An ai knowledge auditing method and system based on a structure syntax field

Abstract

The application provides an AI knowledge auditing method and system based on a structure syntax field, relates to the technical field of artificial intelligence data processing, and comprises the following steps: receiving user input content and preprocessing to obtain a processed text; performing object identification on the processed text to generate an object field; extracting a key proposition to generate a proposition field, and generating an evidence field according to sources, data, references, version information or context; performing evidence layering according to an evidence level rule to generate an evidence layering result; performing boundary detection based on a preset boundary type set to generate a boundary field; identifying a breaking point by combining the object field, the evidence field and the boundary field to generate a breaking field; performing route maturity and risk partitioning based on the evidence layering result, the boundary field and the breaking field to generate a route field; generating a revision suggestion field according to the breaking field and the route field; generating a knowledge auditing report, saving an auditing log, and performing variable updating after user confirmation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, and in particular to an AI knowledge auditing method and system based on structured grammar fields, which performs audit object identification, key claim extraction, evidence layering, boundary detection, breakpoint identification, route partitioning, and structured report generation on AI-generated content, knowledge base fragments, retrieval-enhanced generated results, project materials, or text materials. Background Technology

[0002] With the development of large language models, retrieval-enhanced generation systems, intelligent writing systems, enterprise knowledge base question-answering systems, and personal AI assistants, users are increasingly relying on AI systems to generate answers, write reports, organize research materials, review project proposals, assist in publishing, and manage knowledge bases. While these systems improve the efficiency of text generation, information organization, and knowledge retrieval, they also bring problems such as unclear knowledge sources, confusing evidence levels, insufficient citation support, overly strong conclusions, and difficulties in auditing processes.

[0003] In existing technologies, general large language model question-answering systems typically generate natural language answers based on user input, contextual content, and model parameters, capable of tasks such as question answering, summarizing, rewriting, and solution generation. Retrieval-enhanced generation systems provide contextual support to large language models by retrieving external knowledge bases, document fragments, or online resources, reducing the risks of unsourced generation and content illusion. Text review or compliance review systems are typically used to identify sensitive words, illegal content, copyright risks, formatting issues, or single-point factual errors. Project management and knowledge management systems are primarily used to store documents, tasks, knowledge entries, and collaboration records. Prompt-based engineering tools constrain the output of large language models to specific formats, styles, or workflows using preset prompts. These technologies, to varying degrees, improve the efficiency of text generation, knowledge retrieval, content review, and data management.

[0004] However, existing large language model question-answering systems, retrieval-enhanced generation systems, text review systems, and knowledge management systems typically focus on content generation, data retrieval, compliance checks, or document storage, respectively. They lack a unified, structured knowledge audit framework, making it difficult to clearly distinguish between audit objects, background information, user intent, search materials, and final conclusions. Furthermore, existing systems lack a systematic mechanism for judging the relationship between key claims and evidence, making it difficult to determine whether cited content truly supports the conclusion. They also struggle to differentiate between different levels of evidence, such as facts, data, models, reasoning, hypotheses, and public statements, easily leading to issues like false citations, weak evidence supporting strong conclusions, and the formalization of assumptions. In addition, existing systems lack boundary detection, breakpoint identification, and correction mechanisms, making it difficult to identify structural knowledge risks such as insufficient object definition, inadequate conditions for validity, incomplete verification chains, and unclear responsibility chains. They also lack the ability to automatically generate corrective suggestions based on audit results, such as supplementing sources, supplementing data, supplementing verification conditions, reducing the strength of conclusions, or freezing expressions. Therefore, it is difficult to form a traceable, verifiable, and sustainably iterative AI knowledge audit process. Summary of the Invention

[0005] To address the technical challenges of existing large language model question-answering systems, retrieval enhancement generation systems, text review systems, and knowledge management systems, which typically focus on content generation, data retrieval, compliance checks, or document storage respectively, lacking a unified structured knowledge audit framework, and struggling to clearly distinguish audit objects, background information, user intent, retrieval materials, and final conclusions; furthermore, existing systems lack a systematic judgment mechanism for the relationship between key claims and evidence, making it difficult to determine whether cited content truly supports conclusions, and also difficult to distinguish different levels of evidence such as facts, data, models, reasoning, hypotheses, and public expressions, easily leading to problems such as false citations, weak evidence supporting strong conclusions, and the factualization of hypotheses; in addition, existing systems lack boundary detection, breakpoint identification, and correction and transformation mechanisms, making it difficult to identify structural knowledge risks such as insufficient object definition, insufficient conditions for establishment, incomplete verification chains, and unclear responsibility chains, and also difficult to automatically generate correction suggestions such as supplementary sources, supplementary data, supplementary verification conditions, reducing the strength of conclusions, or freezing expressions based on audit results, thus making it difficult to form a traceable, verifiable, and sustainably iterative AI knowledge audit process. Therefore, this invention provides an AI knowledge audit method and system based on structured grammar fields.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: First aspect: This invention provides an AI knowledge auditing method based on structured syntax fields, the method being executed by at least one processor, and comprising: Receive user input; The user input is preprocessed to obtain preprocessed text; The preprocessed text is subjected to audit object identification to generate object fields; Key claims are extracted from the preprocessed text to generate claim fields for subsequent evidence field generation, evidence layering, boundary detection, breakpoint identification, route partitioning, and structured knowledge audit report generation. Evidence fields are also generated based on the source, data, citation, or context corresponding to the key claims. Based on preset evidence hierarchy rules, the key claims are layered into evidence layers to obtain evidence layering results; Boundary detection is performed on the key claims and the audit objects based on a preset set of boundary types to generate boundary fields; Breakpoint identification is performed based on the object field, the evidence field, and the boundary field to generate a break field; Based on the evidence stratification results, the boundary fields, and the fracture fields, the relevant models, technical routes, project plans, or expression contents are partitioned by route maturity and risk, and route fields are generated. Based on the fracture field and the route field, a correction suggestion field is generated; A structured knowledge audit report is generated based on the object field, the evidence field, the boundary field, the break field, the route field, and the correction suggestion field.

