A "evidence-conclusion" traceable map construction method for information retrieval results

Abstract

The application discloses an "evidence-conclusion" traceable map construction method of intelligence retrieval results, comprising: generating a multi-dimensional identifier for the retrieval results and performing evidence atomization processing; performing fact normalization on the evidence atoms and establishing a record relationship; constructing a source pedigree subgraph and calculating evidence independence; decomposing the conclusion atomization into conclusion claim nodes; constructing semantic relationship edges with direction and semantic types; encapsulating rules, models, prompt words and parameters into reasoning activity nodes to form an auditable reasoning path; calculating the traceable confidence of the conclusion and generating a minimum sufficient evidence set; when the evidence changes, the influence is propagated along the dependent edge and the review is linked. The application realizes accurate backtracking of the conclusion from document-level reference to fragment-level evidence, explicitly depicts the semantic relationship between evidence and conclusion, suppresses false consensus of homologous materials, supports reproducible audit of the large model generation process and dynamic invalidation control of the conclusion, and improves the interpretability and verification efficiency of the research and judgment conclusion.

Description

Technical Field

[0001] This application relates to the fields of information retrieval, intelligence analysis, knowledge graphs, natural language processing, trusted artificial intelligence, and data auditing, and specifically to a method, system, computer-readable storage medium, and electronic device for constructing an "evidence-conclusion" traceable graph of intelligence retrieval results. Background Technology

[0002] In business scenarios such as public opinion analysis, competitive intelligence, risk identification, policy research, military equipment analysis, industry analysis, and event tracking, analysts typically need to retrieve massive amounts of information from multiple sources, including news websites, social media, government announcements, corporate disclosures, industry databases, internal documents, and search engines. Based on this, they then generate analytical conclusions through manual summarization or Large Language Modeling (LLM). However, existing systems mostly present retrieval and analysis results in the form of lists, summaries, or report text, with only hyperlinks, citation numbers, or a general "source list" typically remaining between documents and conclusions.

[0003] Based on analysis of practical application scenarios, the traditional approach has the following pressing technical problems that need to be addressed:

[0004] First, the conclusions cannot be precisely traced back. A complex analytical conclusion is usually supported by multiple pieces of information, multiple sources, and multiple steps of reasoning. Traditional citation techniques can generally only locate the entire document or webpage, and cannot delve into specific pieces of evidence, structured facts, and intermediate reasoning rules, resulting in high verification costs for the conclusions.

[0005] Secondly, the semantic relationship between evidence and conclusion is missing. Most existing methods only record the fact that "the document is cited" and fail to explicitly distinguish the complex semantic types between evidence and conclusion, such as whether the evidence directly supports, indirectly supports, refutes, limits, supplements background, corrects time, or corrects the stance. This results in a lack of interpretability in the formation process of the conclusion.

[0006] Third, multiple sources of evidence are easily weighted repeatedly, leading to false consensus. The phenomena of reposting, translating, excerpting, and secondary commenting are extremely common in the internet ecosystem, with related content often originating from the same original material. Existing systems frequently treat these as multiple independent pieces of evidence, thus creating false multi-source corroboration.

[0007] Fourth, the conclusions cannot be updated in response to changes in evidence. When the original page is updated, links become invalid, reports are corrected, or new counter-evidence emerges, the existing conclusions usually cannot be automatically detected and the affected scope is marked. This requires manual review of each item, resulting in poor timeliness.

[0008] Fifth, the generation process of large models is not auditable. In scenarios where large models assist in judgment, the models may implicitly use unlisted context or present uncertain information as certain conclusions. Existing systems lack effective link tracing mechanisms, making it difficult to determine whether a particular conclusion is supported by sufficient, independent, and valid evidence.

[0009] Sixth, there is a lack of a minimum set of evidence for auditing. Reviewers often want to verify a conclusion in the shortest possible time, but traditional systems can only accumulate massive amounts of data and cannot automatically extract the "minimum sufficient evidence set" and its corresponding reasoning path to support the conclusion.

[0010] Therefore, there is an urgent need to develop a graph construction scheme that deeply correlates original retrieval materials, evidence atoms, normalized facts, reasoning activities, conclusions and their versions, and supports two-way traceability, dynamic failure propagation and interpretable auditing. Summary of the Invention

[0011] To overcome the aforementioned deficiencies in existing technologies, this application proposes a novel method and system for constructing a traceable "evidence-conclusion" graph for multi-source intelligence retrieval results. More specifically, this invention aims to overcome the technical deficiency in multi-source intelligence retrieval where "the source of evidence is visible, but the process of conclusion formation is untraceable," and provides a method and system for constructing a traceable "evidence-conclusion" graph based on evidence atoms, semantic relationships as edges, reasoning activities as intermediary nodes, and version dependence and source lineage as constraints. This graph can systematically answer:

[0012] (1) What specific pieces of evidence support a certain conclusion?

[0013] (2) What is the lineage and version of each piece of evidence?

[0014] (3) What rules or models led to the conclusions from the evidence;

[0015] (4) Whether there is any counter-evidence or invalid evidence;

[0016] (5) Which downstream conclusions will be affected by the change in evidence?

[0017] To achieve the above objectives, the present invention employs the following technical strategies:

[0018] 1. Proposes a four-layer heterogeneous graph model of evidence atoms—normalized facts—reasoning activities—conclusions, breaking through the limitation of traditional document-level citations that cannot accurately explain the source of conclusions.

[0019] 2. Propose evidence-conclusion relationship edges with direction and semantic type, and subdivide them into direct support, indirect support, rebuttal, limitation, time correction and caliber correction, so that the conclusion formation process is calculable and auditable.

