[0100] For each conclusion node in the conclusion node set, find the associated evidence nodes along all supporting edges originating from each conclusion node, and collect these evidence nodes into a candidate evidence node set. Then, the candidate evidence node set is filtered based on the conflict-driving field recorded by each unresolved conflict edge.

[0101] Conflict-driven fields include numerical values, units, applicable conditions, clause references, versions, anchors, etc. Only evidence nodes that are related to conflict-driven fields are retained. For example, if the conflict-driven field is a numerical value, evidence nodes containing the corresponding numerical value or calculation basis are retained; if the conflict-driven field is an applicable condition, evidence nodes containing the condition description are retained.

[0102] At the same time, the comparison judgment factors of the two conclusion objects are considered, that is, whether a certain evidence node affects the distinction between the superiority and inferiority of the two conclusion objects. If deleting a certain evidence node does not change the integrity of the conflict-driving field, nor does it affect the comparison result between the two conclusion objects, then the corresponding evidence node is considered to have no impact on the conflict and can be removed from the set.

[0103] After filtering the conflict-driven field relationships and the comparison and judgment factors of the conclusion objects at both ends, the final set of evidence nodes is obtained.

[0104] Determine the set of supporting edges between the set of conclusion nodes and the set of evidence nodes, that is, retain all supporting edges connecting any node in the set of conclusion nodes to any node in the set of evidence nodes. At the same time, the set of unresolved conflict edges remains unchanged.

[0105] Using the set of conclusion nodes and the set of evidence nodes as the node set, and the set of unresolved conflict edges and the set of supporting edges as the edge set, a minimal conflict subgraph is constructed. This subgraph contains only conclusion nodes and evidence nodes directly related to the conflict to be verified, and evidence nodes that do not contribute to the conflict determination have been pruned, thus defining an accurate range for generating a minimal problem set and initiating targeted verification.

[0106] Based on the generated minimum conflict subgraph, the conflict information in the minimum conflict subgraph is further transformed into a structured and non-redundant verification task by extracting the minimum problem set, providing executable input for subsequent targeted verification.

[0107] The step of extracting the minimum problem set for distinguishing the conclusion object pair judgment result based on the minimum conflict subgraph includes: Extract the conflict type and conflict driving field corresponding to each unresolved conflict edge from the unresolved conflict edges corresponding to the minimum conflict subgraph; Based on the extracted conflict type and conflict driving fields, a corresponding verification task template is generated; Perform a merging process on the generated verification task templates to obtain the minimum set of questions used to distinguish the conclusion objects from the judgment results.

[0108] Iterate through each unresolved conflict edge in the minimum conflict subgraph. For each conflict edge, read the conflict type and conflict driver field recorded in the corresponding conflict relation object.

[0109] Based on the extracted conflict type and conflict driving field, a preset verification task template library is matched. This library stores predefined query statement templates corresponding to each combination of conflict type and driving field. Then, the corresponding information of the structural objects at both ends of the conflict edge is filled into the matched predefined query statements to generate the corresponding verification task template.

[0110] After generating all verification tasks corresponding to unresolved conflict edges, a merging process is performed. The merging criteria include whether multiple verification tasks point to the same document source, the same database record number, and whether they require confirmation of the same field. By comparing the document identifier, page number, table coordinates, database primary key, and the queried field name in the verification tasks, tasks that are duplicated or have a similarity exceeding a preset similarity threshold are identified.

[0111] For tasks judged as duplicates or highly similar, retain only one task, merge the other tasks into the retained task, and summarize all relevant conclusion pair numbers and conflict descriptions in the retained task. After merging, output the minimum problem set.

[0112] Next, a minimal query verification is performed based on the generated minimum question set. Specifically, the corresponding verification method is selected according to the specific content of each verification task. For questions requiring manual confirmation of definitions, units, applicable conditions, or versions, the verification task is sent to the human review end in the form of a structured form or natural language questions, and human feedback is received. For tasks requiring relocating evidence anchors or completing conditional sentences, the original document is re-parsed, including relocating the original text fragments of specified page numbers and paragraphs, re-performing optical character recognition to obtain more complete image text, re-fetching relevant records from the database, or relocating the complete content of standard clauses from the clause library. After collecting all verification feedback information, it is packaged into a verification result object, which at least includes a verification task identifier, verification conclusion, and verification timestamp.