[0007] The second aspect: This invention provides an AI knowledge auditing system based on structured syntax fields, comprising: The user interaction interface module is used to provide human-computer interaction interfaces for user input, audit task initiation, audit report display, user confirmation, and variable update confirmation. The input receiving module is used to receive user input. The preprocessing module is used to preprocess the user input to obtain preprocessed text; The object recognition module is used to identify audit objects in the preprocessed text and generate object fields; The key claim extraction module is used to extract key claims from the preprocessed text and generate claim fields; The structure syntax field generation module is used to generate a set of structure syntax fields based on the source, reference, data, version information or context corresponding to the object field, claim field, and key claim; wherein the set of structure syntax fields includes at least one or more of the following: object field, claim field, evidence field, boundary field, break field, route field, correction suggestion field, report field or log field; The evidence stratification module is used to stratify the key claims based on preset evidence stratification rules to obtain evidence stratification results. The boundary detection module is used to perform boundary detection on the key claims and the audit object based on a preset set of boundary types, and generate boundary fields; The breakpoint identification module is used to identify breakpoints based on the object field, the evidence field, and the boundary field, and generate breakpoint fields. The evidence field includes at least one or more of the following: source information, citation information, data information, version information, evidence level information, and support status information corresponding to the key claim. The route partitioning module is used to partition relevant models, technical routes, project plans, or expression content based on the evidence stratification results, the boundary fields, and the breakpoint fields to determine the route maturity and risk, and generate route fields. The correction suggestion generation module is used to generate correction suggestion fields based on the break field and the route field; The report generation module is used to generate a structured knowledge audit report based on the object field, the evidence field, the boundary field, the break field, the route field, and the correction suggestion field; The audit log module is used to record one or more of the following: audit object, key claims, set of structured syntax fields, evidence stratification results, boundary detection results, breakpoint identification results, route maturity and risk zoning results, remedial suggestions, audit report version, generation time, user confirmation status, and system processing status, forming a traceable audit log; The optional variable update module is used to generate variable update items based on the audit report, break field, route field and correction suggestion field after user confirmation, and write the variable update items into the user variable library, project variable library or historical audit record.

[0008] Third aspect: An embodiment of the present invention provides an AI knowledge auditing device based on structured syntax fields, comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the AI ​​knowledge auditing method based on structured syntax fields as described in the first aspect.

[0009] Fourth aspect: The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI ​​knowledge auditing method based on structured syntax fields as described in the first aspect.

[0010] Fifth aspect: The present invention provides a computer program product, characterized in that: The computer program product includes a computer program that, when executed by a processor, implements the AI ​​knowledge auditing method based on structured syntax fields as described in the first aspect.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, by performing audit object identification, key claim extraction, evidence field generation, evidence layering, boundary detection, breakpoint identification, route maturity and risk partitioning, and correction suggestion generation on the text to be audited, AI-generated content, human-generated text, uploaded materials, knowledge base content, or retrieval-enhanced results can be uniformly transformed into structured audit objects. This facilitates a clear distinction between audit objects, background information, user intent, retrieval materials, and final conclusions, improving the comparability and verifiability of knowledge content. Simultaneously, by establishing a field-based correspondence between key claims and sources, data, citations, or context, and performing layered judgments based on preset evidence level rules, the ability to identify evidence support relationships can be enhanced, reducing risks such as false citations, weak evidence supporting strong conclusions, and the assumption of fact. Furthermore, through boundary detection, breakpoint identification, and route risk partitioning, structural knowledge risks can be marked, and correction suggestions such as supplementing sources, supplementing data, supplementing verification conditions, reducing conclusion strength, or freezing expressions can be automatically generated, thereby forming an AI knowledge audit process that is structurally clear, evidence-corresponding, has well-defined boundaries, and is easy to track and verify. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an AI knowledge auditing method based on structured syntax fields, provided as an embodiment of the present invention.

[0014] Figure 2 This invention provides a structural syntax field relationship diagram.

[0015] Figure 3 This is a diagram illustrating the relationship between evidence layering and boundary detection, provided as an embodiment of the present invention.

[0016] Figure 4 This invention provides a relationship diagram for generating a breakpoint identification and correction suggestion.

[0017] Figure 5 This invention provides a flowchart for route maturity and risk zoning.

[0018] Figure 6 This is a flowchart of a report generation and audit log provided for an embodiment of the present invention.

[0019] Figure 7 This is a diagram illustrating the overall architecture of an AI knowledge auditing system based on structured syntax fields, as provided in this embodiment. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] Example 1: Reference manual attached Figure 1 The diagram illustrates a flowchart of an AI knowledge auditing method based on structured syntax fields provided by an embodiment of the present invention.

[0024] Reference manual attached Figure 2 This illustrates a structural syntax field relationship diagram provided by an embodiment of the present invention.

[0025] This invention provides an AI knowledge auditing method based on structured syntax fields. This method can be implemented by an AI knowledge auditing device based on structured syntax fields, which can be a terminal or a server. The processing flow of the AI ​​knowledge auditing method based on structured syntax fields may include the following steps: Receive user input.

[0026] Specifically, user input includes user natural language questions, AI-generated content, uploaded text documents, PDF documents, Word documents, tabular data, web page materials, knowledge base fragments, search results from the retrieval enhancement generation system, research plans, project proposals, technical roadmap descriptions, publications, white papers, report drafts, course manuscripts, user's historical project materials, or combinations thereof. The system can receive user input through a user interface, file upload interface, knowledge base interface, search interface, API interface, or database interface.

[0027] The user input is preprocessed to obtain preprocessed text.

[0028] Specifically, the user input undergoes one or more of the following processing steps: text extraction, format conversion, paragraph segmentation, table parsing, reference identification, citation mark identification, source link identification, timestamp identification, version information identification, deduplication, language detection, preliminary domain classification, and sensitive field anonymization, to obtain preprocessed text. This preprocessed text is used for subsequent audit object identification, key claim extraction, and structural grammar field generation.