[0020] 3. Propose a joint weighting mechanism of source lineage and independent evidence clusters to reduce false multi-source corroboration caused by reprinting, translation and abstract dissemination.

[0021] 4. A scheme is proposed to encapsulate the model, prompt words, rules, parameters, and input / output as inference activity nodes, so that the large model-assisted judgment results have reproducible link records.

[0022] 5. A method for constructing a minimum sufficient evidence set that simultaneously considers support strength, independence, coverage of limiting conditions, and reading cost is proposed.

[0023] 6. Propose a dependency propagation mechanism from evidence version changes to conclusion invalidation indicators, so that the graph can not only record historical relationships, but also proactively discover conclusions that need to be reviewed.

[0024] Specifically, this application provides the following technical solutions:

[0025] The first aspect of this application provides a method for constructing a traceable "evidence-conclusion" map of intelligence retrieval results, such as... Figure 1 As shown, it includes the following steps:

[0026] S1. Receive the search topic, original search results, and a set of conclusions to be generated or verified; generate an immutable document identifier, acquisition time identifier, source identifier, and content version identifier (data access and multi-dimensional identifier) ​​for each search result; perform content segmentation, page location marking, text recognition, structured field extraction, and time information standardization on the search results to obtain evidence atoms that can be independently cited (evidence atomization processing); perform fact extraction and fact normalization on the evidence atoms to generate normalized fact nodes, and establish a "record" relationship between the evidence atoms and the normalized fact nodes (fact normalization and record association).

[0027] S2. Identify the source lineage relationships between each search result, construct a source lineage subgraph, and calculate the evidence independence accordingly. Group materials from the same source into the same evidence independence cluster for weighting to prevent duplicate weighting of materials from the same source (source lineage construction and independence calculation); perform atomized decomposition on the report conclusion or model output to generate conclusion claim nodes (conclusion atomization decomposition); based on semantic matching, numerical constraints, temporal consistency, entity consistency, and inference rules, construct semantic relationship edges with direction and semantic type between the normalized fact nodes, evidence atoms, and conclusion claim nodes to form a traceable "evidence-conclusion" graph (semantic relationship edge construction).

[0028] S3. Encapsulate the rules used to generate the conclusion, the name of the large model, the version of the prompt words, the set of input evidence, the parameter configuration, and the execution time into reasoning activity nodes, and connect the input evidence or fact nodes to the conclusion claim nodes through the reasoning activity nodes to form an auditable reasoning path (reasoning activity node encapsulation); calculate the traceability confidence of the conclusion based on the supporting edges, rebuttal edges, evidence independence, source authority, time validity, and reasoning rule reliability of the conclusion claim nodes, and automatically generate the minimum sufficient evidence set and structured conclusion explanation cards to support the conclusion claim nodes (confidence calculation and minimum evidence set generation).

[0029] S4. When the evidence atom, original search result, normalized fact node or reasoning activity node is updated, becomes invalid, withdrawn, conflicts or adds counter-evidence, the influence propagation is performed through the dependency edges in the traceable graph, the affected conclusion claim node is automatically identified, and a review task or downgrade mark is generated (dynamic failure propagation and review linkage).

[0030] Furthermore, in the method of this application, the evidence atom at least includes: evidence atom identifier, original document identifier, document version identifier, original text fragment, position of the fragment in the document, semantic type, entity set, applicable time interval, tone or deterministic type, content hash value, and evidence quality vector; the position of the fragment in the document includes at least one of page number, DOM path, or character offset interval.

[0031] Furthermore, in the method of this application, in step S1, the standardized fact node includes subject, predicate, object, time, region, value, unit, limiting conditions and statistical caliber; the fact standardization includes: entity identification of evidence atoms, and performing time normalization, value and unit conversion and statistical caliber identification, and setting unified and comparable status labels for the event life cycle stages.

[0032] Furthermore, in the method of this application, in step S2, the source lineage relationship includes at least one of reprint, citation, translation, abstract, version update and common original source relationship; the identification basis of the source lineage relationship includes at least one of hyperlink citation, citation overlap, publication time sequence, text similarity, image fingerprint and entity time pattern; the calculation of evidence independence includes: clustering homologous evidence into an evidence independence cluster according to the source lineage, and combining evidence from the same evidence independence cluster for weighting.

[0033] Furthermore, in the method of this application, in step S2, the semantic type of the semantic relationship edge includes at least: direct support, indirect support, rebuttal, limitation, caliber correction, time correction, background explanation, deduction, and uncertainty indication; the construction of the semantic relationship edge is implemented using symbolic logic rules, large language models, natural language implication models, or combinations thereof; wherein, for numerical claims, numerical interval inclusion and comparison rules are used, for time claims, the inclusion and intersection relationships of time intervals are verified, and for event stage claims, a preset business state transition diagram is used for logical verification.

[0034] Furthermore, in the method of this application, in step S3, the reasoning activity node records at least: the name and version of the basic model called, the version of the prompt word template, the version of the reasoning rule, the set of input evidence atom identifiers, the set of input fact node identifiers, the set of output conclusion claim identifiers, the execution timestamp, the parameter configuration, and the operator identifier; when the complete prompt word cannot be saved in plaintext, the encrypted hash value, template version, and desensitized parameter digest of the prompt word are saved to achieve deterministic reproduction.

