An archival knowledge graph credible question and answer method and system

Abstract

This invention discloses a reliable question-answering method and system for archival knowledge graphs, relating to the field of knowledge graph processing technology. After receiving an archival query, the method parses the subject entities, relational clues, and constraint sets, constructs an initial evidence subgraph in the archival knowledge graph, and writes it into a graph evidence cache. Based on the initial evidence subgraph, candidate answers are inferred, and an uncertainty score is calculated by combining confidence, multi-sample consistency, support paths, relation coverage, constraint coverage, and candidate competition. When the uncertainty exceeds a threshold, a structured diagnosis of the failure cause is performed, identifying it as an entity gap, relation gap, constraint gap, or support gap. Graph evidence is incrementally supplemented according to the failure type, and the incremental evidence is merged into the graph evidence cache. After further inference and support verification, a reliable answer with evidence sources is output. This scheme can reduce the noise introduced by blindly expanding the scope of the query, significantly improving the reliability, accuracy, and traceability of archival question answering.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph processing technology, and in particular to a reliable question-answering method and system based on archival knowledge graphs, including uncertainty triggering, structured failure diagnosis, and targeted incremental repair. Background Technology

[0002] Archival management departments face numerous demands for intelligent question-and-answer applications based on archival content in scenarios such as government services, collection utilization, historical event verification, thematic archive retrieval, and cross-file correlation analysis. Users often wish to query information such as the formation time of a specific event, the responsible party, the approval basis, related document numbers, the change process, upstream and downstream processing links, and the source of evidence. Traditional archival retrieval relies primarily on keyword matching, catalog search, or manual consultation, which suffers from problems such as long retrieval links, weak cross-file correlation capabilities, and poor interpretability of results.

[0003] With the development of large language models, retrieval-enhanced generative techniques, and knowledge graph technologies, existing solutions attempt to use archival text, archival catalog data, or knowledge graphs as external knowledge sources. After receiving a user's question, these solutions first perform a retrieval, and then input the retrieval results into a generative model to output the answer. While these solutions can improve question-answering efficiency to some extent, they still have limitations in reliable archival question-answering scenarios.

[0004] First, archival Q&A differs from general open-domain Q&A, demanding higher standards of authenticity, traceability, consistency, and constraint compliance. Archival answers cannot merely be linguistically sound; they must also pinpoint the relevant archival entry, file relationship, creation date, responsible department, document number, or upstream / downstream processing stage. If the system directly outputs an answer when evidence is insufficient, it can easily lead to distorted responses, severely impacting the reliability of archival access.

[0005] Second, existing technologies typically employ a uniform approach of broadening the search scope, repeated recall, or re-inference of the entire package when responses fail, lacking a detailed analysis of the reasons for failure. For example, unreliable answers may stem from a lack of file alias mapping, cross-file relationships, time constraints, or departmental relationships, or from scattered evidence that fails to form a complete supporting chain. If a uniform broad search is conducted without distinguishing the reasons for failure, irrelevant file fragments, noise edges, and redundant paths are easily introduced.

[0006] Third, archival question-and-answer systems exhibit distinct structured constraints. Many questions involve not only entities but also constraints such as chronological order, document number correspondence, institutional hierarchy, subject affiliation, processing procedures, and version changes. While existing knowledge graph-enhanced question-and-answer solutions can perform graph retrieval and generate answers, they typically lack a complete closed loop encompassing evidence coverage, uncertainty triggering, failure type diagnosis, targeted supplementary evidence, and result verification. This makes it difficult to meet the core requirements of accuracy and auditability in credible archival question-and-answer scenarios.

