Knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method and device

Abstract

This application proposes a knowledge graph-based bidirectional linkage visualization method and apparatus for evidence chains and reasoning logic, relating to the field of natural language processing. The method includes: structurally decomposing the input question using a semantic big data model to generate evidence triples; constructing an initial evidence chain based on the evidence triples; generating reasoning nodes using the semantic big data model's logic reasoning engine based on the initial evidence chain; constructing an initial reasoning logic path based on the reasoning nodes; constructing a bidirectional linkage mapping relationship between the evidence triples and the reasoning nodes based on the knowledge graph; performing factual detection and targeted correction on the initial evidence chain based on the bidirectional linkage mapping relationship to obtain a corrected evidence chain; performing validity detection and collaborative correction on the initial reasoning logic path based on the bidirectional linkage mapping relationship to obtain a corrected reasoning logic path; and visualizing the corrected evidence chain, the corrected reasoning logic path, and the bidirectional linkage mapping relationship through a hierarchical visualization system to obtain visualized query results.

Description

Technical Field

[0001] This application relates to the field of natural language processing, and in particular to a method and apparatus for visualizing the bidirectional linkage between knowledge graph evidence chain and reasoning logic. Background Technology

[0002] With the widespread application of semantic analysis and knowledge graph technologies in intelligent querying and logical reasoning, users' demands for structured parsing of unstructured problems, traceability of reasoning processes, and visualization of results are increasing. Existing technologies, when handling related tasks, suffer from low reliability of reasoning results and insufficient accuracy in correcting errors in evidence chains and reasoning logic paths, as detailed below.

[0003] Existing methods construct evidence chains and reasoning logic paths by making reasoning nodes and evidence triples independent of each other, lacking an effective association mechanism. This results in insufficient interpretability of the reasoning process, an inability to clearly define the factual basis for the reasoning conclusions, and a reduction in the credibility of the reasoning results.

[0004] At the error correction level, existing technologies mostly use global verification to correct the evidence chain and reasoning path, without establishing a precise connection between evidence and reasoning. This leads to vague error location, and the correction process requires a lot of computing power to traverse all the data. The correction efficiency is low and the accuracy is difficult to guarantee. Slight deviations are easily ignored, which in turn affects the accuracy of the overall conclusion. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a bidirectional linkage visualization method for knowledge graph evidence chains and reasoning logic, comprising:

[0006] The input question is obtained and a semantic big model is constructed. The input question is then decomposed into evidence triples using the semantic big model. An initial chain of evidence is then constructed based on the evidence triples.

[0007] Based on the initial chain of evidence, reasoning nodes are generated through the logical reasoning engine of the semantic big model, and the initial reasoning logic path is constructed based on the reasoning nodes;

[0008] Obtain the knowledge graph and construct a two-way linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph;

[0009] Based on the two-way linkage mapping relationship, the initial chain of evidence is subjected to factual detection and targeted correction to obtain a corrected chain of evidence;

[0010] Based on the bidirectional linkage mapping relationship, the initial reasoning logic path is validated and collaboratively corrected to obtain a corrected reasoning logic path;

[0011] A hierarchical visualization system is constructed to visualize the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship, thereby obtaining visualized query results.

[0012] Optionally, the step of structurally decomposing the input question using a semantic big data model to generate evidence triples, and constructing an initial evidence chain based on the evidence triples, includes:

[0013] By using a large semantic model, entity recognition, semantic parsing, and relation extraction are performed on the input question to extract the subject entity, object entity, and related relations from the input question, and to generate evidence triples in the form of subject entity-related relations-object entity.

[0014] Obtain the semantic logical order of the input question, and then connect each evidence triple in sequence according to the semantic logical order to generate an initial evidence chain.

[0015] Optionally, the step of generating reasoning nodes based on the initial chain of evidence using the logical reasoning engine of the semantic large model, and constructing an initial reasoning logic path based on the reasoning nodes, includes:

[0016] The semantic big model's logical reasoning engine performs association analysis on all evidence triples in the initial evidence chain, extracting the association features of each evidence triple.

[0017] Obtain the reasoning rule base, retrieve the reasoning nodes of each evidence triple in the reasoning rule base according to the association features, and connect each reasoning node in sequence according to the chain order of the evidence triples to generate the initial reasoning logic path.

[0018] Optionally, the step of constructing a bidirectional linkage mapping relationship between evidence triples and inference nodes based on the knowledge graph includes:

[0019] Based on the knowledge graph, a mapping benchmark data set is constructed. The evidence triples are positively mapped to the reasoning nodes through the mapping benchmark data set, and a fact support relationship is established between the evidence triples and the corresponding reasoning nodes.

[0020] The reasoning node performs a reverse mapping of the evidence triples by mapping the baseline data set, establishing a logical dependency relationship between the reasoning node and the evidence triples; the factual support relationship and the logical dependency relationship are linked and fused to obtain a two-way linkage mapping relationship.

[0021] Optionally, the step of performing factual detection and targeted correction on the initial evidence chain based on the bidirectional linkage mapping relationship to obtain a corrected evidence chain includes:

[0022] Based on the bidirectional linkage mapping relationship, each evidence triple in the initial evidence chain is subjected to basic fact detection, integrity detection, and semantic consistency detection to obtain basic errors, integrity errors, and consistency errors.