[0035] Furthermore, in the method of this application, in step S3, the calculation of the traceability confidence of the conclusion includes: calculating it according to the following model based on the supporting evidence set, the rebuttal evidence set, and the set of independent evidence clusters:

[0036]

[0037] Wherein, P(C) is the traceability confidence of the conclusion, G is the set of independent evidence clusters, R is the set of rebuttal evidence, Wg is the comprehensive weight of the independent evidence clusters, Sup(g,C) is the support strength of the independent evidence clusters for the conclusion, Wr is the weight of the rebuttal evidence, Ref(r,C) is the rebuttal strength, Amb(C) is the ambiguity of the conclusion statement, and θ is the penalty coefficient; evidence from the same source lineage is weighted together in the same evidence cluster.

[0038] Furthermore, in the method of this application, step S3, the generation of the minimum sufficient evidence set includes: sorting the evidence nodes supporting the conclusion claim nodes according to source independence, directness, authority, timeliness, and coverage; under the conditions of satisfying the preset support threshold, key limiting condition coverage threshold, and independent source cluster number threshold, selecting the evidence subset with the fewest evidence nodes and the lowest source lineage overlap, and attaching at least one rebuttal or limiting evidence; the minimum sufficient evidence set does not contain any high-intensity rebuttal evidence that has not been included in the interpretation.

[0039] Furthermore, in the method of this application, step S4 specifically includes: when the original search result is updated, creating a new version of the evidence atom and connecting the old version of the evidence atom through a "replacement" edge; starting from the replaced old version of the evidence atom, traversing along the "record", "support", and "derive" dependency edges in the traceability graph to identify the affected conclusion claim nodes; marking the affected conclusion claim nodes as pending review, recalculating their conclusion traceability confidence, and generating a structured review task notification containing the affected conclusion, the difference between the original text fragments before and after the change, and the suggested modification actions.

[0040] The second aspect of this application provides a system for constructing a traceable "evidence-conclusion" map of intelligence retrieval results. The system, when running, implements the steps of the aforementioned method for constructing a traceable "evidence-conclusion" map of intelligence retrieval results. The system includes:

[0041] The data access module is used to receive search results, web page snapshots, document files, database records and manually entered information, and to generate an immutable document identifier, acquisition time identifier, source identifier and content version identifier for each search result;

[0042] The evidence atom generation module is used to segment, locate, extract semantics, hash content, and label quality of search results to generate evidence atoms.

[0043] The fact normalization module is used to generate normalized fact nodes, and to normalize entities, time, values ​​and statistical standards, as well as to establish a "record" relationship between evidence atoms and normalized fact nodes;

[0044] The source lineage identification module is used to identify the relationships of reprinting, citation, translation, derivation and version update among search results, construct source lineage subgraphs and determine independent clusters of evidence;

[0045] The conclusion decomposition module is used to decompose the report conclusions or model outputs into verifiable conclusion claim nodes;

[0046] The relation determination module is used to identify the supporting, refuting, limiting, modifying and deducing relationships between evidence and conclusions, and to construct semantic relation edges with direction and semantic type.

[0047] The inference activity recording module is used to encapsulate rules, models, prompt word versions, inputs, outputs, and parameters into inference activity nodes, forming an auditable inference path;

[0048] The confidence calculation module is used to calculate the traceability confidence of the conclusion and generate the minimum sufficient evidence set.

[0049] The version impact analysis module is used to detect changes to evidence atoms, original search results, normalized fact nodes, or reasoning activity nodes, and propagate the changes to the affected conclusion claim nodes along the dependency edges.

[0050] The graph query and audit module provides reverse graph tracing of conclusions, positive impact analysis of evidence, time slice playback, and audit report output. It supports reverse graph traversal starting from any conclusion claim node to view the original source, the original text of the underlying evidence, the transformation rules, and the context of the reasoning activity nodes. When it is impossible to save the complete prompt words in plaintext, it supports saving the encrypted hash value, template version, and desensitized parameter summary of the prompt words to achieve deterministic reproduction.

[0051] A third aspect of this application provides an electronic device, including: a memory and a processor;

[0052] Memory: Used to store computer programs;

[0053] Processor: Used to execute the computer program to implement the steps of the aforementioned method for constructing an evidence-conclusion traceability map of intelligence retrieval results.

[0054] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the aforementioned method for constructing an "evidence-conclusion" traceability map of intelligence retrieval results.

[0055] In summary, compared with the prior art, the solution of the present invention has the following advantages:

[0056] (1) From document-level citation to fragment-level evidence tracing: Each conclusion can be precisely located to the original text fragment, page number, paragraph, webpage snapshot location or structured field, which completely breaks the traditional limitation of "only seeing the whole document and not the specific paragraph", and significantly improves the efficiency of manual verification.

[0057] (2) Explicitly map the relationship of evidence: The system not only records "what was cited", but also explains "how the evidence supports, refutes, limits or modifies the conclusion" through edges with direction and semantic type, making the complex judgment process highly interpretable.

[0058] (3) Suppress false consensus through source lineage constraints: Clustering and weighting of a large number of reprinted, translated, and excerpted materials on the Internet effectively avoids the problem of multiple derivative reports artificially increasing the credibility of the conclusion.

[0059] (4) Dynamic failure control of supporting conclusions: When evidence becomes invalid, webpage is updated, content is withdrawn or new counter-evidence is added, the changes can be automatically propagated along the dependency edge, so that the affected downstream conclusions are marked as "pending review", "downgraded" or "invalid" in real time, ensuring the timeliness and accuracy of intelligence.

[0060] (5) Significantly improve the auditability of the output of the large model: Introduce "inference activity nodes" to fully record the name of the large model, the version of the prompt words, the set of input evidence and hyperparameters, so that the conclusion generated by each large model can be replayed, reproduced and subject to strict compliance audit.