[0007] Therefore, a technical solution is needed for the credible question-answering scenario of archives. This solution can first construct a small-scale initial evidence subgraph on the archive knowledge graph and complete the first answer. When the current answer is detected to have evidence gaps or instability, the uncertainty is calculated and the cause of failure is diagnosed. Then, necessary archive graph evidence is supplemented in a targeted manner according to the failure type. Finally, based on the repaired evidence subgraph, the solution is re-inferred and a credible answer with supporting evidence is output. Summary of the Invention

[0008] This invention aims to address the problem that existing archival question-answering systems may still output answers without archival basis when initial retrieval evidence is insufficient. To achieve the above objective, this invention provides a reliable question-answering method based on archival knowledge graphs, which specifically includes the following steps: Receive the user's file query question q, perform semantic parsing on the question, and obtain a set of subject entities. Relationship Clues Collection and constraint sets Subject entities can include file name, item name, organization name, document number, personnel, location, and time; relationship clues can include fields such as "formed at", "belongs to", "based on", "handling department", "superior authority", "changed to", and "related file"; constraint sets can include time constraints, document number constraints, organization level constraints, file category constraints, version constraints, sorting constraints, or comparison constraints.

[0009] based on , and An initial retrieval is performed in the archival knowledge graph to construct an initial evidence subgraph. The initial search may include entity neighborhood expansion search, relation focusing search, constraint attribute search, and bridging path search, to enable... To satisfy the minimum necessary subgraph required for the first inference, thereby controlling the cue length and the first response delay.

[0010] The minimum necessary subgraph refers to the smallest graph structure selected from the archival knowledge graph that supports the initial inference, based on the subject entities, relational clues, and constraints in the user profile query question. In other words, when answering the user question for the first time, the entire archival knowledge graph is not retrieved, nor is the search scope blindly expanded; instead, only the smallest possible evidence subgraph necessary to complete the initial inference is extracted. This subgraph prioritizes retaining entities, relations, attributes, and evidence source nodes that are directly related to the question, have clear sources, strong constraint matching, and strong path support. This reduces the probability of irrelevant graph data entering the inference context and decreases retrieval noise, prompt length, and initial response latency.

[0011] The initial evidence subgraph Write the graph evidence cache M. The cache includes at least a TripleStore, a PathStore, an AliasStore, a ConstraintStore, and a VerificationStore. Each file's triple can be recorded as follows: Where h is the head entity, r is the relation, o is the tail entity or attribute value, s is the evidence source identifier, w is the retrieval score or verification score, ρ is the round of entering the cache, and σ is the status label.

[0012] Implement state management for triples in the cache. State labels should include at least retrieved, alias, bridge, constraint, supportive, verified, and conflict. State management allows for clear differentiation of evidence sources, repair purposes, supporting roles, and conflict situations based on state, facilitating subsequent incremental reuse, prioritization, evidence compression, and manual auditing.

[0013] Connect the user question q with the initial evidence subgraph The structured representation is input into the reasoning model to obtain the initial answer. Initial reasoning traces and initial confidence information The reasoning model can be a large language model, or a combination of a large language model and a graph reasoning module. The structured representation can employ one or more of the following: a triplet list, path summary, constraint summary, evidence source summary, and conflict marker summary.

[0014] After the initial reasoning, the uncertainty score U of the current answer is calculated. U can be composed of the answer confidence inverse term u1, multi-sample inconsistency term u2, missing support path term u3, missing relation coverage term u4, missing constraint coverage term u5, and candidate competition term u6, which can be specifically represented as: , where λ1 to λ6 are weight parameters, satisfying that each weight is non-negative and their sum is 1.

[0015] When U is not greater than the preset threshold τ, the system can directly proceed to support verification and output the current answer; when U is greater than τ, the failure diagnosis step is triggered. The failure diagnosticator D obtains the gap score vector based on the question q, the current answer a, the reasoning trace r, the current evidence subgraph G, and the graph evidence cache M. The gap type with the highest score is designated as the failure type g. Failure types include at least entity gaps, relation gaps, constraint gaps, and support gaps.

[0016] Based on the failure type g, the corresponding repair operator π(g) is selected to generate incremental graph evidence ΔG from the archive knowledge graph. Entity gaps correspond to entity alias expansion, same-name disambiguation, document number mapping, and bridging entity exploration; relation gaps correspond to relation focusing retrieval, relation direction expansion, and relation bridge search; constraint gaps correspond to time attribute completion, document number attribute completion, institutional hierarchy attribute completion, type edge completion, and version edge completion; support gaps correspond to multi-hop path recovery, broken path splicing, and intermediary node completion.