[0023] By using knowledge graphs to identify and correct basic errors, supplement data for integrity errors, and replace conflicting elements, delete semantic contradictions, and adjust matching relationships for consistency errors, a corrective evidence chain is obtained.

[0024] Optionally, the step of performing validity detection and collaborative correction on the initial reasoning logic path based on the bidirectional linkage mapping relationship to obtain a corrected reasoning logic path includes:

[0025] Based on the bidirectional linkage mapping relationship, each reasoning node in the initial reasoning logic path is subjected to rule compliance detection, evidence support validity detection, and logical coherence detection to obtain compliance errors, evidence support errors, and coherence errors.

[0026] By using knowledge graphs, we can correct the reasoning rules for compliance errors, correct the logical dependencies for evidence support errors, and correct the logical paths for coherence errors, thereby obtaining corrected reasoning logical paths.

[0027] Optionally, the hierarchical visualization system includes a fact layer, a logic layer, and a linkage layer. The evidence triples that correct the evidence chain are mapped to the fact layer according to their priority, and the reasoning nodes that correct the reasoning logic path are mapped to the logic layer according to their derivation order. The linkage layer binds the bidirectional linkage mapping relationship between the evidence triples and the reasoning nodes.

[0028] The visual query results include: evidence triples displayed in different colors; corrected reasoning logic paths displayed by visual connecting lines; clicking on an evidence triple automatically highlights the corresponding reasoning node; and clicking on a reasoning node automatically highlights the corresponding evidence triple.

[0029] This invention also provides a knowledge graph evidence chain-reasoning logic bidirectional linkage visualization device for implementing the aforementioned knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method. The device includes:

[0030] The initial evidence chain acquisition module is used to acquire the input question and build a semantic big model. The semantic big model is used to structurally decompose the input question to generate evidence triples, and the initial evidence chain is constructed based on the evidence triples.

[0031] The initial reasoning logic path acquisition module is used to generate reasoning nodes based on the initial evidence chain through the logical reasoning engine of the semantic big model, and to construct the initial reasoning logic path based on the reasoning nodes.

[0032] The two-way linkage mapping relationship acquisition module is used to acquire the knowledge graph and construct a two-way linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph;

[0033] The revised evidence chain acquisition module is used to perform factual detection and targeted correction on the initial evidence chain based on the two-way linkage mapping relationship, and obtain the revised evidence chain.

[0034] The module for obtaining the corrected reasoning logic path is used to perform validity checks and collaborative corrections on the initial reasoning logic path based on the bidirectional linkage mapping relationship, and to obtain the corrected reasoning logic path.

[0035] The visualization query result acquisition module is used to build a hierarchical visualization system. Through the hierarchical visualization system, the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship are visualized to obtain the visualization query results.

[0036] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method.

[0037] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the aforementioned knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method.

[0038] The present invention has the following beneficial effects:

[0039] 1. By decomposing the input problem into evidence triples through a semantic big data model, and constructing an initial evidence chain according to semantic logic, the efficiency of structuring unstructured problems is improved, and the cost of evidence extraction is reduced. Based on the initial evidence chain, the semantic big data model's logical reasoning engine generates reasoning nodes, and constructs an initial reasoning logic path based on the reasoning nodes, improving the interpretability of the reasoning process. Based on a knowledge graph, a two-way linkage mapping relationship is constructed between evidence triples and reasoning nodes, realizing two-way linkage and mutual verification of evidence and reasoning, and strengthening the overall reliability of evidence and reasoning. The initial evidence chain and initial reasoning logic path are corrected according to the two-way linkage mapping relationship without global verification, improving the efficiency and accuracy of correction. The corrected evidence chain, corrected reasoning logic path, and two-way linkage mapping relationship are visualized through a hierarchical visualization system, intuitively displaying the evidence chain, reasoning logic, and the relationship between the two, reducing the understanding threshold for non-technical personnel, and improving the analysis efficiency of the semantic big data model.

[0040] 2. Evidence triples provide factual support for inference nodes through forward mapping, while inference nodes provide logical dependencies for evidence triples through reverse mapping. Based on the two-way linkage, the root cause can be quickly located when errors are detected, achieving more accurate error location. Through cross-validation of the two-way linkage, even if there is a slight deviation in a certain link, it can be discovered through reverse verification of the inference logic, reducing the impact of deviation in a single link and improving the credibility of the inference results.

[0041] 3. Perform basic fact detection, integrity detection, and semantic consistency detection on the initial evidence chain to identify basic errors, integrity errors, and consistency errors; perform rule compliance detection, evidence support validity detection, and logical coherence detection on the initial reasoning logic path to identify compliance errors, evidence support errors, and coherence errors; call the corresponding resources of the knowledge graph to correct different error types without global reconstruction, thereby improving the credibility and efficiency of the correction results. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0044] Figure 2 This is a structural diagram of the device according to an embodiment of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0046] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0047] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0048] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Reference Figure 1 This invention provides a method for visualizing the bidirectional linkage between knowledge graph evidence chain and reasoning logic, comprising:

[0050] The input question is obtained and a semantic big model is constructed. The input question is then decomposed into evidence triples using the semantic big model. An initial chain of evidence is then constructed based on the evidence triples.