[0061] (6) Automatically generate minimum sufficient evidence set: While ensuring the support strength and coverage of key limiting conditions, automatically reduce the amount of reading materials, significantly reduce the cognitive load of the reviewer, and improve the generation and auditing efficiency of the assessment report. Attached Figure Description

[0062] To more clearly illustrate the technical solution of this application, the accompanying drawings involved in the description of this invention will be briefly introduced below. It should be noted that the drawings only show some embodiments of the invention. For those skilled in the art, other related drawings can be derived from these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram illustrating the operational steps of the "evidence-conclusion" traceability map construction method for the intelligence retrieval results of this invention.

[0064] Figure 2 This is a schematic diagram of the overall architecture of the "evidence-conclusion" traceability map construction system for intelligence retrieval results in an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram illustrating the process of generating evidence atoms and normalized fact nodes from retrieved documents in an embodiment of the present invention.

[0066] Figure 4 This is a schematic diagram of the source lineage identification and evidence independent cluster construction process in an embodiment of the present invention.

[0067] Figure 5 This is a schematic diagram illustrating the traceability relationship between evidence, reasoning activities, and conclusions / claims in an embodiment of the present invention.

[0068] Figure 6 This is a schematic diagram illustrating the process of evidence alteration affecting the propagation of the evidence and the review of conclusions in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0070] In this document, the term "comprising" and any variations thereof (such as "including," "including," etc.) are open-ended expressions and should be understood as "including but not limited to," meaning that the listed content is not exhaustive and may include other content not explicitly mentioned. The term "based on" should be understood as "at least partially based on," meaning that the basis or condition referred to may not be the only factor and may involve other relevant factors. The term "one embodiment" should be understood as "at least one embodiment," meaning that the described embodiment is not the only possible implementation, and other similar embodiments may exist.

[0071] In this application, the terms "a" and "a plurality of" are used to modify related elements or features, and their expression is illustrative rather than restrictive. Unless otherwise expressly stated in the context, "a" should be understood as "at least one," and "a plurality of" should be understood as "at least two." Those skilled in the art should reasonably interpret these terms based on the semantic and logical relationships of the context to ensure that they cover the possibility of "one or more."

[0072] This invention provides the following technical solution:

[0073] The first aspect of this invention provides a method for constructing an evidence-conclusion traceability map of intelligence retrieval results, the core steps of which include:

[0074] 1 (Data Access and Multidimensional Identification): Receives the search topic, original search results, and a set of conclusions to be generated or verified; generates an immutable document identifier (d_id), retrieval time identifier, source identifier, and content version identifier (v_id) for each search result.

[0075] 2 (Evidence Atomization Processing): The search results are segmented into content blocks, marked with page positions, text recognized, structured fields extracted, and time information standardized to obtain "evidence atoms" that can be cited independently; each evidence atom contains at least the original text fragment, document position, semantic type, entity set, time range, credibility, and content hash value.

[0076] 3 (Fact Standardization and Record Association): Extract facts and standardize facts from evidence atoms to generate standardized fact nodes that include subject, predicate, object, time, region, value, unit, limiting conditions and statistical caliber; establish a "record" relationship between evidence atoms and standardized fact nodes.

[0077] 4 (Source lineage construction and independence calculation): Identify the relationships of reprinting, citation, translation, abstracting, version updates and common original sources among documents, and construct a source lineage subgraph; calculate the independence of evidence based on the source lineage to prevent homologous materials from being repeatedly used as independent evidence.

[0078] 5 (Conclusion Atomization Decomposition): Atomize the report conclusions or model outputs to generate conclusion claim nodes; each conclusion claim node includes the conclusion content, scope of application, time validity period, level of certainty, and conclusion version identifier.

[0079] 6 (Construction of Semantic Relationship Edges): Based on semantic matching, numerical constraints, temporal consistency, entity consistency, and reasoning rules, normalized fact nodes, evidence atom nodes, and conclusion claim nodes are connected into semantic relationship edges with direction and type; the semantic relationship edges include at least direct support, indirect support, rebuttal, limitation, caliber correction, time correction, background explanation, derivation, and uncertainty hints.

[0080] 7 (Encapsulation of Reasoning Activity Nodes): The rules for generating conclusions, the name of the main model, the version of the prompt words, the set of input evidence, the parameter configuration, and the execution time are encapsulated as "reasoning activity nodes"; the input evidence or fact nodes are connected to the conclusion claim nodes through the reasoning activity nodes, thereby forming a complete and auditable reasoning path.

[0081] 8 (Confidence Calculation and Minimum Evidence Set Generation): Based on the supporting and rebuttal sides of the conclusion's claim nodes, the independence of evidence, the authority of the source, the validity of time, and the reliability of the reasoning rules, calculate the traceability confidence of the conclusion; automatically generate the minimum sufficient evidence set supporting the conclusion and a structured conclusion explanation card.

[0082] 9 (Dynamic Failure Propagation and Review Linkage): When evidence atoms, original documents, fact nodes, or reasoning activities are updated, become invalid, withdrawn, conflicted, or new counter-evidence is added, the impact propagation is performed through the dependency edges in the graph, automatically identifying the affected conclusion claim nodes and generating review tasks or downgrade markers.

[0083] Core data structure definition:

[0084] Evidence of atomic structure (E):

[0085] E=(e_id,d_id,v_id,span,loc,entity_set,time_interval,modality,hash,q)

[0086] Where e_id is the evidence atom identifier, d_id is the original document identifier, v_id is the document version identifier, span is the original text fragment, loc is the position of the fragment in the document (such as page number, DOM path or coordinates), entity_set is the entity set, time_interval is the applicable time interval, modality is the tone or deterministic type, hash is the content hash value, and q is the evidence quality vector.