[0017] The incremental evidence ΔG is merged into the current evidence subgraph to obtain the repaired subgraph G1, denoted as follows: ,in This indicates a rule-based incremental merge operation, rather than a simple union. Incremental merge includes deduplication, redundancy filtering, conflict marking, and state updates.

[0018] The triplet retention function can be expressed as: ,in This indicates the relevance of the triplet to the file retrieval problem. This indicates the degree of support the triplet provides for the candidate answer. This indicates the contribution of the triple to the constraint set. As a redundancy penalty item, , , , These are the weight parameters.

[0019] Based on the repaired subgraph Through further reasoning, we arrive at the final answer. Before output, support validation is performed, including at least: whether a supporting path exists from the subject entity to the answer entity; whether the supporting path covers key relationship clues; whether the supporting path meets constraints such as time, document number, and institutional level; and whether the triples participating in the supporting path have been marked as supportive or verified. Output is considered complete when the current uncertainty score drops below a threshold, a complete supporting path has been formed, all constraints have been met, or the maximum number of repair rounds has been reached. Repair should be stopped immediately.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: 1. By comprehensively considering model confidence, support paths, relationship coverage, constraint coverage, and candidate competition through an uncertainty scoring mechanism, the probability of erroneous outputs that seem reasonable but lack archival basis can be reduced.

[0021] 2. By first diagnosing the cause of failure and then performing repairs according to the failure type, we avoid the large number of irrelevant triples entering the context due to unified expansion of the test, thereby reducing the noise of supplementary evidence and improving the quality of re-inference.

[0022] 3. By using graph evidence caching and incremental merging mechanisms to reuse acquired and valid archive triples, the number of graph queries, network transmission volume, hint construction length, and model inference cost are reduced.

[0023] 4. Record the source of evidence, supporting roles, conflict situations and verification status through triplet status tags, which facilitates manual review, accountability and auditing during the use of archives.

[0024] 5. Treating constraint gaps as an independent failure type for diagnosis and repair can better handle complex archival questions involving time, document number, and institutional level, such as which document number was the earliest and which department created the document.

[0025] 6. This invention prioritizes passing the filtered key triples, supporting paths, and constraint summaries to the generative model, rather than the full text of the underlying archives or large-scale graph data. This helps to reduce the exposure scope of the original archive data and improve data security. Detailed Implementation

[0026] The present invention will be further described below with reference to preferred embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0027] Example The reliable question-answering method for archival knowledge graphs mainly includes the following steps: Step S1: Receive the file query question input by the user, perform semantic parsing on the file query question, and obtain the set of subject entities, the set of relation clues, and the set of constraints; The system receives a natural language archive query question q from the user. The parsing module performs word segmentation, named entity recognition, entity linking, relation clue extraction, and constraint extraction on question q, resulting in... , and The parsing module can parse abbreviations, aliases, and old names in the field of archives, and call a dedicated alias dictionary and file number mapping table for entity normalization; for expressions such as earliest, latest, superior, basis, before and after change, the problem parsing module converts them into sorting constraints, time constraints, organizational hierarchy constraints, document number constraints, or version constraints.

[0028] Step S2: Based on the set of topic entities, the set of relation clues, and the set of constraints, perform an initial retrieval in the archive knowledge graph to construct an initial evidence subgraph; The initial retrieval module is based on For seed entities, query the one-hop neighborhood, specified relation neighborhood, and candidate bridging paths in the archival knowledge graph; simultaneously, based on... Filter edges that are irrelevant to the problem's relational clues, based on Supplement necessary attributes. To control the initial subgraph size, the system can set the maximum number of entities, the maximum number of relations, the maximum path length, and the retention limit for each type of relation, and prioritize retaining triples with clear evidence sources, high retrieval scores, and relevance to constraints.

[0029] Step S3: Write the initial evidence subgraph into the graph evidence cache, and set a status label for the triples in the graph evidence cache; The system will Write to the graph evidence cache M. TripleStore stores triples and their sources, scores, rounds, and states; PathStore stores the paths between topic entities and candidate answers; AliasStore stores file aliases, old names, abbreviations, and file number mappings; ConstraintStore stores constraint attributes such as time, document number, institutional level, and version; VerificationStore stores evidence fragments that have passed verification.