[0029] It should be noted that text extraction is used to extract processable text from input files or input data. Paragraph segmentation is used to divide input text into headings, chapters, paragraphs, sentences, table cells, or citation fragments. Table parsing is used to identify table headers, row and column relationships, numerical fields, unit fields, and note fields. Reference identification is used to identify bibliographic entries, authors, titles, publication information, years, links, or bibliographic numbers. Citation mark identification is used to identify citation numbers, footnotes, links, source descriptions, or RAG citation fragments in the text. Source link identification is used to extract URLs, document IDs, knowledge base sources, database sources, or internal system sources. Timestamp identification is used to identify publication time, update time, generation time, upload time, data collection time, or applicable time range. Version information identification is used to identify document versions, model versions, rule versions, policy versions, product versions, or knowledge base versions. Deduplication is used to identify duplicate paragraphs, duplicate documents, duplicate citations, or duplicate data. Sensitive field masking is used to replace or mask personal information, sensitive corporate information, account information, contact information, or other fields that need to be hidden. Preliminary domain classification is used to determine the domain to which the input content belongs, such as AI knowledge auditing, publication review, medical discussion, technical roadmap, materials engineering, energy systems, platform governance, RAG output review, or project proposal review.

[0030] The preprocessed text is subjected to audit object identification, and object fields are generated.

[0031] In one possible implementation, the object field includes at least the object name, object type, domain, audit target, and object boundary.

[0032] The object types include at least one of the following: theoretical models, technical routes, engineering systems, medical texts or health information content, published texts, AI answers, knowledge base entries, platform mechanisms, project plans, and public communication texts.

[0033] It should be noted that: Object Name is used to identify the object being audited. Object Type is used to identify the category to which the object belongs. Domain is used to identify the domain or application scenario in which the object exists. Audit Objective is used to identify the audit task the user wants the system to perform, such as determining credibility, identifying risks, reviewing evidence levels, stress testing technical routes, rewriting public statements, checking citation support status, or generating audit reports. Audit Task Type is used to identify whether this task belongs to AI response auditing, publication auditing, research material auditing, technical route auditing, RAG output auxiliary review, project plan review, or long-term project variable update. Object Boundaries are used to limit the scope of the object being audited. Excluded Content is used to exclude background information, historical materials, generalized comments, attachment descriptions, or content not requested by the user that is irrelevant to the audit objective.

[0034] Specifically, the system identifies audit targets based on titles, user questions, summaries, chapter titles, keywords, model judgments, technical route descriptions, project names, document metadata, user commands, and contextual semantics. The system can determine audit targets through rule matching, semantic classification, entity recognition, large language model extraction, or manual verification.

[0035] For example, when the content to be audited is a report on the "credibility of a certain AI knowledge base," the system can identify the "credibility assessment scheme of the AI ​​knowledge base" as the audit object, rather than the entire AI industry. As another example, when the content to be audited is a promotional material stating that "new materials will completely replace existing high-temperature alloys," the system can identify the "routes for replacing high-temperature alloys with new materials" as the audit object, rather than all materials science.

[0036] In this embodiment of the invention, by using the above-described audit object identification method, this implementation can accurately determine the specific object targeted by the knowledge audit based on information such as the title, user questions, summary, chapter titles, keywords, project names, document metadata, user instructions, and contextual semantics in the text to be audited, and generate object fields containing object name, object type, domain, audit target, and object boundaries. This avoids misidentifying domain background, the industry as a whole, the full text of related materials, or non-audit content as audit objects, reducing the risk of audit target confusion and audit scope overgeneralization.

[0037] Key claims are extracted from preprocessed text to generate claim fields for subsequent evidence field generation, evidence stratification, boundary detection, breakpoint identification, route partitioning, and structured knowledge audit report generation. Evidence fields are also generated based on the source, data, citation, or context corresponding to the key claims.

[0038] In one possible implementation, key claim extraction specifically includes: Identify at least one of the following in user input: factual judgment, data judgment, model judgment, causal judgment, route judgment, risk judgment, conclusion judgment, and public communication expression.

[0039] Establish connections between each key claim and the text paragraph containing the key claim, its source, citation, timestamp, and context.

[0040] It should be noted that a source refers to an information object that can provide factual, data-driven, model-based, background, contextual, or cited evidence for key claims. Sources can include publicly available papers, standard documents, regulations and policies, corporate announcements, database records, web pages, news reports, reports, auditable documents, internal knowledge base snippets, RAG search snippets, tabular data, experimental data, historical versions of documents, contextual paragraphs, or cited snippets in AI responses.

[0041] Specifically, the system extracts key claims from preprocessed text. Key claims include at least one of the following: factual assertions, data assertions, model assertions, causal assertions, route assertions, risk assertions, conclusion assertions, and public communication expressions. The system associates each key claim with its corresponding text paragraph, source, citation, timestamp, version information, and context to form the basis for subsequent evidence stratification and boundary detection.

[0042] Furthermore, the system generates audit fields based on a preset set of structured syntax fields. The set of structured syntax fields includes at least one of the following: object field, claim field, evidence field, boundary field, break field, route field, correction suggestion field, report field, log field, and optional variable field.

[0043] The claim field includes the key claim number, key claim text, claim type, paragraph, claim strength, and associated audit object.

[0044] The evidence fields include key claim number, key claim text, source of evidence, source type, source date, source version, evidence level, support status, citation location, and evidence gap. Boundary fields include scale boundary, time boundary, information boundary, evidence boundary, stable boundary, solvable boundary, closed-loop boundary, main boundary, and boundary justification. Breakpoint fields include breakpoint number, breakpoint type, breakpoint description, related claims, related evidence, related boundaries, risk level, and whether manual review is required. Route fields include route object, route type, partitioning basis, maturity, evidence strength, closed-loop degree, and public communication risk. Correction fields include correction suggestion number, correction type, suggested action, required supplementary data, required supplementary model, required supplementary experiment or verification, required closed-loop, and public expression rewriting suggestions. Report fields include report number, report type, summary, object identification results, evidence table, boundary table, breakpoint list, route partitioning results, correction suggestion list, and audit log summary. Log fields include input time, input content hash value, document version, model version, rule version, field generation result, boundary detection result, report version, report generation time, user modification records, and manual review records. Optional variable fields include user project number, historical object definition, historical evidence status, historical breakpoint, historical route partition, currently added data, currently changed variables, and user confirmation status.