[0051] In some embodiments, the step of structurally decomposing the input question using a semantic large model to generate evidence triples, and constructing an initial chain of evidence based on the evidence triples, includes:

[0052] By using a large semantic model, entity recognition, semantic parsing, and relation extraction are performed on the input question to extract the subject entity, object entity, and related relations from the input question, and to generate evidence triples in the form of subject entity-related relations-object entity.

[0053] In some embodiments, a pre-trained large model adapted to Chinese semantic understanding is first selected, such as ERNIE or a fine-tuned LLaMA-2. Specific fine-tuning parameters for entity recognition and relation extraction are loaded onto the model, and a professional terminology dictionary for the target scene is imported to ensure that the model can accurately identify special entities and relations within the domain. Next, the input question is preprocessed: redundant interjections without actual semantic meaning, such as "please ask" and "whether," are removed; punctuation marks are standardized to full-width Chinese characters; if the question length exceeds the model's limit, long sentences are split according to semantic completeness, while preserving the contextual relationship between clauses during splitting.

[0054] The model employs a dual approach of automatic model recognition and rule validation to extract entities. First, the model annotates the preprocessed input question with entities labeled with [subject] and [object], along with their corresponding types. Second, it validates the validity of entities using predefined rules, removing entities irrelevant to the core semantics and merging synonymous entities. Finally, it outputs a set of main entities and labels each entity with a semantic role.

[0055] First, the syntactic structure of the input question is decomposed using a semantic large model to clarify the logical relationship types between entities. Then, based on the syntactic structure and logical types, the relationship is extracted. The model first outputs candidate relationships, and then performs standardization processing through a pre-set relationship dictionary to unify the synonyms of natural language expressions into standard terms to ensure the consistency of relationship expressions. Finally, the relationship is determined and bound to the corresponding subject entity and object entity.

[0056] Strictly follow the fixed format of subject entity-relationship-object entity to combine the bound entities and relations into triples, ensuring that each triple contains only 1 subject, 1 relationship, and 1 object, avoiding the mixing of multiple entities or multiple relations;

[0057] Obtain the semantic logical order of the input question, and connect each evidence triple in sequence according to the semantic logical order to generate an initial evidence chain;

[0058] In some embodiments, the text features of the input question are analyzed by a semantic big data model, and the logical relationship type is determined by combining keywords to determine the sequence order of the triples. Common logical types include time sequence, causal sequence, primary and secondary sequence, and parallel sequence. If the question has multiple logical combinations, the core logic is used as the main line, and the secondary logic is nested in the main line to ensure logical coherence.

[0059] The triples are linked together according to the identified semantic logical order, with the following specific rules: If adjacent triples have a common entity, they are linked together directly in logical order without additional explanation; if adjacent triples do not have a common entity, a semantic bridging explanation is required to clarify the relationship between the two; after the linking is completed, an initial evidence chain is formed, the structure of which is the sequential connection of triples; for example, if the input problem is "the user selects products and places an order, the merchant ships the goods after payment, and the user reports a problem due to logistics delay", the triple set is T001[user-selects products-goods], T002[user-places an order-order], T003[user-pays-order], T004[merchant-ships-goods], T005[user-feedback-logistics problem], and the initial evidence chain is T001→T002→T003→T004→T005.

[0060] Based on the initial chain of evidence, reasoning nodes are generated through the logical reasoning engine of the semantic big model, and the initial reasoning logic path is constructed based on the reasoning nodes;

[0061] In some embodiments, the step of generating reasoning nodes through the logical reasoning engine of the semantic large model based on the initial chain of evidence, and constructing an initial reasoning logic path based on the reasoning nodes, includes:

[0062] The semantic big model's logical reasoning engine performs association analysis on all evidence triples in the initial evidence chain, extracting the association features of each evidence triple.

[0063] In some embodiments, the three core functional modules of entity association analysis, relation association mining, and logical dependency parsing are first installed on the logical reasoning engine of the semantic big model. The parsing templates for the corresponding scenarios are imported, and the confidence threshold for association recognition is set to 0.8. At the same time, the evidence triples in the initial evidence chain are organized in a simple format of ID, subject entity, association relationship, object entity, and logical segment to which they belong, so that the engine can read them easily.

[0064] Next, the engine parses the association features of the triples. For adjacent triples, it checks whether their entities co-occur directly or indirectly, whether the relationship is action progression, condition or trigger, and whether the logical dependency is chronological or causal. For non-adjacent but related triples, it finds the global association through entity tracing and logical propagation. Then, it extracts these association features into a structure containing feature ID, association range, entity, relationship, and logical features, and checks whether it is consistent with the logic of the initial evidence chain. If it is inconsistent, it is parsed again.

[0065] Obtain the reasoning rule base, retrieve the reasoning nodes of each evidence triple in the reasoning rule base according to the association features, and connect each reasoning node in sequence according to the chain order of the evidence triple to generate the initial reasoning logic path;

[0066] In some embodiments, the inference rule base is divided into general rules and scenario-specific rules, which are stored according to rule ID, conditions, conclusion templates, and confidence thresholds. After being imported into the logic inference engine, an index is built according to feature type to facilitate matching. During matching, candidate rules are first searched according to the core conditions of the features, and then it is checked whether the association strength is sufficient for the rule's threshold. If it is sufficient, it is kept; otherwise, it is discarded. If there are multiple candidate rules, scenario-specific rules are used first. After matching, the feature information is fitted into the conclusion template of the rule to generate inference nodes, and information such as node ID, inference basis, and confidence level are added.