[0087] The conclusion advocates for a node structure (C):

[0088] C=(c_id,statement,scope,valid_interval,certainty,c_version)

[0089] Where c_id is the conclusion identifier, statement is the conclusion description, scope is the applicable objects and conditions of the conclusion, valid_interval is the valid time interval, certainty is the certainty level, and c_version is the conclusion version identifier.

[0090] Core computational model:

[0091] The conclusion can be traced back to the confidence calculation model P(C):

[0092] Calculate based on the supporting evidence set S, the rebuttal evidence set R, and the set of independent evidence clusters G:

[0093]

[0094] Wherein, Wg is the overall weight of the independent evidence cluster, Sup(g,C) is the support strength of the independent evidence cluster for the conclusion, Wr is the weight of the rebuttal evidence, Ref(r,C) is the rebuttal strength, Amb(C) is the ambiguity of the conclusion statement, and θ is the penalty coefficient. Evidence from the same source lineage is weighted together in the same evidence cluster, effectively suppressing "false consensus".

[0095] Minimal Sufficient Evidence Set Generation Strategy:

[0096] From all the evidence nodes supporting the conclusion, sort them according to the independence of the source, directness, authority, timeliness and coverage; under the premise of meeting the preset support threshold and the key limiting condition coverage threshold, select the evidence subset with the fewest nodes and the lowest overlap of source lineages; at the same time, attach at least one high-risk rebuttal or limiting evidence to avoid one-way interpretation bias.

[0097] The second aspect of this invention provides a system for constructing a traceable "evidence-conclusion" map of intelligence retrieval results, such as... Figure 2 As shown, it includes:

[0098] Data access module: used to receive search results, webpage snapshots, document files, database records, and manually entered information;

[0099] Evidence Atom Generation Module: Used to complete block segmentation, location positioning, semantic extraction, content hashing, and quality labeling;

[0100] Fact normalization module: used to generate normalized fact nodes and complete the normalization of entities, time, values ​​and calibers;

[0101] Source lineage identification module: used to construct relationships between reprints, citations, translations, derivations, and version updates;

[0102] Conclusion Decomposition Module: Used to decompose report conclusions or model outputs into verifiable conclusion claim nodes;

[0103] Relationship determination module: used to identify the supporting, refuting, limiting, modifying, and deductive relationships between evidence and conclusions;

[0104] Inference Activity Recording Module: Used to record rules, models, prompt word versions, inputs, outputs, and parameters;

[0105] Confidence Calculation Module: Used to calculate the traceability confidence of the conclusion and the minimum sufficient evidence set;

[0106] Version Impact Analysis Module: Used to detect changes and propagate the findings to affected parties;

[0107] The graph query and audit module is used to provide reverse tracing of conclusions, positive impact analysis of evidence, time slice playback, and audit report output.

[0108] To more clearly illustrate the technical solution of this application, the following will provide further explanation through specific scenario embodiments.

[0109] Example 1: Multi-source intelligence retrieval and evidence atom construction

[0110] like Figure 3 As shown, in a corporate strategy and investment risk assessment task, analysts performed a comprehensive search using the theme "the scale and implementation status of a company's project investment in a target region." The data access module integrated heterogeneous documents such as government announcements, corporate annual reports, press releases, industry media reports, and social media announcements. The system first assigned an immutable document identifier (d_id) to each document and fully saved the original URL, crawl time, page snapshot, content hash, and publisher metadata.

[0111] Next, the evidence atom generation module meticulously segments the document based on its layout and semantic features, dividing it into semantic paragraphs, table units, image captions, quoted statements, and time expressions. For example, the statement "plans to invest $1 billion in the target region, with an initial investment of $200 million" from a company's annual report is precisely segmented into two independent evidence atoms: the investment plan atom and the initial investment atom. Each atom simultaneously saves the page number, DOM path, or character offset range (loc), allowing reviewers to jump back to the precise location on the original webpage or document with a single click.

[0112] Example 2: Fact Normalization and Source Lineage Identification

[0113] like Figure 4 As shown, in order to eliminate the differences in expression methods between heterogeneous texts, the fact normalization module transforms the above-mentioned evidence atoms into normalized semantic triples and multidimensional attributes:

[0114] Fact Node 1: Company A – Planned Investment – ​​$1 billion – Target Region – 2026-2028

[0115] Fact Node 2: Company A – Invested – $200 million – Target Region – As of Q2 2026

[0116] During this process, the system will uniformly convert different expressions such as "hundred million US dollars" and "billion US dollars" into standard currency and quantity formats, and set uniform and comparable status labels for different life cycle stages of events such as planning, signing, commencement, investment, and completion.

[0117] Meanwhile, the source genealogy identification module identifies the derivative relationships between multiple reports based on hyperlink citations, citation overlap, publication time sequence, text similarity, image fingerprints, and entity time patterns. If the system finds that three independent news reports all explicitly or implicitly cite the same corporate annual report, the system will create a "derived from" edge in the genealogy and group these three reports into the same independent evidence cluster. These clusters will then be weighted together in subsequent calculations, avoiding false consensus caused by "multiple publications of the same article."

[0118] Example 3: Conclusion Proposal Decomposition and Semantic Relation Determination

[0119] like Figure 5 As shown, in response to the comprehensive conclusion given in the assessment report: "Company A has completed a $1 billion investment in the target region", the conclusion decomposition module identifies that the statement contains three core verifiable elements: "completed", "$1 billion", and "target region", and generates the corresponding conclusion claim node (C) accordingly.