[0030] Then, the triples obtained from the initial retrieval are set to the retrieved state; the triples obtained by alias expansion or document number mapping are set to the alias state; the triples obtained by bridging across case files or matters are set to the bridge state; the triples that supplement constraints such as time, document number, and institutional level are set to the constraint state; the triples that actually participate in the answer support path are set to the supportive state; the triples that pass source verification, path verification, or manual review are set to the verified state; and the triples that conflict with existing evidence are set to the conflict state.

[0031] In some implementations, the graph evidence cache M not only stores retrieved evidence, but also the round, source, purpose, and status of each piece of evidence. During each round of repair, the system first checks the cache for reusable triples or paths to avoid repeatedly accessing the underlying archive knowledge graph.

[0032] Incremental merging is not simply about... Instead of adding it to G, the criteria are filtered based on relevance, support, constraint contribution, and redundancy. If a triple is highly relevant to the question and can supplement the supporting path or necessary constraints, it can be retained even if the retrieval score is not the highest; if a triple overlaps with existing evidence and does not increase the supporting path or constraint coverage, it can be filtered or only the source information can be merged.

[0033] When incremental evidence conflicts with existing evidence, the system does not directly delete the conflicting evidence. Instead, it marks it as conflict, records the conflict type and source identifier, and prompts the model to handle it with caution during re-inference. For example, if there are two formation times for the same event, the system can retain both time attributes and indicate during re-inference and output that manual verification is required or that verified evidence should be given priority.

[0034] Step S4: Input the structured representation of the file query question and the initial evidence subgraph into the reasoning model to obtain the initial answer, reasoning traces, and confidence information; Next, the initial inference is performed. The inference module will process question q. Input the triplet summary, path summary, constraint summary, and evidence source summary into the inference model. The inference model outputs the initial answer. Reasoning traces And confidence information Reasoning traces can be identified by the entity relationships, paths, and constraints upon which they are based.

[0035] Step S5: Calculate the uncertainty score based on the initial answer, reasoning traces, confidence information, and initial evidence subgraph, and determine whether to trigger repair based on the uncertainty score; In some implementations, the system uses an uncertainty score U as the basis for deciding whether to proceed with the repair process. The higher the U, the less reliable the current answer is. u1 reflects the confidence of the model's own output, u2 reflects the consistency of the results of multiple sampling inferences, u3 reflects whether the answer has a supporting path on the graph, u4 reflects whether the relationship required by the question is covered by the current subgraph, u5 reflects whether constraints such as time, document number, and institutional level are covered, and u6 reflects the degree of competition among multiple candidate answers.

[0036] For example, when a question asks which department first initiated a matter and based on which document number, if the current subgraph only contains the matter entity and department entity, but lacks the initiation time and the document number, then even if the inference model provides a department name, u5 will still be high. If there is no traceable path between the subject entity and the answer entity, then u3 will be high. If the scores of the top two candidate departments are close, then u6 will be high. All these indicators collectively cause U to exceed the threshold, thus triggering targeted correction, rather than directly outputting an answer without sufficient evidence.

[0037] The threshold τ can be set by the system administrator based on the application scenario of the archives, or it can be trained based on historical question and answer logs, manually labeled credible answers, and audit feedback. Weights λ1 to λ6 can be fixed or dynamically adjusted according to the question type. For example, for questions about which document number to rely on, the weight of the constraint coverage missing item u5 can be increased; for questions about which files a certain matter is related to, the weights of the support path missing item u3 and the relationship coverage missing item u4 can be increased. When specifically calculating the uncertainty fraction U, u1 can be taken as 1- u2 can be taken Where N is the number of sampling inferences, The number of times the same answer is given; u3 can be determined based on... The value is 1 when no supporting path exists, and 0 or a smaller value when a supporting path exists; u4 can take... u5 can be taken u6 can be taken ,in and The top two candidate answers will be scored. The above items can be normalized according to the business scenario.

[0038] Step S6: When the uncertainty score exceeds a preset threshold, perform a structured diagnosis of the failure cause to obtain the failure type; when Furthermore, the system outputs the answer when a supporting path exists for the current solution and the necessary constraints are met; otherwise, , or although However, when there are situations such as missing key constraints, broken supporting paths, or conflicting evidence, the system will enter the structured failure diagnosis stage.