[0067] Then, the inference nodes are arranged in the order of the evidence triples of the initial evidence chain. For example, node N001 corresponding to T001-T002 is placed before node N002 corresponding to T002-T003. Adjacent inference nodes are marked as premise dependencies, same logical segments, or interruptions according to their logical relationships. Finally, the initial inference logic path is output in the format of [connection type] node ID.

[0068] Obtain the knowledge graph and construct a two-way linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph;

[0069] In some embodiments, constructing a bidirectional linkage mapping relationship between evidence triples and inference nodes based on a knowledge graph includes:

[0070] Based on the knowledge graph, a mapping benchmark data set is constructed. The evidence triples are positively mapped to the reasoning nodes through the mapping benchmark data set, and a fact support relationship is established between the evidence triples and the corresponding reasoning nodes.

[0071] In some embodiments, core data is extracted from the knowledge graph as a baseline template for mapping. The entity library and relation library in the knowledge graph are extracted, and the common reasoning logic corresponding to entities and relations in the knowledge graph is recorded and organized into an entity-relationship-reasoning rule association table. The data is standardized according to the structure of baseline item ID, entity information, relation information, corresponding reasoning rule identifier, and association weight to form a structured mapping baseline data set, which is convenient for subsequent matching and calling.

[0072] The fact support score between the evidence triples and the reasoning nodes is calculated based on the mapping baseline dataset. The calculation formula is as follows:

[0073]

[0074] in, For evidence triples, , As the main entity, It is a related relationship. As an object entity, The base confidence level for the evidence triplet t; The reasoning rule for the i-th reasoning node; Score based on facts; Score for type consistency; Score for word matching; Score the structure matching. Score for time consistency; This is the prior template for the i-th inference node; Non-negative weights; when At that time, establish factual support relationships between evidence triples and reasoning nodes;

[0075] The reasoning node performs a reverse mapping of the evidence triples by mapping the baseline data set, establishing a logical dependency between the reasoning node and the evidence triples; the factual support relationship and the logical dependency relationship are linked and fused to obtain a two-way linkage mapping relationship;

[0076] The logical dependency score between inference nodes and evidence triples is calculated based on the mapping baseline dataset. The calculation formula is as follows:

[0077]

[0078] in, This represents an instantiated inference node; For the j-th evidence triple; Scoring based on logical dependencies; This serves as evidence of the fit between the node and the required slot. For consistency score; The proportion of nodes covered by evidence; For the evidence triple The baseline confidence level; To restrain punishment; Non-negative weights; when Establish logical dependencies in a timely manner;

[0079] The bidirectional linkage score is calculated based on the factual support score and the logical dependency score, expressed as follows:

[0080]

[0081]

[0082] in, Indicates the evidence triple With reasoning nodes The two-way linkage score; To achieve the harmonic mean, Score based on facts. Scoring based on logical dependencies It is a numerically stable term; The intensity of the conflict; This is the conflict penalty coefficient; For the consensus gain of multiple evidences; This is the redundancy gain coefficient; when At that time, in the evidence triad With reasoning nodes A two-way linkage mapping relationship is established between them.

[0083] Based on the two-way linkage mapping relationship, the initial chain of evidence is subjected to factual detection and targeted correction to obtain a corrected chain of evidence;

[0084] In some embodiments, the step of performing fact-checking and targeted correction on the initial evidence chain based on the bidirectional linkage mapping relationship to obtain a corrected evidence chain includes:

[0085] Based on the bidirectional linkage mapping relationship, each evidence triple in the initial evidence chain is subjected to basic fact detection, integrity detection, and semantic consistency detection to obtain basic errors, integrity errors, and consistency errors.

[0086] In some embodiments, the detection process takes a two-way linkage mapping relationship as the core link, and combines the entity library, relation library and rule library of the knowledge graph to perform basic fact detection, integrity detection and semantic consistency detection on each evidence triple in sequence, and generate error type labels and detailed descriptions;

[0087] The basic fact detection based on knowledge graphs is carried out in the following steps: First, the knowledge graph entity database is queried to verify whether the subject entity S and the object entity O exist and whether their attributes are valid. Second, the knowledge graph relation database is queried to verify whether the relationship R truly exists between S and O and whether it conforms to the domain rules. Finally, basic error results are output, specifying the error type, error location, and verification basis in the knowledge graph.

[0088] When performing integrity checks, the integrity judgment criteria are first defined. Based on the target scenario rule base, the required attributes corresponding to each relation R are clarified. When adjacent triples have logical associations, they must have common association entities or clear semantic bridging information. Then, integrity verification is performed: First, check whether the S and O of the triples are missing required attributes; Second, combine the logical dependency score LD of the bidirectional linkage mapping relationship. If the LD value is less than the threshold τ_l1, it is determined that the association information is missing; Third, verify whether the triples contain core semantic elements; Finally, output the integrity error results, clarifying the error type, the name of the missing item, and the source of supplementary evidence.

[0089] When performing semantic consistency detection, the consistency verification dimensions are first set, including entity consistency, relation consistency, and temporal and causal consistency. Then, multi-dimensional verification is performed: First, the representation of the same entity in all triples is compared globally. Second, the adaptability of relations and entities is verified. Third, the semantic logic is verified through the knowledge graph rule base. Finally, consistency error results are output, specifying the error type, conflict location, and conflict description.