[0045] Optionally, key claim extraction can be achieved through keyword rules, syntactic analysis, semantic classification, textual implication identification, large language model extraction, or manual review. The system can determine the priority of key claims based on claim strength, risk level, source status, and audit objectives.

[0046] For example, when processing the text "An AI system will completely replace human reviewers," the system can extract "AI system will completely replace human reviewers" as a conclusive claim and mark it as a high-intensity conclusion. The system can then check whether the claim has sufficient evidence, whether there is over-extension, whether the conclusion strength needs to be reduced, and whether it falls under a cautious approach or a frozen expression.

[0047] In this embodiment of the invention, through the aforementioned key claim extraction and evidence field generation methods, this implementation can identify core content such as factual judgments, data judgments, model judgments, causal judgments, route judgments, risk judgments, conclusion judgments, and public communication expressions from the text to be audited, and establish associations between each key claim and its corresponding paragraph, source, citation, timestamp, version information, and context. Thus, the main viewpoints, technical conclusions, risk judgments, and public expressions in unstructured text can be transformed into numberable, locatable, traceable, and auditable claim fields and evidence fields, avoiding the omission of key judgments or the misinterpretation of background descriptions as core conclusions during the audit process.

[0048] Reference manual attached Figure 3 This diagram illustrates a relationship between evidence layering and boundary detection provided by an embodiment of the present invention.

[0049] Based on the pre-defined evidence hierarchy rules, the key claims are stratified into different levels of evidence to obtain the evidence stratification results.

[0050] In one possible implementation, the evidence stratification result includes evidence stratification markers generated for each key claim; the evidence stratification markers include at least two of the following: fact stratification markers, data stratification markers, model stratification markers, reasoning stratification markers, hypothesis stratification markers, metaphor stratification markers, public expression stratification markers, and freeze stratification markers.

[0051] The evidence field includes supporting status information, which includes at least one of the following: strong support, weak support, indirect support, relevant but not supported, conflicting, outdated, unsourced, and requiring manual review.

[0052] Specifically, the system stratifies key claims and their corresponding sources, data, citations, or contexts according to preset evidence hierarchy rules, obtaining evidence hierarchy results. The evidence hierarchy results include at least two of the following: fact layer, data layer, model layer, reasoning layer, hypothesis layer, metaphor layer, public expression layer, and frozen layer.

[0053] It should be noted that the following layers are used: Fact Layer: Marks content with clear sources or verifiable factual basis. Data Layer: Marks content with data, experimental, observational, statistical, or measurement basis. Model Layer: Marks content based on theoretical, technical, or computational models. Reasoning Layer: Marks content derived from facts, data, or models. Hypothesis Layer: Marks content that has not yet been fully verified but is valuable for discussion. Metaphor Layer: Marks explanatory, analogical, or expressive language. Public Expression Layer: Marks simplified expressions intended for public communication. Freeze Layer: Marks content with insufficient current evidence, high risk, or unsuitable as a definitive conclusion. The freeze layer or frozen expression described in this invention refers to the system marking content with insufficient current evidence, high risk, excessive conclusion strength, or requiring manual review as temporarily not a definitive conclusion; this status does not indicate deletion of the content or denial of its research value, but rather serves to limit its output strength, trigger manual review, generate correction suggestions, or record audit logs.

[0054] The system can determine the support status of key claims and their corresponding sources. Support status can include at least one of the following: Strong_Support, Weak_Support, Indirect_Support, Related_Not_Support, Conflict, Outdated, No_Source, and Need_Review. When the cited content directly and currently effectively supports the key claim, the system can mark it as Strong_Support. When the citation only supports part of the content or its scope is smaller than the key claim, the system can mark it as Weak_Support. When the citation only provides background information and cannot directly prove the conclusion, the system can mark it as Indirect_Support. When the citation is only thematically related to the claim but does not support the core conclusion, the system can mark it as Related_Not_Support. When the citation conflicts with the claim in terms of facts, time, object, or conclusion, the system can mark it as Conflict. When the citation comes from obsolete versions or outdated materials, the system can mark it as Outdated. When a key claim cannot be linked to a clear source, the system can mark it as No_Source. When the system cannot determine supporting relationships or involves a high-risk scenario, the system can mark it as Need_Review. The system can automatically generate at least one corrective suggestion based on the support status: reduce the strength of the conclusion, supplement sources, enter manual review, or freeze the expression.

[0055] It should be noted that different fields can set different rules for refining the level of evidence. For example, in the audit of technical routes, laboratory sample data, engineering prototype data, long-term service data, industrialization data, and maintenance closed-loop data can be further subdivided. In the review of RAG outputs, strong support, weak support, relevant but not supporting, conflicting, outdated, unsourced, and requiring manual review can be classified as support statuses. In the review of published manuscripts, factual citations, model reasoning, public statements, and high-risk propaganda statements can be distinguished.

[0056] In this embodiment of the invention, through the aforementioned evidence stratification method, this implementation can classify key claims into different evidence levels based on preset evidence stratification rules, according to their source support, data foundation, model basis, reasoning relationship, hypothesis attributes, public communication attributes, and risk level. This distinguishes the hierarchical differences between facts, data, models, reasoning, hypotheses, metaphors, public expressions, and frozen content, avoiding the misinterpretation of hypothetical content, reasoning content, or public communication expressions as factual conclusions, and also preventing the support of strong conclusions with weak, indirect, outdated, or unsupported sources.

[0057] Boundary checks are performed on key claims and audit objects based on a preset set of boundary types, generating boundary fields.

[0058] In one possible implementation, boundary detection includes: Identify scale boundaries based on the scale range, applicable objects, sample range, or application scope in the key claims.

[0059] Identify time boundaries based on the time scope, data timeliness, lifecycle, disease course, engineering life, or feedback cycle in the key claims.

[0060] Identify information boundaries based on the information sources, data gaps, unobservable areas, source contamination, or version status corresponding to key claims.

[0061] Identify the boundaries of evidence based on the source type of the key claims, the degree of citation matching, the level of evidence, and the strength of the conclusion.