[0113] Based on the verification feedback, the relevant data in the conflict map is updated. For feedback involving confirmation of units or values, the unit and value fields of the corresponding conclusion object are adjusted, replaced with the verified standard values, and the unit conversion consistency sub-item in the evidence strength score of that conclusion object is recalculated. For feedback involving confirmation of applicable conditions, the applicable conditions field of the corresponding conclusion object is updated, and the evidence coverage score is reassessed. For feedback that supplements evidence anchors or conditional sentences, anchor fields and original text fragments are added or modified in the evidence object, and the location integrity score and coverage score of the evidence object are recalculated, thereby updating the evidence strength score of the associated conclusion object. Simultaneously, based on the updated conclusion object values ​​and evidence object information, the conflict severity score of the relevant conflict relationships is recalculated, including recalculating conflict difference values, evidence missingness, and other indicators.

[0114] After the score update is completed, the preset adjudication rules are re-executed. If the status of retention and rejection can be distinguished after the adjudication, the adjudication status of the corresponding conclusion object is updated. If it is still impossible to distinguish, the status of pending verification continues.

[0115] Next, determine whether the verification termination conditions are met. Verification termination conditions include: all verification tasks corresponding to unresolved conflict edges in the minimum conflict subgraph have received feedback, and after re-adjudication, there are no more conclusion objects to be verified (i.e., all conflicts have been resolved); and although there are still unresolved conflicts, the preset re-question termination conditions have been met, such as reaching the verification round limit, exhausting verification resources, or being marked as unverifiable by the manual review end. When any termination condition is met, the minimum re-question loop terminates.

[0116] Collect all conclusion objects whose adjudication status is reserved, group them according to their fact key, and output a final conclusion for each fact key group.

[0117] The output conclusion includes the unique identifier of the corresponding fact item, the value and status of the retained conclusion, the chain of evidence on which it is based, and the adjudication path.

[0118] The evidence chain consists of the anchor point and original text summary of each piece of evidence extracted from the evidence set of the conclusion object, and the adjudication path consists of the rules triggered in the adjudication process of each conflict, the scoring comparison results, and the verification feedback records.

[0119] If some fact items still have unresolved conflicts due to exhaustion of verification resources, output a conclusion that cannot be determined or requires manual review, and attach a list of missing evidence, listing the missing anchor points, conditional clauses or version information, and binding them to the corresponding fact primary key.

[0120] Simultaneously, based on the information from the entire adjudication process, corresponding version signatures, evidence signatures, rule signatures, and output signatures are generated and output as part of the audit information along with the document consistency conclusion. Specifically, first, the version numbers of each document in the document set used for the current verification, the version numbers of the standard clause library, and the version identifiers of the historical version library are read. These version information is concatenated in a fixed order and a hash value is calculated to obtain the version signature. Next, all evidence objects involved in the current verification are traversed, and the digest fingerprint, original text fragment hash value, and anchor point positioning information of each evidence object are extracted. These are concatenated in the order of the evidence object numbers and a hash value is calculated to obtain the evidence signature. Then, the version of the preset conflict rule library used for the current adjudication, the preset adjudication rule sequence identifier, and the threshold parameters of each rule are read. These rule configuration information is concatenated and a hash value is calculated to obtain the rule signature. Finally, each conclusion, its evidence chain, and adjudication path in the final output document consistency conclusion are serialized in a fixed format and a hash value is calculated to obtain the output signature.

[0121] During subsequent verification, by recalculating the four sub-signatures in the current operating environment and comparing them one by one with the original output sub-signatures, the source of the change can be quickly located, improving the audit capability for document consistency results. If the version signature changes, it indicates that the cause of the discrepancy is a change in the report document version or clause version. If the evidence signature changes, it indicates that the original evidence, database record, or optical character recognition result has changed. If the rule signature changes, it indicates that the version of the adjudication rule or the rule parameter has changed. If the version signature, evidence signature, and rule signature all remain unchanged, but the output signature changes, it indicates that there was an anomaly or change in the generation process of the output content.

[0122] Therefore, the conclusions corresponding to the factual primary keys in the final output, as well as the adjudication process information, are integrated to output the final document consistency conclusion.