[0090] By using knowledge graphs to identify and correct basic errors, supplement data for integrity errors, and replace conflicting elements, delete semantic contradictions, and adjust matching relationships for consistency errors, a corrective evidence chain is obtained.

[0091] In some embodiments, the correction process is supported by knowledge graphs, and differentiated correction operations are performed for different error types. After correction, a second verification is required through bidirectional linkage mapping to ensure that the correction results meet the requirements of facts and logic.

[0092] When performing basic error correction, the error identification process begins by adding a red error label to detected basic error triples, indicating the error type and verification basis, and pausing the logical transmission of the triple in the evidence chain. Then, targeted correction is performed: First, entity error correction: the correct entity with the highest matching degree is retrieved from the knowledge graph entity database. If no replacement entity is available, it is marked as an entity to be supplemented and manual review is triggered. Second, relation error correction: based on the knowledge graph relation database, illegal relations are replaced with standard relations suitable for S and O. Third, post-correction verification: the fact support score of the triple is recalculated, requiring FS ≥ threshold τ_f2; otherwise, correction is repeated.

[0093] When correcting integrity errors, the process begins with attribute supplementation. For missing attribute errors, the necessary attribute data corresponding to the entity is extracted from the knowledge graph, and the data source is annotated after supplementation. Next, association information is supplemented: for the problem of adjacent triples having no common entity, semantic bridging statements are supplemented based on the semantic association rules of the knowledge graph. The bridging statements must conform to the semantic logical order within the chain. Finally, core element supplementation is performed: if core semantic elements are missing, data is retrieved from the entity-relationship-core element association table of the knowledge graph to ensure that the supplemented triples have complete semantic expression.

[0094] When correcting consistency errors, the following steps are taken: First, conflicting elements are replaced: For entity conflicts, non-standard expressions are replaced with unified entity identifiers based on the entity synonym mapping table of the knowledge graph; for relationship conflicts, they are replaced with standard relationships that are compatible with the entity type and the logic within the chain. Next, semantic conflict deletion is performed: If there are irreconcilable semantic conflicts in triples, triples with higher degree of conflict are deleted. Before deletion, it is necessary to verify whether the remaining triples can still support the logic within the chain. If it affects the logical coherence, manual intervention is triggered. Finally, matching relationship adjustment is performed: For inconsistencies in time and causality, the order of triples is adjusted based on the temporal rule base of the knowledge graph, or the time attribute is corrected. After adjustment, the semantic consistency within the chain is verified to ensure that there are no new logical conflicts.

[0095] All corrected triples are reconnected in their original semantic and logical order, and the correction records are retained to form an initial corrected evidence chain. Based on the bidirectional linkage mapping relationship, the fact support score (FS) and logical dependency score (LD) of each corrected triple are recalculated, requiring FS ≥ 0.85 and LD ≥ 0.7; otherwise, the correction is returned. The evidence chain integrity rule base of the knowledge graph is used to verify whether the corrected evidence chain has complete semantic logic and no missing core links. After the verification is passed, the formal corrected evidence chain is output, and a correction report is generated simultaneously for traceability.

[0096] Based on the bidirectional linkage mapping relationship, the initial reasoning logic path is validated and collaboratively corrected to obtain a corrected reasoning logic path;

[0097] In some embodiments, the step of performing validity detection and collaborative correction on the initial reasoning logic path based on the bidirectional linkage mapping relationship to obtain a corrected reasoning logic path includes:

[0098] Based on the bidirectional linkage mapping relationship, each reasoning node in the initial reasoning logic path is subjected to rule compliance detection, evidence support validity detection, and logical coherence detection to obtain compliance errors, evidence support errors, and coherence errors.

[0099] In some embodiments, when performing rule compliance checks, the reasoning basis of reasoning node n_i is found based on the bidirectional linkage mapping relationship, including the rule ID used and the instantiation content of the rule conditions. Then, the standard definition corresponding to the rule ID is retrieved from the knowledge graph reasoning rule base. Next, rule conditions are compared: if the instantiation preconditions of n_i are fewer than the standard rules, it is marked as a condition missing compliance error; if the instantiation preconditions conflict with the standard rules, it is marked as a condition conflict compliance error; if the rule ID has no corresponding record in the knowledge graph reasoning rule base, it is marked as a rule non-existent compliance error. Finally, confidence is checked: the reasoning confidence of n_i is calculated. If the result is less than the threshold τ_rule, it is marked as an insufficient confidence compliance error. After the detection is completed, the compliance error result is output, specifying the error type, the node ID of n_i, the error location, and the specific reason.

[0100] The criteria for valid evidence support include the logical dependency score (LD) between n_i and evidence triplet t_j in the bidirectional linkage mapping relationship, the evidence-node binding relationship, and the correction state of t_j in the knowledge graph.

[0101] The first step in the detection process is the location of associated evidence. This involves extracting all t_j bound to n_i from the bidirectional linkage mapping relationship, excluding t_j that have been marked as invalid. The second step is the LD value threshold verification: a threshold τ_LD is set for the validity of evidence support. If the LD (n_i→t_j) of all associated t_j is less than τ_LD, it is marked as an error of no valid support. If the LD value of some t_j is greater than or equal to τ_LD, but the number of valid t_j is less than 2, it is marked as an error of insufficient support. The third step is a secondary verification of the validity of evidence: even if the LD value of t_j meets the threshold, if t_j has an uncorrected underlying error, it is still marked as an error of invalid evidence. Finally, the results of the evidence support error are output, specifying the error type, the node ID of n_i, the list of associated evidence, and the reason for the error.