[0039] The initial answer Initial reasoning traces Initial evidence subgraph Introducing a failure diagnostic tool. Analyze whether the required entities in the problem have been linked, whether key relationships have been covered, whether constraints such as time / document number / organizational level have been completed, and whether a complete supporting path exists in the answer. Output .like The largest, determined to be a solid gap; if The largest value is determined to be a relationship gap; if If the maximum value is reached, it is determined to be a constraint gap; if The largest value is identified as a support gap.

[0040] In some implementations, the failure diagnostic tool can be implemented using rules, classification models, or a combination of rules and models. Rules may include: increasing the likelihood of failure if a subject entity is not linked to a unique archive entity, or if multiple entities with the same candidate entity name exist. ;like If the key relationship in G is not present in G, or the direction of the relationship is inconsistent with the problem requirements, then the requirement is increased. ;like If time, document number, organizational level, version, or sorting constraints are missing, then improve... If there is no continuous path from the topic entity to the candidate answer, or if there are broken nodes in the path, then the score will be improved. .

[0041] Step S7: Select the corresponding repair operator according to the failure type and generate incremental graph evidence from the archive knowledge graph; Targeted incremental repair is performed, with the system selecting the repair operator π(g) based on the failure type. When g represents an entity gap, aliases, old names, abbreviations, file numbers, and disambiguation information for entities with the same name are retrieved first; when g represents a relation gap, disambiguation information is retrieved around... The missing relation retrieval function retrieves the target relation edge, reverse edge, and bridging relation; when g is a constraint gap, attributes such as time, document number, institutional level, archive category, and version status are added; when g is a support gap, multi-hop paths are restored or broken paths are pieced together between the subject entity, candidate answer, and intermediary node.

[0042] Specifically, the entity gap repair operator first queries the AliasStore and file number mapping table to obtain candidate mappings such as entity alias, old name, abbreviation, full name, file number, and dossier number, and then uses context constraints for disambiguation. For example, it excludes items with the same name based on the creation time, archiving department, and archive category.

[0043] The relation gap repair operator can be used for The system focuses on searching for missing relationships, prioritizing edges that are synonymous with or closely related to the predicate in the question, and allows bridging relationship searches within a limited number of hops. For example, for the question of which document number to rely on, in addition to directly searching for the basis relationship, it can also search for equivalent or similar relationships such as cited document number, approval document number, and processing basis.

[0044] The constraint gap repair operator can fill in attributes such as time, document number, organizational level, archive category, and version status. For sorting constraints, it can fill in the time attribute of all candidate objects before comparison; for organizational level constraints, it can fill in the superior authority, subordinate relationship, and historical evolution of the department; for version constraints, it can fill in or change to obsolete, continued, or other versions.

[0045] The gap repair operator supports constrained multi-hop path recovery between subject entities and candidate answers. Path search can be configured with a maximum number of hops, a relation whitelist, a time window, and a lower limit on source credibility. When a path break is detected, intermediate nodes at the break point are prioritized for completion, such as case files, documents, responsible departments, document numbers, or processing steps.

[0046] Step S8: Perform a rule-based incremental merge of the incremental graph evidence and the current evidence subgraph, and update the state label to obtain the repaired subgraph; Repair and merge module will Perform rule-based merging with the current subgraph. For duplicate triples, merge their source identifier, score, and state; for redundant triples, filter according to η(t); for conflicting triples, retain them but set a conflict state, and prompt the inference model to handle them carefully during re-inference; for strongly supporting triples, set a supportive or verified state. Then, based on... Reason again and output candidate answers. .

[0047] Step S9: Based on the repaired subgraph, reason again and output a credible answer with evidence sources after verification.

[0048] The verification and result output module verifies whether the answer entity or answer attribute can be reached from the subject entity along the supporting path, whether the supporting path covers key relationship clues, whether constraints such as time, document number, and institutional level are met, and whether the triples participating in the path have source identifiers and are in a supporting or verified state. After passing the verification, it outputs the answer text, evidence path, source identifier, uncertainty score, and necessary conflict prompts.