[0062] Identify stability boundaries based on the long-term operating conditions, material conditions, system conditions, physical conditions, platform conditions, or institutional conditions of the audited entity.

[0063] Identify solvable boundaries based on the computational difficulty of the problem, the proof status, the observation conditions, the engineering implementation conditions, or resource constraints.

[0064] Identify the closed-loop boundaries based on whether the audited entity has feedback, maintenance, versioning, accountability, write-back, rollback, or long-term correction mechanisms.

[0065] It should be noted that the scale boundary is used to identify risks arising from the scale range, applicable objects, sample range, application scope, or extrapolation of overall conclusions from partial conclusions in key claims. The time boundary is used to identify risks arising from the time range, data timeliness, lifecycle, disease progression, engineering lifespan, feedback cycle, or substitution of long-term judgments for short-term results in key claims. The information boundary is used to identify risks arising from the information sources, data gaps, unobservable areas, source contamination, version status, or incomplete information corresponding to key claims. The evidence boundary is used to identify whether there is a match between the source type, citation matching degree, evidence level, support status, and conclusion strength of key claims. The stability boundary is used to identify whether the audit object possesses stability under long-term operating conditions, material conditions, system conditions, physical conditions, platform conditions, or institutional conditions. The solvability boundary is used to identify whether the relevant problem is currently solvable in terms of computational difficulty, proof status, observation conditions, engineering implementation conditions, or resource constraints. The closed-loop boundary is used to identify whether the audit object has feedback, maintenance, versioning, accountability, write-back, rollback, or long-term correction mechanisms.

[0066] Specifically, the system can perform boundary detection based on preset trigger conditions.

[0067] The system can trigger scale boundary detection when a key claim involves extrapolating from a local sample to a global conclusion, from experimental conditions to engineering applications, or from a single object to all objects. The system can trigger time boundary detection when a key claim lacks valid time, version status, long-term records, or lifecycle conditions. The system can trigger information boundary detection when a key claim has an unknown source, insufficient source level, missing version, or missing citation location. The system can trigger evidence boundary detection when the strength of a key claim exceeds its level of evidence, or when it exhibits a Weak_Support, Indirect_Support, Related_Not_Support, Conflict, Outdated, No_Source, or Need_Review status.

[0068] Furthermore, the system can associate the boundary detection results with key claim numbers, evidence numbers, breakpoint numbers, and route numbers, and write them into the boundary field, breakpoint field, and audit log.

[0069] In this embodiment of the invention, through the aforementioned boundary detection method, this implementation can perform structured detection on key claims and audit objects from multiple dimensions such as scale, time, information, evidence, stability, solvability, and closed loop based on a preset set of boundary types, and generate boundary fields containing boundary type, boundary risk, boundary description, and suggested actions. This allows for the identification of structural risks such as extrapolating overall conclusions from local samples, extrapolating engineering applications from experimental conditions, supporting current judgments with outdated data, supporting strong conclusions with weak evidence, lack of long-term stable conditions, exceeding the current solvability range, and lack of feedback, maintenance, accountability, write-back, or rollback mechanisms. This avoids overlooking deeper knowledge risks by relying solely on surface text review or single-point fact-checking.

[0070] Reference manual attached Figure 4 This illustrates a relationship diagram for breaking point identification and correction suggestion generation provided by an embodiment of the present invention.

[0071] Breakpoint identification is performed based on the object field, evidence field, and boundary field to generate a breakpoint field.

[0072] In one possible implementation, the breakpoint includes at least one of the following: insufficient object definition, insufficient conditions for establishment, unstable structural constraints, incomplete generation chain, insufficient verification chain, missing write-back chain, unclear chain of responsibility, insufficient level of evidence, excessive strength of conclusion, and overreach of public expression.

[0073] Optionally, when a key claim lacks supporting sources or citations, the system identifies it as insufficient evidence or excessively strong conclusion. When a technical approach only demonstrates laboratory sample performance without specifying engineering scale-up, lifespan, maintenance, and cost, the system identifies it as having a break in scale boundaries, time boundaries, and closed-loop boundaries. When an AI response cannot trace the model version, source version, or human review record, the system identifies it as having an unclear chain of responsibility or a missing write-back chain.

[0074] Furthermore, when performing key claim extraction, evidence stratification, boundary detection, and breakpoint identification, the system can adopt a combination of field-based input, rule-based judgment, and structured output.

[0075] In this embodiment of the invention, through the aforementioned breakpoint identification method, this implementation can identify structural gaps in the audit object's definition, establishment conditions, structural constraints, generation chain, verification chain, write-back chain, responsibility chain, evidence level, conclusion strength, and public expression based on the correspondence between object fields, evidence fields, and boundary fields. It then generates recordable, associative, and verifiable breakpoint fields. This avoids merely providing a single risk warning, instead specifically pinpointing issues such as insufficient evidence, boundary overstepping, reasoning jumps, lack of verification, unclear responsibility, lack of write-back mechanisms, and excessive public expression to corresponding key claims, evidence sources, and boundary types. This provides a clear basis for subsequent route maturity and risk zoning, revision suggestion generation, manual review, and audit log recording.

[0076] Reference manual attached Figure 5 This illustrates a flowchart of route maturity and risk zoning provided by an embodiment of the present invention.

[0077] Based on the evidence stratification results, boundary fields, and fracture fields, the relevant models, technical routes, project plans, or expressed content are partitioned according to route maturity and risk, and route fields are generated.

[0078] In one possible implementation, route maturity and risk zoning includes classifying relevant models, technical routes, research routes, project plans, or expressions into at least one of main routes, strategic routes, cautious routes, and frozen expressions.

[0079] It should be noted that the "Main Route" is used to mark routes or expressions with strong evidence, high maturity, good closure, controllable risks, or stable application conditions. The "Strategic Route" is used to mark routes with potential but still requiring supplementary data, models, experiments, engineering conditions, or long-term validation. The "Cautious Route" is used to mark routes with high risks, insufficient evidence, unclear conditions, or that are not currently suitable for strong expression. The "Frozen Expression" is used to mark content with insufficient current evidence, conclusions that clearly exceed limits, high risks of public dissemination, or that is not suitable as definitive output.