[0102] The logic coherence detection is based on the association features between reasoning nodes in the bidirectional linkage mapping relationship, as well as the logic coherence rule base of the knowledge graph.

[0103] The detection process begins with a node sequence and association feature verification: following the node order of the initial reasoning logical path, the association features of adjacent nodes are analyzed. If the conclusion of n_i is not related to the premise of n_{i+1}, it is marked as a connection discontinuity error; if there is a causal or temporal conflict between n_i and n_{i+1}, it is marked as a logical conflict coherence error. Next, a global logical loop closure verification is performed: the logical association between the first and last nodes of the path is checked. If the premise of the first node and the conclusion of the last node do not have a global logical loop closure, it is marked as a global logical deficiency coherence error. Finally, a cross-node evidence consistency verification is performed: if multiple reasoning nodes are associated with the same t_j, the interpretation of t_j by each node is checked for consistency. If inconsistent, it is marked as an evidence interpretation conflict coherence error. After the detection is completed, the coherence error results are output, specifying the error type, conflicting node ID, inter-node association features, and standard logical rules.

[0104] By using knowledge graphs, we can correct the reasoning rules for compliance errors, correct the logical dependencies for evidence support errors, and correct the logical paths for coherence errors, thereby obtaining corrected reasoning logical paths.

[0105] In some embodiments, the correction of inference rules for compliance errors follows the principle of replacing erroneous rules, supplementing missing conditions, and ensuring that the corrected nodes conform to standard rules and achieve the required confidence level. The operations for different error types are as follows: For rule non-existence errors, first extract the inference premises and conclusions of n_i, then find the rule with the closest premise-conclusion match in the knowledge graph inference rule base using a rule similarity matching algorithm, replace the rule ID of n_i with the matching rule ID, and re-instantiate the premises; For condition missing errors, first clarify the missing premises based on standard rules, then search for t_j that supports the condition in the evidence-node association library of the bidirectional linkage mapping relationship, supplement t_j to the premises of n_i, and update the inference basis; For condition conflict errors, first delete the conflicting premises, then retrieve t_j that conforms to the standard rule premises from the valid evidence library of the knowledge graph, replace it with the correct premise, and recalculate the inference confidence level; For insufficient confidence errors, increase the number of associated evidence triples, or if the rule itself has low confidence, replace it with a rule with higher confidence in the same scenario.

[0106] The logical dependency correction for evidence support errors follows the principle of strengthening evidence support, correcting logical dependencies, and ensuring that the LD values ​​of n_i and t_j meet the standard and that t_j is valid evidence. For errors with no valid support or insufficient support, the evidence-node association library of the knowledge graph is used to find t_new that matches the reasoning logic of n_i and has an LD value ≥ τ_LD. t_new is then bound to n_i, and the logical dependency in the bidirectional linkage mapping is updated. If no matching t_new is found, manual supplementation of evidence is triggered. For invalid evidence errors, invalid t_j associated with n_i is first deleted. Then, the semantically identical and corrected t_corrected is retrieved from the corrected evidence chain of the knowledge graph. The LD values ​​of n_i and t_corrected are recalculated, and the binding relationship is updated after ensuring that the values ​​are ≥ τ_LD. After correction, the number of valid associated evidences of n_i is verified to be ≥ 2, and the LD values ​​of all associated t_j are ≥ τ_LD. Otherwise, supplementation or replacement of evidence continues.

[0107] The logical path correction for inconsistency errors follows the principles of adjusting node order, supplementing intermediate nodes, and deleting conflicting nodes to ensure smooth logical connections between nodes and compliance with global rules. For inconsistency errors caused by connection gaps, based on the knowledge graph's logical consistency rule base, the intermediate reasoning node n_mid that needs to be supplemented at the gap is identified, an association t_j is matched for n_mid, inserted between n_i and n_{i+1} in the path, and the node order is updated. For inconsistency errors caused by logical conflicts, the temporal and causal rules in the knowledge graph are retrieved, and conflicting nodes are reordered according to the rules. If conflicts still exist after reordering, logically contradictory nodes are deleted. For inconsistency errors caused by missing global logic, based on the logic of the first and last nodes of the path, the missing core node is found in the knowledge graph's logical closed-loop rule base. Evidence is matched to supplement the node, and a complete logical chain is constructed. For inconsistency errors caused by conflicting evidence interpretation, the standard semantic interpretation of t_j in the knowledge graph is retrieved, the interpretation of t_j by all associated nodes is unified, and the reasoning basis of the nodes is corrected. After correction, local verification is performed first, followed by global verification.

[0108] Reconnect all the corrected inference nodes in the order they passed the verification, and retain the correction record of each node to form a draft of the corrected inference logic path; then perform a two-way linkage overall verification: based on the two-way linkage mapping relationship, recalculate the FS value, LD value, and Link value of all nodes in the path, requiring that the FS value of all nodes is ≥0.85, LD value is ≥0.7, and Link value is ≥0.75, and the average Link value between nodes is ≥0.8; after the verification passes, output the formal corrected inference logic path.