[0049] The output of a credible answer includes not only the natural language answer but also supporting paths and sources of evidence. Supporting paths can be represented as: subject entity relationship—intermediary entity relationship—answer entity, or subject entity attribute relationship—answer attribute value. Sources of evidence can point to archival entry numbers, case file numbers, document numbers, document numbers, page numbers, the department that created the document, or source nodes in the archival knowledge graph.

[0050] Supported verification can include path existence verification, relationship coverage verification, constraint consistency verification, and status verification. Path existence verification confirms that the answer can be deduced from archival evidence; relationship coverage verification confirms that key relationship clues in the problem are covered; constraint consistency verification confirms that constraints such as time, document number, organizational level, and version are satisfied; and status verification confirms that key triples are at least supportive, preferably verified. If a support verification fails but the maximum number of repair rounds has been reached... The system can output a prompt indicating that an answer cannot be determined based on the current archival evidence, and list the relevant evidence found, the types of missing evidence, and suggested directions for manual verification. This avoids forcibly generating a definitive answer when evidence is insufficient.

[0051] Furthermore, this invention also provides a trusted question-answering system based on archival knowledge graphs. The question parsing module is responsible for extracting... , and The initial retrieval module is responsible for constructing G0 from the archival knowledge graph; the graph evidence caching module is responsible for storing triples, paths, aliases, constraints, and validation information; and the triple state management module is responsible for setting and updating... The initial reasoning module and the re-reasoning module are responsible for generating the answer based on the current evidence subgraph; the uncertainty assessment module is responsible for calculating U; the failure diagnosis module is responsible for outputting g; the directional incremental repair module is responsible for calling π(g) to generate ΔG; and the repair and merging module is responsible for execution. The verification and result output module is responsible for outputting credible answers and evidence sources.

[0052] Each module can be deployed on the same server or in a distributed architecture. For example, the archival knowledge graph and caching modules can be deployed on the intranet data service side, the inference model can be deployed in a controlled computing environment, and the result output module can be embedded in the archive's business system, government service platform, or electronic archive management system.

[0053] This invention can also be implemented as an electronic device. The electronic device includes a processor, a memory, a communication interface, and a bus. The memory stores a computer program, and when the processor executes the computer program, it implements the aforementioned reliable question-and-answer method for archival knowledge graphs.

[0054] The present invention can also be implemented as a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-described method steps. The storage medium can be a read-only memory, random access memory, magnetic disk, optical disk, solid-state drive, or other medium capable of storing program code.

[0055] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A reliable question-answering method for archival knowledge graphs, characterized in that, Includes the following steps: Step S1: Receive the file query question input by the user, perform semantic parsing on the file query question, and obtain the set of subject entities, the set of relation clues, and the set of constraints; Step S2: Based on the set of topic entities, the set of relation clues, and the set of constraints, perform an initial retrieval in the archive knowledge graph to construct an initial evidence subgraph; Step S3: Write the initial evidence subgraph into the graph evidence cache, and set a status label for the triples in the graph evidence cache; Step S4: Input the structured representation of the file query question and the initial evidence subgraph into the reasoning model to obtain the initial answer, reasoning traces, and confidence information; Step S5: Calculate the uncertainty score based on the initial answer, reasoning traces, confidence information, and initial evidence subgraph, and determine whether to trigger repair based on the uncertainty score; Step S6: When the uncertainty score exceeds a preset threshold, perform a structured diagnosis of the failure cause to obtain the failure type; Step S7: Select the corresponding repair operator according to the failure type, and generate incremental graph evidence from the archive knowledge graph; Step S8: Perform a rule-based incremental merge of the incremental graph evidence and the current evidence subgraph, and update the state label to obtain the repaired subgraph; Step S9: Based on the repaired subgraph, reason again and output a credible answer with evidence sources after verification; The structured diagnostics are achieved through a failure diagnostic tool. Obtain the gap score vector The gap type with the highest score is designated as the failure type; where q represents the file query question, a represents the current answer, r represents the reasoning trace, G represents the current evidence subgraph, and M represents the graph evidence cache. , , , These correspond to entity gaps, relational gaps, constraint gaps, and support gaps, respectively.

2. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: The subject entity set includes at least one of the following: file name, matter name, organization name, document number, personnel, location, and time expression; the relationship clue set includes at least one of the following fields: "formed at", "belongs to", "based on", "handling department", "superior authority", "changed to", and "related file"; the constraint set includes at least one of the following: time constraint, document number constraint, organization level constraint, file category constraint, version constraint, sorting constraint, and comparison constraint.

3. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: The initial retrieval includes at least one of entity neighborhood expansion retrieval, relation focusing retrieval, constraint attribute retrieval, and bridging path retrieval, such that the initial evidence subgraph is the minimum necessary subgraph required to satisfy the first reasoning.

4. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: The graph evidence cache includes at least a triple cache, a path cache, an alias cache, a constraint cache, and a validation cache; the triple is denoted as... Where h represents the head entity, r represents the relation, o represents the tail entity or attribute value, s represents the evidence source identifier, w represents the retrieval score or verification score, ρ represents the round of entering the cache, and σ represents the status label; the status label includes at least retrieved, alias, bridge, constraint, supportive, verified, and conflict.

5. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: The uncertainty fraction U is according to Calculate, where λ1 to λ6 are non-negative weight parameters and their sum is 1, u1 is the answer confidence reverse term, u2 is the multi-sample inconsistency term, u3 is the support path missing term, u4 is the relation coverage missing term, u5 is the constraint coverage missing term, and u6 is the candidate competition term; Specifically, the answer confidence inverse term is determined based on confidence information; the multi-sample inconsistency term is determined based on the consistency ratio of multiple inference output answers; the support path missing term is determined based on whether there is a support path from the topic entity to the answer entity in the current evidence subgraph; the relation coverage missing term is determined based on the intersection ratio of the relation clue set and the relation set in the current evidence subgraph; the constraint coverage missing term is determined based on the intersection ratio of the constraint set and the constraint attribute set in the current evidence subgraph; and the candidate competition term is determined based on the score difference between the first candidate answer and the second candidate answer.

6. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: When the failure type is an entity gap, the repair operator performs at least one of entity alias expansion, same-name disambiguation, document number mapping, and bridging entity exploration; when the failure type is a relation gap, the repair operator performs at least one of relation focusing retrieval, relation direction expansion, and relation bridge search; when the failure type is a constraint gap, the repair operator performs at least one of time attribute completion, document number attribute completion, organizational level attribute completion, type edge completion, and version edge completion; when the failure type is a support gap, the repair operator performs at least one of multi-hop path recovery, broken path splicing, and intermediary node completion.

7. The reliable question-answering method for archival knowledge graphs according to claim 1, characterized in that: The rule-based incremental merging includes deduplication, redundancy filtering, collision marking, and state updating; among these, a triplet retention function is utilized. Screening the triples, when The triplet is retained if it is not less than the retention threshold; otherwise, the triplet is filtered out. rel(t,q) represents the relevance of the triplet to the file query question. This indicates the degree of support the triplet provides for the candidate answer. This indicates the contribution of the triple to the constraint set. Indicates a redundancy penalty term. , , , These are the weight parameters.

8. A trusted question-and-answer system based on archival knowledge graphs, characterized in that, include: The problem parsing module is used to perform semantic parsing on file query problems, and obtain the set of subject entities, the set of relation clues, and the set of constraints; The initial retrieval module is used to construct an initial evidence subgraph in the archival knowledge graph; the graph evidence caching module is used to cache triples, paths, aliases, constraints, and verification information. The triplet state management module is used to set and update triplet state labels; The reasoning module generates answers and reasoning traces based on the current evidence subgraph; the uncertainty assessment module calculates the uncertainty score of the answer and triggers repairs. The failure diagnosis module is used to output entity gaps, relation gaps, constraint gaps, or support gaps; The targeted incremental repair module is used to generate incremental graph evidence based on the failure type. The repair and merging module is used to perform rule-based incremental merging of incremental graph evidence with the current evidence subgraph; It supports a verification and result output module, which is used to verify the consistency of supported paths and constraints and output a reliable answer; The system is configured to perform the method of any one of claims 1-7.