[0120] Subsequently, the relationship determination module conducted an in-depth comparison of the evidence and conclusion: the system found that the annual report evidence only supported "planned investment of $1 billion", and the initial investment evidence only supported "already invested $200 million". Neither of them could logically support "the $1 billion investment has been completed". Therefore, the system automatically established "restriction" and "refutation" edges, automatically downgraded the certainty level of the conclusion to "low", and intelligently recommended a more accurate revised statement to the user: "Company A plans to invest $1 billion from 2026 to 2028, and has invested $200 million as of the second quarter of 2026".

[0121] The aforementioned relationship determination can be achieved using symbolic logic rules, Large Language Model (LLM), Natural Language Entailment (NLI) models, or combinations thereof. For numerical claims, the system prioritizes numerical interval inclusion and comparison rules; for time-related claims, it prioritizes verifying the inclusion and intersection relationships of time intervals; for event phase-related claims, it uses a pre-defined business state transition diagram for automated logical verification.

[0122] Example 4: Inference Activity Nodes and Reproducible Auditing

[0123] When the system uses a large model or automated rule chain to generate a conclusion from multiple pieces of evidence, the reasoning activity recording module automatically creates a reasoning activity node. This node records at least:

[0124] The name and version of the underlying model being called;

[0125] Prompt Template version;

[0126] System core inference rule version;

[0127] The input set of evidence atom IDs and the set of fact node IDs;

[0128] The output conclusion asserts the ID set;

[0129] Execution timestamp, temperature parameter, confidence threshold, and operator identifier.

[0130] The complete traceable path in the graph can be represented as:

[0131]

[0132] The reviewer can perform a reverse graph traversal starting from any conclusion claim node to view its original source, underlying evidence text, transformation rules, and the context of the large model execution with a single click. If the complete prompt cannot be saved in plaintext due to corporate compliance or data privacy reasons, the system supports saving the encrypted hash of the prompt, template version, and de-identified parameter summary, thereby achieving 100% deterministic reproduction in a controlled environment.

[0133] Example 5: Generation of Minimal Sufficient Evidence Set

[0134] Regarding the conclusion that "Company A will complete a $1 billion investment by 2028", the system retrieved a massive amount of materials in the knowledge graph (including an authoritative corporate annual report, a regulatory agency project approval, and multiple news articles reprinted from the same source).

[0135] The confidence calculation and evidence simplification module executes the following optimization process:

[0136] 1. First, multiple news reports from the same source are aggregated and weighted into a single evidence cluster according to their source lineage;

[0137] 2. Select company annual reports as direct "evidence of the plan";

[0138] 3. Select regulatory approvals as independent "binding evidence";

[0139] 4. Retain one piece of supporting material explaining that "project implementation is affected by local approval progress".

[0140] As a result, the system automatically generates a minimum set of sufficient evidence containing three core supporting materials, completely changing the inefficient model of traditional reports that require listing dozens of redundant reprinted articles.

[0141] In addition, the generation of the minimum sufficient evidence set must strictly satisfy the following multidimensional constraints:

[0142] (1) The conclusion must fully cover the subject, action, object, value, time and limiting conditions in the claim;

[0143] (2) It must contain no fewer than the preset number of independent source clusters;

[0144] (3) There is no strong rebuttal evidence in the graph that has not been included in the interpretation;

[0145] (4) Under the premise of satisfying the above conditions, the total number of evidence, the degree of overlap of source lineage, and the cost of manual reading should be minimized globally.

[0146] Example 6: Version Changes and the Propagation of Dynamic Impacts

[0147] like Figure 6As shown, the system supports 24 / 7 dynamic monitoring and a failure propagation mechanism. When the backend monitoring detects that a company's annual report has been updated from version v1 to version v2, and the key data has been changed from "planned investment of $1 billion" to "planned investment of no more than $800 million":

[0148] (1) The evidence atom generation module automatically creates new versions of evidence atoms and connects the old version atoms through the "Supersede" edge.

[0149] (2) The version impact analysis module starts from the replaced old evidence atom and performs a breadth-first traversal along the dependency edges such as "record", "support", and "derive" in the graph.

[0150] (3) The test results revealed that several downstream companies have concluded that they are deeply dependent on the original investment range of "800 million to 1 billion US dollars".

[0151] (4) The system automatically marks these affected conclusions as “evidence has been changed and is pending review”, re-triggers the confidence calculation model to update its traceable confidence P(C), and automatically generates a structured review task notification for the analyst, which includes the affected conclusions, the differences between the original text fragments before and after the change, and the suggested modification actions.

[0152] The flowcharts and block diagrams in the accompanying drawings illustrate possible implementations of systems, methods, and computer program products according to various embodiments of this application, including architecture, functionality, and operation. In these figures, each block may represent a module, program segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should be noted that each block in the block diagrams and / or flowcharts, and combinations thereof, can be implemented using either a dedicated hardware-based system or a combination of dedicated hardware and computer instructions to achieve the specified function or operation.

[0153] Embodiments of this application also disclose an electronic device, including: a processor, a communication interface, a memory for storing a processor-executable computer program, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor executes the executable computer program to implement the steps of the aforementioned method for constructing an "evidence-conclusion" traceability map of intelligence retrieval results.

[0154] It is understood that, in addition to memory and a processor, this electronic device may also include input devices (such as a keyboard), output devices (such as a display), and other communication modules. These input devices, output devices, and other communication modules all communicate with the processor through I / O interfaces (i.e., input / output interfaces).