[0080] Specifically, when a key claim has clear object boundaries, currently valid sources, a Strong_Support status, and low boundary risk, the system can classify it as the main route. When a key claim has research value but still requires supplementary experiments, long-term data, engineering validation, or feedback mechanisms, the system can classify it as the strategic route. When a key claim has insufficient evidence, unclear version, high boundary risk, or requires manual review, the system can classify it as the cautious route. When a key claim has strong conclusions without sources, unsupported citations, serious source conflicts, or overstepping public expression boundaries, the system can classify it as a frozen expression. The system can generate route fields including Route_Type, Risk_Level, Support_Status, Boundary_Risk, and Suggested_Action, which can be used to generate structured audit reports.

[0081] Generate a correction suggestion field based on the break field and the route field.

[0082] In one possible implementation, the suggested correction fields include at least one of the following: supplementing sources, supplementing data, supplementing model conditions, supplementing experiments or observations, supplementing long-term records, supplementing engineering closure loops, supplementing the chain of responsibility, reducing the strength of conclusions, labeling the hypothesis layer, adjusting public expression, and entering manual review.

[0083] Specifically, when the system identifies risks to the boundaries of evidence, it generates suggestions for supplementing sources, data, citations, or reducing the strength of the conclusion. When the system identifies risks to the closed-loop boundaries, it generates suggestions for supplementing feedback mechanisms, version records, maintenance mechanisms, chains of responsibility, or write-back mechanisms. When the system identifies that public expression has gone too far, it generates suggestions for downgrading or rewriting public expression.

[0084] Reference manual attached Figure 6 The diagram illustrates a report generation and audit log flowchart provided by an embodiment of the present invention.

[0085] A structured knowledge audit report is generated based on the object field, evidence field, boundary field, break field, route field, and correction suggestion field.

[0086] In one possible implementation, the structured knowledge audit report includes at least two of the following: audit object summary, object definition results, evidence level table, boundary detection table, breakpoint list, route partitioning results, correction suggestion list, and audit log summary.

[0087] The structured knowledge audit report is output in at least one of the following formats: webpage, PDF, Word, Markdown, table, JSON, or API response.

[0088] Specifically, the system generates a structured knowledge audit report based on object fields, evidence fields, boundary fields, breakpoint fields, route fields, and correction suggestion fields. The structured knowledge audit report includes at least two of the following: audit object summary, object definition results, evidence level table, boundary detection table, breakpoint list, route partitioning results, correction suggestion list, and audit log summary.

[0089] In this embodiment of the invention, through the aforementioned structured knowledge audit report generation method, this implementation can uniformly summarize object fields, evidence fields, boundary fields, breakpoint fields, route fields, and correction suggestion fields to form a structured knowledge audit report containing audit object summaries, object definition results, evidence level tables, boundary detection tables, breakpoint lists, route partitioning results, correction suggestion lists, and audit log summaries. Thus, the dispersed audit process and audit results can be transformed into output files with a unified format, clear hierarchy, ease of viewing, comparison, and review, enabling users to clearly understand the core objects, evidence status, boundary risks, structural breaks, route partitioning, and correction directions within the audited content.

[0090] In one possible implementation, it also includes: Save audit logs, which include at least one of the following: input content, input content hash value, document version, model version, rule version, field generation result, boundary detection result, report version, report generation time, user modification record, and manual review record.

[0091] Audit logs can be used to relocate affected key claims, evidence fields, boundary fields, break fields, and route fields when document versions, model versions, rule versions, knowledge base versions, or user variables change, and generate re-audit, error write-back, or rollback tasks.

[0092] Specifically, the system can save audit logs to record input content, model calls, rule versions, knowledge base versions, field generation results, manual review processes, and report output processes. When document versions, rule versions, model versions, knowledge base versions, or user variables change, the system can relocate the affected key claims, evidence fields, boundary fields, and route fields based on the audit logs, and generate re-audit, error write-back, or rollback tasks.

[0093] In one possible implementation, it also includes: When changes are detected in object definitions, evidence status, breakpoints, route partitions, or output boundaries in a user's long-term project, a variable update item to be confirmed is generated.

[0094] After receiving user confirmation, the variable update item to be confirmed will be written into the user's project variable library.

[0095] Specifically, when the system detects changes in object definitions, evidence status, breakpoints, route partitions, or output boundaries in a user's long-term project, it generates a variable update item to be confirmed. This item includes the variable name, original variable value, new variable value, reason for change, associated document, associated audit report, associated breakpoint, associated route partition, and user confirmation status. Upon receiving user confirmation, the variable update item is written to the user project variable library.

[0096] The variable update items include at least one of the following: user project number, historical object definition, historical evidence status, historical breakpoint, historical route partition, newly added data, current variable change, and user confirmation status.

[0097] Example 2: The published version in this embodiment is only to illustrate one application of the present invention in the scenario of auditing texts in public communication. The published version can also be replaced by a white paper, research report, project specification, corporate knowledge base output or other text materials containing judgmental expressions.

[0098] In this embodiment, the content to be audited is a published manuscript about AI-assisted scientific literature review. The manuscript includes a title, main text, several paragraph descriptions, citation markers, and judgmental statements regarding the capabilities of the AI ​​system, the credibility of the retrieval enhancement generation system, the efficiency of automatically generated reviews, and the effectiveness of scientific literature review. After receiving the manuscript, the system performs format recognition, text extraction, paragraph segmentation, title recognition, citation marker recognition, source link recognition, timestamp recognition, and version information recognition to obtain preprocessed text.

[0099] The system can identify claims in published manuscripts regarding the capabilities of AI-assisted research literature review as audit targets. It can categorize these targets as public communication texts or assessments of AI capabilities, identify the relevant field as AI knowledge auditing or research assistance tools, and determine whether the manuscript faces risks such as insufficient evidence, overly strong conclusions, inadequate citation support, overreaching public expression, or missing chains of responsibility. The system further defines the target boundaries as content in the manuscript related to AI-assisted research literature review, credibility enhancement systems for search generation, and the automatic generation of reviews and peer reviews. Background information, general promotional material, or paragraphs unrelated to the audit objectives are marked as excluded.