[0109] A hierarchical visualization system is constructed to visualize the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship, thereby obtaining visualized query results.

[0110] In some embodiments, the hierarchical visualization system includes a fact layer, a logic layer, and a linkage layer. Evidence triples that correct the evidence chain are mapped to the fact layer according to their priority, and inference nodes that correct the reasoning logic path are mapped to the logic layer according to their derivation order. The linkage layer binds the bidirectional linkage mapping relationship between evidence triples and inference nodes.

[0111] The visual query results include: evidence triples displayed in different colors; corrected reasoning logic paths displayed by visual connecting lines; clicking on an evidence triple automatically highlights the corresponding reasoning node; and clicking on a reasoning node automatically highlights the corresponding evidence triple.

[0112] In some embodiments, the layered visualization system adopts an architecture design of front-end layer overlay + back-end data linkage. The front-end realizes independent rendering and overlay display of three layers through the layer management module, and supports users to manually switch the visibility of each layer. The back-end provides structured data through data interfaces, including corrected evidence chains, corrected reasoning logic paths, and bidirectional linkage mapping relationships. All data is transmitted in JSON format, adapting to different front-end technology selections. For lightweight scenarios, ECharts can be used to quickly build charts, while for complex customized scenarios, D3.js can be used to implement custom node and connection styles. For native knowledge graph scenarios, Neo4jBloom can be directly connected to simplify data calls.

[0113] To visually represent the different correction states of t_corr, a design combining color coding and texture differentiation is adopted: triples without errors are represented by dark green, textureless nodes; basic error corrections are represented by sky blue diagonal thin stripe nodes; integrity error corrections are represented by orange horizontal thin stripe nodes; consistency error corrections are represented by purple vertical thin stripe nodes; and nodes awaiting manual verification are represented by light yellow grid-patterned nodes. A color legend is added to the lower right corner of the interface, including color blocks, texture examples, and status descriptions.

[0114] In addition to distinguishing styles by type, the connection lines of the logic layer inference path also reflect the inference confidence score_n through color depth: the connection line color is darkest when score_n≥0.9, medium when 0.8≤score_n<0.9, and lightest when 0.7≤score_n<0.8. An inference type filter box is added to the top of the interface, allowing users to select the desired inference type. After selection, only the connection lines and associated nodes of the corresponding type are displayed, while nodes and connection lines of other types are semi-transparent, making it easy to focus on specific inference logic.

[0115] When the evidence triple t_corr in the fact layer is clicked, the border of the node turns red and increases in thickness to 5px, and is slightly enlarged by 1.1 times. At the same time, based on the association table, all associated inference nodes n_corr are found, and their borders turn red and increase in thickness to 4px, and gray dashed linkage lines are dynamically displayed. If t_corr is associated with multiple n_corr, all associated nodes are highlighted simultaneously, and a label for the number of associated evidence is displayed next to each n_corr.

[0116] When you click on the inference node n_corr in the logic layer, the node's fill color changes to light red and it is slightly enlarged by 1.1 times. At the same time, all associated t_corrs are found, their fill colors are changed to light red, and gray dashed linkage lines are dynamically displayed. Associated t_corrs are highlighted differently according to priority to help identify core supporting evidence.

[0117] Clicking on a blank area of ​​the interface or the reset highlight button at the top can restore the original style of all nodes and hide the linkage lines; it supports holding down the Ctrl key and clicking on multiple nodes to trigger highlighting in batches, meeting the needs of correlation analysis of multiple evidence and multiple nodes.

[0118] Reference Figure 2 This invention provides a knowledge graph evidence chain-reasoning logic bidirectional linkage visualization device 20, used to realize a knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method. The device includes:

[0119] The initial evidence chain acquisition module 21 is used to acquire the input question and build a semantic big model. The input question is decomposed in a structured manner through the semantic big model to generate evidence triples, and the initial evidence chain is constructed based on the evidence triples.

[0120] The initial reasoning logic path acquisition module 22 is used to generate reasoning nodes through the logic reasoning engine of the semantic big model based on the initial evidence chain, and to construct the initial reasoning logic path based on the reasoning nodes.

[0121] The bidirectional linkage mapping relationship acquisition module 23 is used to acquire the knowledge graph and construct a bidirectional linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph.

[0122] The modified evidence chain acquisition module 24 is used to perform factual detection and targeted correction on the initial evidence chain according to the two-way linkage mapping relationship to obtain the modified evidence chain;

[0123] The module 25 for obtaining the corrected reasoning logic path is used to perform validity detection and collaborative correction on the initial reasoning logic path according to the bidirectional linkage mapping relationship, so as to obtain the corrected reasoning logic path.

[0124] The visualization query result acquisition module 26 is used to construct a hierarchical visualization system. Through the hierarchical visualization system, the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship are visualized to obtain the visualization query results.

[0125] This application provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method of any of the above schemes.