[0155] The operations described in this application can be implemented by writing computer program code using one or more programming languages ​​or a combination thereof. The programming languages ​​include, but are not limited to, the following types:

[0156] Object-oriented programming languages, such as Java, Smalltalk, C++, etc.

[0157] Conventional procedural programming languages, such as "C" or similar programming languages.

[0158] The execution methods of program code include, but are not limited to:

[0159] It runs entirely on the user's computer;

[0160] Part of it executes on the user's computer, and part of it executes on a remote computer;

[0161] Execute as a standalone software package;

[0162] It is executed entirely on a remote computer or server.

[0163] In scenarios involving remote computers, the remote computer can connect to the user's computer via any type of network, including but not limited to local area networks (LANs) or wide area networks (WANs). Furthermore, the remote computer can also connect to external computers through an internet service provider, for example, by utilizing the internet for connection.

[0164] Furthermore, this application also discloses a computer-readable storage medium, which, when the instructions in the computer-readable storage medium are executed by the processor of an electronic device, enables the electronic device to perform the various steps of the "evidence-conclusion" traceability map construction method for intelligence retrieval results disclosed in this application.

[0165] In the context of this application, a computer-readable storage medium refers to a tangible medium capable of storing computer program code and related data. Specific examples include, but are not limited to, the following:

[0166] (1) Portable computer disk: such as floppy disks and other removable magnetic storage media.

[0167] (2) Hard disk: including mechanical hard disks and solid-state hard disks and other fixed storage devices.

[0168] (3) Random Access Memory (RAM): A volatile storage medium used for temporary storage of data and program code.

[0169] (4) Read-only memory (ROM): a non-volatile storage medium used to store fixed programs and data.

[0170] (5) Erasable programmable read-only memory (EPROM) or flash memory: non-volatile storage media that supports multiple erasures and reprogrammings.

[0171] (6) Fiber optic storage devices: storage media based on fiber optic technology.

[0172] (7) Portable compact disc read-only memory (CD-ROM): a read-only medium that stores data in the form of an optical disc.

[0173] (8) Optical storage devices: such as DVDs, Blu-ray discs and other storage media based on optical principles.

[0174] (9) Magnetic storage devices: such as magnetic tapes, disks and other storage media based on magnetic principles.

[0175] (10) Any suitable combination of the above: for example, combining multiple storage media to meet different storage needs.

[0176] These computer-readable storage media can be used to store the program code and related data described in this application to support program execution and persistent data storage.

[0177] Specifically, according to embodiments of this application, the processes described in the flowcharts can be implemented as computer software programs. For example, embodiments of this application relate to a computer program product comprising a computer program carried on a non-transitory computer-readable medium. This computer program contains program code for executing the "evidence-conclusion" traceability map construction method for the intelligence retrieval results disclosed in this application. When this computer program is executed by a processing system, it can achieve the functions defined in the embodiments of this application.

[0178] While the foregoing discussion contains several specific implementation details, these details should not be construed as limiting the scope of this application. The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features. Furthermore, this application should also cover other technical solutions formed by any combination of the above-described technical features or their equivalents without departing from the foregoing disclosed concept.

[0179] Those skilled in the art should also understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features, without departing from the spirit and scope of the technical solutions of the embodiments of this application. These modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the core spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for constructing an evidence-conclusion traceability map of intelligence retrieval results, characterized in that, Includes the following steps: S1. Receive the search topic, original search results, and a set of conclusions to be generated or verified; generate an immutable document identifier, acquisition time identifier, source identifier, and content version identifier for each search result; perform content segmentation, page location marking, text recognition, structured field extraction, and time information standardization on the search results to obtain evidence atoms that can be independently cited; perform fact extraction and fact normalization on the evidence atoms to generate normalized fact nodes, and establish a "record" relationship between the evidence atoms and the normalized fact nodes; S2. Identify the source lineage relationships between the search results, construct a source lineage subgraph, and calculate the evidence independence accordingly. Group materials from the same source into the same evidence independence cluster for merging and weighting. Perform atomized decomposition on the report conclusions or model outputs to generate conclusion claim nodes. Based on semantic matching, numerical constraints, temporal consistency, entity consistency, and reasoning rules, semantic relationship edges with direction and semantic type are constructed between the standardized fact nodes, evidence atoms, and conclusion claim nodes to form a traceable "evidence-conclusion" graph. S3. Encapsulate the rules used to generate the conclusion, the name of the large model, the version of the prompt words, the set of input evidence, the parameter configuration, and the execution time into reasoning activity nodes, and connect the input evidence or fact nodes to the conclusion claim nodes through the reasoning activity nodes to form an auditable reasoning path; calculate the traceability confidence of the conclusion based on the supporting edges, rebuttal edges, evidence independence, source authority, time validity, and reasoning rule reliability of the conclusion claim nodes, and automatically generate the minimum sufficient evidence set and structured conclusion explanation cards to support the conclusion claim nodes; S4. When the evidence atom, original search result, normalized fact node or reasoning activity node is updated, becomes invalid, withdrawn, conflicts or adds counter-evidence, the influence propagation is performed through the dependency edges in the traceable graph, the affected conclusion claim node is automatically identified, and a review task or downgrade identifier is generated.

2. The method according to claim 1, characterized in that, The evidence atom includes at least: evidence atom identifier, original document identifier, document version identifier, original text fragment, fragment position in document, semantic type, entity set, applicable time interval, tone or deterministic type, content hash value, and evidence quality vector; the fragment position in document includes at least one of page number, DOM path, or character offset interval.