[0100] The system generates object fields. Object fields may include object number, object name, object type, domain, audit objective, audit task, object boundaries, exclusions, and contextual summary. Object fields are used to define the specific scope of this audit and serve as the basis for subsequent key claim extraction, evidence stratification, boundary detection, breakpoint identification, route maturity and risk zoning, and the generation of corrective recommendations.

[0101] The system then extracts key claims from the preprocessed text. In this embodiment, the system can extract the following types of key claims: conclusive claims about whether artificial intelligence systems can solve the problem of reviewing scientific literature; technical claims about whether search enhancement generation systems can ensure credible answers; roadmap claims about whether AI-generated reviews can replace human peer review; and public communication expressions about how AI tools improve efficiency, reduce costs, or change the review process.

[0102] The system generates claim fields for each key claim. Claim fields may include claim number, claim text, claim type, paragraph, claim strength, and associated audit object. The system further associates each key claim with its paragraph, citation location, source object, source version, and context. Sources may include published manuscript paragraphs, references, web links, research reports, open papers, system documentation, knowledge base fragments, search-enhanced fragments, or contextual paragraphs. The system can generate a source number for each source and record the source type, source date, source version, source location, and source status. The system can generate a citation number for each citation and record the citation text, citation location, citation context, and corresponding source number. Based on object consistency, scope consistency, time consistency, conclusion consistency, and contextual consistency between key claims and citations, the system generates the association status between claims and citations.

[0103] For example, for strong conclusions such as "AI systems will completely solve the problem of scientific literature review," if the published manuscript does not provide sufficient sources, or only provides general cases, technical introductions, or descriptions of efficiency improvements, the system can mark the claim as having no clear supporting sources or weak support, and record the evidence gap. This evidence gap can include the lack of long-term effect data, error rate assessments, human review mechanisms, disciplinary difference analysis, peer review responsibility chains, and verification through real scientific research processes.

[0104] For statements such as "search enhancement generation systems can ensure completely credible answers," if the source can only explain that the search enhancement generation system can improve the traceability of answers or reduce the risk of generating answers without a source, but cannot prove that it reaches a completely credible level, the system can mark the claim as partially supported, weakly supported, or relevant but not supported, and record gaps in source classification, version verification, citation support verification, manual review, error write-back, and chain of responsibility recording. For statements such as "AI-generated reviews can replace human peer review," if the source can only explain that AI-generated reviews can improve the efficiency of literature review, but cannot prove that it can assume the functions of complete academic judgment, ethical responsibility, conflict of interest identification, and community evaluation, the system can mark the claim as an assumption, a public expression, or content requiring manual review.

[0105] The system performs evidence stratification. In this embodiment, the system marks basic facts directly supported by clear sources as the fact layer. Content with data, statistics, or evaluation as the data layer. Content based on model capabilities, system architecture, or technical specifications as the model layer. Content further derived from facts, data, or models as the reasoning layer. Content that is not yet fully verified but has discussion value as the hypothesis layer. Analogical, rhetorical, or explanatory expressions as the metaphor layer. Simplified expressions intended for public communication as the public expression layer. Expressions with insufficient current evidence, significantly excessive conclusion strength, or that may mislead the public are marked as the frozen layer or content requiring review.

[0106] The system performs boundary checks. In this embodiment, the system performs checks on key claims and audit objects based on a preset set of boundary types. For strong conclusions expressed by the AI ​​system in addressing research literature review issues, the system can identify risks related to evidence boundaries, time boundaries, stability boundaries, and closed-loop boundaries. Evidence boundary risk lies in the published manuscript not providing sufficient evidence to support a definitive conclusion. Time boundary risk lies in the published manuscript not explaining the long-term operation, error accumulation, version changes, and stability of the AI ​​system. Stability boundary risk lies in the published manuscript not explaining the applicable conditions of the AI ​​system in different disciplines, different document types, and different review tasks. Closed-loop boundary risk lies in the published manuscript not explaining the human review of the AI ​​review results, error feedback, responsibility attribution, and version write-back mechanisms.

[0107] For retrieval enhancement systems that generate credibility expressions, the system can identify risks related to information boundaries, evidence boundaries, and closed-loop boundaries. Information boundary risks include issues such as unclear source level, ambiguous version status, old data supporting new conclusions, source contamination, or retrieval omissions. Evidence boundary risks lie in the fact that the existence of a citation does not equate to its support for every key claim. Closed-loop boundary risks occur when the published manuscript fails to explain how errors in the response are written back to the knowledge base for maintenance, retrieval strategies, manual review queues, or rule update mechanisms.

[0108] Regarding AI-generated reviews replacing peer review, the system can identify the boundaries of evidence, responsibility, and the risk of public expression overstepping these boundaries. The risk of the evidence boundary lies in the fact that increased efficiency cannot directly replace complete peer review. The risk of the responsibility boundary lies in the fact that peer review involves not only textual summarization but also professional judgment, academic responsibility, ethical review, conflict of interest identification, and community evaluation. The risk of public expression overstepping these boundaries lies in the fact that related expressions may mislead readers into believing that the AI ​​system possesses the full capacity for academic review responsibility.

[0109] The system identifies breakpoints based on object fields, evidence fields, and boundary fields. In this embodiment, the system can identify the following breakpoints: Overly strong conclusions, i.e., published manuscripts use strong expressions such as "completely solves," "completely credible," or "replaces human review," but fail to provide sufficient evidence. Insufficient evidence levels, i.e., lack of long-term data, real-world scenario verification, error rate statistics, human review results, and chain of responsibility records. Incomplete generation chain, i.e., directly deriving "AI will solve the problem of scientific literature review" from "AI can assist in retrieval or generate reviews." Insufficient verification chain, i.e., failing to explain how AI output undergoes expert review, peer review, version comparison, domain adaptation testing, or verification through real-world research processes. Missing write-back chain, i.e., failing to explain how errors in AI generation, citation mismatches, or retrieval errors are written back to the knowledge base, rule base, model hint strategies, or human review process. Overreaching public expression, i.e., expressing potentially valid auxiliary capabilities as definitive replacement capabilities.