[0126] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0127] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0128] This application also provides a computer-readable medium storing a computer program thereon, which, when executed by a processor, implements the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method of any of the above schemes. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0129] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0130] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method, characterized in that, include: The input question is obtained and a semantic big model is constructed. The input question is then decomposed into evidence triples using the semantic big model. An initial chain of evidence is then constructed based on the evidence triples. Based on the initial chain of evidence, reasoning nodes are generated through the logical reasoning engine of the semantic big model, and the initial reasoning logic path is constructed based on the reasoning nodes; Obtain the knowledge graph and construct a two-way linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph; Based on the two-way linkage mapping relationship, the initial chain of evidence is subjected to factual detection and targeted correction to obtain a corrected chain of evidence; Based on the bidirectional linkage mapping relationship, the initial reasoning logic path is validated and collaboratively corrected to obtain a corrected reasoning logic path; A hierarchical visualization system is constructed to visualize the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship, thereby obtaining visualized query results.

2. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that, The process involves structurally decomposing the input question using a large semantic model to generate evidence triples, and constructing an initial chain of evidence based on these triples, including: By using a large semantic model, entity recognition, semantic parsing, and relation extraction are performed on the input question to extract the subject entity, object entity, and related relations from the input question, and to generate evidence triples in the form of subject entity-related relations-object entity. Obtain the semantic logical order of the input question, and then connect each evidence triple in sequence according to the semantic logical order to generate an initial evidence chain.

3. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that, The process of generating reasoning nodes based on the initial chain of evidence using the semantic big data model's logical reasoning engine, and constructing an initial reasoning logic path based on these nodes, includes: The semantic big model's logical reasoning engine performs association analysis on all evidence triples in the initial evidence chain, extracting the association features of each evidence triple. Obtain the reasoning rule base, retrieve the reasoning nodes of each evidence triple in the reasoning rule base according to the association features, and connect each reasoning node in sequence according to the chain order of the evidence triples to generate the initial reasoning logic path.

4. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that, The construction of a bidirectional linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph includes: Based on the knowledge graph, a mapping benchmark data set is constructed. The evidence triples are positively mapped to the reasoning nodes through the mapping benchmark data set, and a fact support relationship is established between the evidence triples and the corresponding reasoning nodes. The reasoning node performs a reverse mapping of the evidence triples by mapping the baseline data set, establishing a logical dependency relationship between the reasoning node and the evidence triples; the factual support relationship and the logical dependency relationship are linked and fused to obtain a two-way linkage mapping relationship.

5. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that, The step of performing fact-checking and targeted correction on the initial evidence chain based on the bidirectional linkage mapping relationship to obtain a corrected evidence chain includes: Based on the bidirectional linkage mapping relationship, each evidence triple in the initial evidence chain is subjected to basic fact detection, integrity detection, and semantic consistency detection to obtain basic errors, integrity errors, and consistency errors. By using knowledge graphs to identify and correct basic errors, supplement data for integrity errors, and replace conflicting elements, delete semantic contradictions, and adjust matching relationships for consistency errors, a corrective evidence chain is obtained.

6. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that, The step of performing validity checks and collaborative corrections on the initial reasoning logic path based on the bidirectional linkage mapping relationship to obtain a corrected reasoning logic path includes: Based on the bidirectional linkage mapping relationship, each reasoning node in the initial reasoning logic path is subjected to rule compliance detection, evidence support validity detection, and logical coherence detection to obtain compliance errors, evidence support errors, and coherence errors. By using knowledge graphs, we can correct the reasoning rules for compliance errors, correct the logical dependencies for evidence support errors, and correct the logical paths for coherence errors, thereby obtaining corrected reasoning logical paths.

7. The knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method according to claim 1, characterized in that: The hierarchical visualization system includes a fact layer, a logic layer, and a linkage layer. The evidence triples that correct the evidence chain are mapped to the fact layer according to their priority, and the reasoning nodes that correct the reasoning logic path are mapped to the logic layer according to their derivation order. The linkage layer binds the bidirectional linkage mapping relationship between the evidence triples and the reasoning nodes. The visual query results include: evidence triples displayed in different colors; corrected reasoning logic paths displayed by visual connecting lines; clicking on an evidence triple automatically highlights the corresponding reasoning node; and clicking on a reasoning node automatically highlights the corresponding evidence triple.

8. A knowledge graph evidence chain-reasoning logic bidirectional linkage visualization device, used to implement the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method as described in any one of claims 1 to 7, characterized in that, The device includes: The initial evidence chain acquisition module is used to acquire the input question and build a semantic big model. The semantic big model is used to structurally decompose the input question to generate evidence triples, and the initial evidence chain is constructed based on the evidence triples. The initial reasoning logic path acquisition module is used to generate reasoning nodes based on the initial evidence chain through the logical reasoning engine of the semantic big model, and to construct the initial reasoning logic path based on the reasoning nodes. The two-way linkage mapping relationship acquisition module is used to acquire the knowledge graph and construct a two-way linkage mapping relationship between evidence triples and reasoning nodes based on the knowledge graph; The revised evidence chain acquisition module is used to perform factual detection and targeted correction on the initial evidence chain based on the two-way linkage mapping relationship, and obtain the revised evidence chain. The module for obtaining the corrected reasoning logic path is used to perform validity checks and collaborative corrections on the initial reasoning logic path based on the bidirectional linkage mapping relationship, and to obtain the corrected reasoning logic path. The visualization query result acquisition module is used to build a hierarchical visualization system. Through the hierarchical visualization system, the corrected evidence chain, the corrected reasoning logic path, and the two-way linkage mapping relationship are visualized to obtain the visualization query results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the knowledge graph evidence chain-reasoning logic bidirectional linkage visualization method as described in any one of claims 1 to 7.

Images