[0123] Another aspect of this embodiment also provides a document consistency adjudication device based on conflict graphs and evidence strength, such as... Figure 3 As shown, it includes: The document processing module is used to perform document unit segmentation and field normalization on the document collection and the preset rule collection, generate evidence objects and conclusion objects containing fact keys, group the conclusion objects according to the fact keys, and perform conflict detection on each group of conclusion objects to obtain conflict relationships. The conflict graph generation module is used to construct conflict graphs based on evidence objects and conclusion objects; The document deterministic adjudication module is used to calculate the conflict severity score and the evidence strength score of each conclusion object through the conflict graph. Based on the preset adjudication rules, it performs a deterministic adjudication on the conclusion object corresponding to each conflict relationship, obtains the corresponding adjudication status, and outputs the final document consistency conclusion based on the adjudication status. The minimize verification module is used to minimize the corresponding conflict relationship when a conclusion object with a judgment status pending verification is identified, until the preset re-questioning termination condition is met.

[0124] The output of the document processing module is connected to the input of the conflict graph generation module. The output of the conflict graph generation module is connected to the first input of the document deterministic decision module. The first output of the document deterministic decision module is connected to the input of the minimization verification module. The output of the minimization verification module is connected to the feedback input of the document deterministic decision module. The document deterministic decision module also includes a second output for outputting the final document consistency conclusion.

[0125] The document processing module, conflict graph generation module, document deterministic adjudication module, and minimal verification module are all microprocessors, computers, and other components with corresponding data processing capabilities.

[0126] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A document consistency adjudication method based on conflict graphs and evidence strength, characterized in that, include: The document set and the preset rule set are processed by document unit segmentation and field normalization to generate evidence objects and conclusion objects containing fact primary keys; The conclusion objects are grouped based on the fact primary key, and conflict detection is performed on each group of conclusion objects to obtain conflict relationships; A conflict graph is constructed based on the evidence object and the conclusion object, and the conflict severity score of each conflict relationship and the evidence strength score of each conclusion object are calculated through the conflict graph. Based on the preset adjudication rules, and combined with the corresponding conflict severity score and evidence strength score, a deterministic adjudication is made on the conclusion object corresponding to each conflict relationship, and the corresponding adjudication status is obtained; Identify conclusions in the adjudication status as pending verification, minimize the corresponding conflict relationships through re-questioning, until a preset re-questioning termination condition is met, and output the final document consistency conclusion based on the final adjudication status. Filter out conflict relationships with a pending adjudication status and construct the corresponding set of unresolved conflict edges; Extract the endpoint conclusion object nodes of all conflicting edges in the set of unresolved conflicting edges to form a set of conclusion nodes; Select evidence nodes that are connected to the conclusion node set, filter target evidence nodes based on the conflict-driven field association relationship and the comparison judgment factors of the conclusion objects at both ends, and form a corresponding evidence node set; Determine the set of supporting edges between the set of conclusion nodes and the set of evidence nodes. Using the set of conclusion nodes and the set of evidence nodes as the node set, and the set of unresolved conflict edges and the set of supporting edges as the edge set, form a minimum conflict subgraph. Extract the minimum problem set used to distinguish the conclusion object from the judgment result based on the minimum conflict subgraph; Perform minimal query verification based on the minimum question set, and obtain the corresponding verification feedback information; Update the corresponding evidence strength score and conflict severity score based on the verification feedback information, and re-execute the preset adjudication rules to update the adjudication status; When the minimum conflict subgraph verification is completed and the minimum problem set is fully confirmed, or when the preset re-question termination condition is met, the final document consistency conclusion is generated and output based on all conclusion objects with the decision status set to "reserved".

2. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The process of grouping conclusion objects based on fact keys and performing conflict detection on each group to obtain conflict relationships includes: Using the same fact key as the matching rule, all conclusion objects are traversed, and conclusion objects with the same fact key are grouped into the same fact key group; Based on preset conflict rules, conclusion objects in the same fact key group are compared pairwise. All fact key groups are traversed to identify conclusion object pairs that meet the preset conflict rules and are in the same fact key group. Based on the corresponding preset conflict rules, determine the conflict type of each conclusion object pair, and extract the corresponding conflict driving field and conflict difference. Based on the corresponding conflict type, conflict-driving field, and conflict difference, the corresponding conflict relationships are formed for each conclusion object pair.

3. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The construction of the conflict graph based on evidence objects and conclusion objects includes: Based on the field content of each conclusion object, and combined with the preset support relationship judgment rules, the evidence objects are matched to obtain the associated evidence objects of each conclusion object. Using the conclusion object as the conclusion node and the evidence object as the evidence node, conflict edges are established for conflicting conclusion object pairs grouped under the same fact primary key, and support edges are established for each conclusion object and its associated evidence object. A conflict graph is constructed based on the conclusion nodes, evidence nodes, and their corresponding conflict edges and support edges.

4. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The calculation of the conflict severity score for each conflict relationship and the evidence strength score for each conclusion object through the conflict map includes: Based on preset conflict index dimensions, the association information of each conflict relationship is extracted from the conflict map. The preset conflict index dimensions include at least the scope of conflict impact, conflict difference value, conflict spread scope, and evidence missing degree. Based on the preset weights of each conflict indicator dimension, the correlation information of each conflict indicator dimension is weighted and summed to obtain the conflict severity score of each conflict relationship. Extract evidence information from all evidence objects associated with each conclusion object from the conflict graph. The evidence information includes at least the completeness of the evidence location field, the degree to which the evidence content covers the conclusion field, the level of evidence source, the consistency of evidence, and the degree of matching of evidence version. Based on preset rules, the evidence information corresponding to each conclusion object is scored item by item, and the scores are weighted and summed according to the preset weights of the evidence items to obtain the evidence strength score of each conclusion object.

5. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The process involves making a definitive ruling on the conclusion object corresponding to each conflict relationship based on preset ruling rules, combined with corresponding conflict severity scores and evidence strength scores, and obtaining the corresponding ruling status, including: Based on the preset adjudication rules, and in combination with the corresponding conflict severity score and evidence strength score, the conclusion object pairs corresponding to each conflict relationship are verified according to the preset verification priority order, which is in the following order: hard rule verification, minimum evidence set verification, evidence strength score comparison verification, and source level and version verification. If the hard rule validation fails, set the ruling status of the corresponding conclusion object to the rejection status; When passing the hard rule verification, the next verification is executed according to the preset verification priority order; After one of the verifications, if the conclusion object pair does not meet the corresponding equal judgment conditions, select the conclusion object to be retained from the conclusion object pair according to the corresponding verification results, set the adjudication status of the selected conclusion object to the retained status, and set the adjudication status of the other conclusion object in the conclusion object pair to the rejection status. If the conclusion object pair meets the corresponding equal judgment condition, the next verification will continue until all verifications are completed. If the conclusion object pair meets the equal judgment condition corresponding to each verification, the adjudication status of the corresponding conclusion object will be set to the pending verification status.

6. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The extraction of the minimum problem set based on the minimum conflict subgraph for distinguishing the conclusion object from the judgment result includes: Extract the conflict type and conflict driving field corresponding to each unresolved conflict edge from the unresolved conflict edges corresponding to the minimum conflict subgraph; Based on the extracted conflict type and conflict driving fields, a corresponding verification task template is generated; Perform a merging process on the generated verification task templates to obtain the minimum set of questions used to distinguish the conclusion objects from the judgment results.

7. The document consistency adjudication method based on conflict graphs and evidence strength according to claim 1, characterized in that, The process of segmenting documents and normalizing fields in the document set and the preset rule set to generate evidence objects and conclusion objects containing fact primary keys includes: The document collection is divided into corresponding document units according to data type, and the fields of each document unit are normalized according to a set of preset rules to generate evidence objects; For each document unit, candidate conclusions are extracted, and fact keys are generated by combining them with preset tuple rules to form a conclusion object containing the fact keys.

8. A document consistency adjudication apparatus based on conflict graphs and evidence strength, used to execute the document consistency adjudication method according to any one of claims 1 to 7, characterized in that, include: The document processing module is used to perform document unit segmentation and field normalization on the document collection and the preset rule collection, generate evidence objects and conclusion objects containing fact keys, group the conclusion objects according to the fact keys, and perform conflict detection on each group of conclusion objects to obtain conflict relationships. The conflict graph generation module is used to construct conflict graphs based on evidence objects and conclusion objects; The document deterministic adjudication module is used to calculate the conflict severity score and the evidence strength score of each conclusion object through the conflict graph. Based on the preset adjudication rules, it performs a deterministic adjudication on the conclusion object corresponding to each conflict relationship, obtains the corresponding adjudication status, and outputs the final document consistency conclusion based on the adjudication status. The minimize verification module is used to minimize the corresponding conflict relationship when a conclusion object with a judgment status pending verification is identified, until the preset re-questioning termination condition is met.

Images