3. The method according to claim 1, characterized in that, In step S1, the standardized fact node includes subject, predicate, object, time, region, value, unit, limiting conditions and statistical caliber; the fact standardization includes: entity recognition of evidence atoms, and execution of time normalization, value and unit conversion and statistical caliber recognition, and setting unified and comparable status labels for the event life cycle stages.

4. The method according to claim 1, characterized in that, In step S2, the source lineage relationship includes at least one of reprint, citation, translation, abstract, version update and common original source relationship; the identification criteria for the source lineage relationship include at least one of hyperlink citation, citation overlap, publication time sequence, text similarity, image fingerprint and entity time pattern. The calculation of evidence independence includes: clustering homologous evidence into an evidence independence cluster according to the source lineage, and combining evidence from the same evidence independence cluster for weighting.

5. The method according to claim 1, characterized in that, In step S2, the semantic types of the semantic relationship edges include at least: direct support, indirect support, rebuttal, limitation, caliber correction, time correction, background explanation, deduction, and uncertainty indication; the construction of the semantic relationship edges is implemented using symbolic logic rules, large language models, natural language implication models, or combinations thereof; among them, numerical claims are implemented using numerical interval inclusion and comparison rules, time claims are implemented by verifying the inclusion and intersection relationships of time intervals, and event phase claims are implemented by using a preset business state transition diagram for logical verification.

6. The method according to claim 1, characterized in that, In step S3, the reasoning activity node shall record at least: the name and version of the base model called, the version of the prompt word template, the version of the reasoning rule, the set of input evidence atom identifiers, the set of input fact node identifiers, the set of output conclusion claim identifiers, the execution timestamp, the parameter configuration, and the operator identifier; When the complete prompt word cannot be saved in plaintext, the encrypted hash value, template version, and de-identified parameter digest of the prompt word are saved to achieve deterministic reproduction.

7. The method according to claim 1, characterized in that, In step S3, the calculation of the traceability confidence of the conclusion includes: based on the supporting evidence set, the rebuttal evidence set, and the set of independent evidence clusters, calculated according to the following model: Wherein, P(C) is the traceability confidence of the conclusion, G is the set of independent evidence clusters, R is the set of rebuttal evidence, Wg is the comprehensive weight of the independent evidence clusters, Sup(g,C) is the support strength of the independent evidence clusters for the conclusion, Wr is the weight of the rebuttal evidence, Ref(r,C) is the rebuttal strength, Amb(C) is the ambiguity of the conclusion statement, and θ is the penalty coefficient; evidence from the same source lineage is weighted together in the same evidence cluster.

8. The method according to claim 1, characterized in that, In step S3, the generation of the minimum sufficient evidence set includes: sorting the evidence nodes supporting the conclusion claim nodes according to source independence, directness, authority, timeliness, and coverage; under the conditions of satisfying the preset support threshold, key limiting condition coverage threshold, and independent source cluster number threshold, selecting the evidence subset with the fewest evidence nodes and the lowest source lineage overlap, and attaching at least one rebuttal or limiting evidence; the minimum sufficient evidence set does not contain any high-intensity rebuttal evidence that has not been included in the interpretation.

9. The method according to claim 1, characterized in that, Step S4 specifically includes: when the original search results are updated, creating a new version of the evidence atom and connecting the old version of the evidence atom through a "replacement" edge; starting from the replaced old version of the evidence atom, traversing along the "record", "support", and "derive" dependency edges in the traceability graph to identify the affected conclusion claim nodes; marking the affected conclusion claim nodes as pending review, recalculating their conclusion traceability confidence, and generating a structured review task notification containing the affected conclusions, the differences between the original text fragments before and after the change, and the suggested modification actions.

10. A system for constructing a traceable "evidence-conclusion" graph of intelligence retrieval results, characterized in that, The system operation implements the steps of constructing the "evidence-conclusion" traceability map of the intelligence retrieval results as described in any one of claims 1 to 9, including: The data access module is used to receive search results, web page snapshots, document files, database records and manually entered information, and to generate an immutable document identifier, acquisition time identifier, source identifier and content version identifier for each search result; The evidence atom generation module is used to segment, locate, extract semantics, hash content, and label quality of search results to generate evidence atoms. The fact normalization module is used to generate normalized fact nodes, and to normalize entities, time, values ​​and statistical standards, as well as to establish a "record" relationship between evidence atoms and normalized fact nodes; The source lineage identification module is used to identify the relationships of reprinting, citation, translation, derivation and version update among search results, construct source lineage subgraphs and determine independent clusters of evidence; The conclusion decomposition module is used to decompose the report conclusions or model outputs into verifiable conclusion claim nodes; The relation determination module is used to identify the supporting, refuting, limiting, modifying and deducing relationships between evidence and conclusions, and to construct semantic relation edges with direction and semantic type. The inference activity recording module is used to encapsulate rules, models, prompt word versions, inputs, outputs, and parameters into inference activity nodes, forming an auditable inference path; The confidence calculation module is used to calculate the traceability confidence of the conclusion and generate the minimum sufficient evidence set. The version impact analysis module is used to detect changes to evidence atoms, original search results, normalized fact nodes, or reasoning activity nodes, and propagate the changes to the affected conclusion claim nodes along the dependency edges. The graph query and audit module provides reverse graph tracing of conclusions, positive impact analysis of evidence, time slice playback, and audit report output. It supports reverse graph traversal starting from any conclusion claim node to view the original source, the original text of the underlying evidence, the transformation rules, and the context of the reasoning activity nodes. When it is impossible to save the complete prompt words in plaintext, it supports saving the encrypted hash value, template version, and desensitized parameter summary of the prompt words to achieve deterministic reproduction.

Images