[0110] System execution roadmap maturity and risk zoning. In this embodiment, the system can classify "AI-assisted literature retrieval," "AI-assisted literature organization," and "AI-assisted preliminary review generation" as main roadmap enhancements or strategic roadmaps. "AI-automated review generation assisting human review under certain conditions" is classified as a strategic roadmap. "AI replacing human peer review" is classified as a cautious roadmap. Absolute expressions such as "AI completely solves the problem of scientific literature review" or "the retrieval enhancement generation system ensures completely credible answers" are classified as frozen expressions or high-risk public expressions.

[0111] The system generates correction suggestions based on the breakpoints and route partitioning results. In this embodiment, the system can generate the following correction suggestions: rewrite absolute expressions into limiting expressions, such as changing "completely solves" to "helps reduce some risks" or "improves traceability, auditability, and correctability." Rewrite "completely credible" to "improves credibility under the conditions of traceable sources, verifiable citations, verifiable versions, and the existence of human review mechanisms." Rewrite "replaces human peer review" to "assists in data preparation, preliminary review, and consistency checks before human peer review." Supplement data sources, literature years, knowledge base versions, model versions, citation locations, and supporting relationships. Add mechanisms for human review, citation verification, error marking, version updates, and audit log recording. Add a description of the boundaries of public expression to avoid describing AI-assisted capabilities as complete replacement capabilities.

[0112] The system generates structured knowledge audit reports based on object fields, evidence fields, boundary fields, breakpoint fields, route fields, and correction suggestion fields. The report may include an audit object summary, a list of key claims, an evidence level table, a boundary detection table, a breakpoint list, route partitioning results, a correction suggestion list, and an audit log summary. The system saves audit logs, which may include input hash values, document versions, model versions, rule versions, field generation results, boundary detection results, report versions, report generation time, user modification records, and manual review records.

[0113] Through this embodiment, the present invention can convert a published manuscript with strong conclusions from natural language text into an auditable data object, and generate a verifiable structured knowledge audit report through object recognition, key claim extraction, evidence stratification, boundary detection, breakpoint identification, route maturity and risk zoning, and correction suggestions. This reduces the risks of excessive promotion of artificial intelligence capabilities, insufficient evidence, overreach of public expression, and missing chain of responsibility in the published manuscript.

[0114] Example 3: In this embodiment, the content to be audited is a technical route document regarding the potential of a novel material to replace existing high-temperature alloys. This technical route document includes laboratory sample performance data, performance comparison charts, application prospect descriptions, and judgmental statements regarding its engineering substitution prospects. After receiving the technical route document, the system performs format recognition, text extraction, paragraph segmentation, table parsing, chart description recognition, source identification, timestamp identification, and version information identification to obtain preprocessed text.

[0115] The novel material in this embodiment is merely an example illustrating the technical process of the present invention. This embodiment does not require specific material composition, preparation process, experimental parameters, test thresholds, real project data, customer information, or undisclosed technical details. The system can generate structured syntax fields based on the text, data, source, citation, version, and context in the input content, and perform auditing according to preset rules.

[0116] The system identifies the audit target based on the preprocessed text. In this embodiment, the system identifies the core audit target as "the route for replacing existing high-temperature alloys with new materials," identifies the object type as a technical route or an engineering materials route, identifies the field as materials engineering, and identifies the audit objective as determining whether the technical route has risks such as insufficient evidence, engineering extrapolation, overestimation of route maturity, lack of long-term service data, inadequate manufacturing and maintenance conditions, or overreaching public expression. The system further defines the object boundary as the relevant content in the material regarding the performance of the new material, its replacement of existing high-temperature alloys, engineering application conditions, long-term service prospects, and the maturity of the technical route.

[0117] The system generates object fields. These fields may include object number, object name, object type, domain, audit objective, audit task, object boundaries, exclusions, and a contextual summary. The object fields ensure that the system audits the specific technical route of "new materials replacing high-temperature alloys," rather than a general evaluation of the entire materials science field or all high-temperature alloy systems. The system then extracts key claims from the preprocessed text. In this embodiment, the system may extract the following key claims: the laboratory sample performance of the new material is superior to some control indicators under specified conditions; the new material has the potential to replace existing high-temperature alloys in engineering; the new material can comprehensively replace existing high-temperature alloy systems; and the new material can be rapidly applied in industrial applications.

[0118] The system generates evidence fields for each key claim. For claims related to laboratory sample performance, the system can associate them with test data, charts, test conditions, sample batches, sample sizes, test environments, and data sources, classifying them into a data layer. For claims related to engineering substitution, the system identifies that the claim requires support from engineering component verification, manufacturing processes, long-term service, oxidation corrosion, fatigue creep, maintenance cycles, cost conditions, and real-world service environments, thus classifying them into a reasoning layer or an assumption layer. For strong expressions such as "complete substitution" or "rapid industrialization," the system identifies them as high-intensity public expressions, and the scope of the conclusions significantly exceeds the range that laboratory sample data can directly support, thus marking them as public expression layers, cautious routes, or frozen expressions. In one implementation, the system can generate evidence fields for key claims of the following types: claim number, claim text, source object, source type, source date, source version, citation location, evidence level, support status, and evidence gap. For example, laboratory performance data may support the statement "possesses performance potential under limited test conditions," but is insufficient to directly support the conclusions "comprehensive engineering substitution conditions are already met" or "can comprehensively replace the existing system."

[0119] The system performs evidence stratification. In this embodiment, the system divides laboratory sample performance into a data layer; engineering substitution judgments into a reasoning layer or hypothesis layer; and high-intensity expressions such as "complete replacement of the existing system" or "rapid industrialization" are identified as public expression layers or frozen layers. The system further determines that the content of the data layer can support performance judgments under limited conditions, but cannot directly support a comprehensive substitution conclusion across scales, time periods, operating conditions, and industrial chains.

[0120] The system performs boundary checks. In this embodiment, the system checks key claims and audit objects based on a preset set of boundary types. For laboratory sample performance claims, the system can identify information boundary and evidence boundary risks. Information boundary risks lie in the possibility that experimental data may lack information on sample origin, testing conditions, batch variations, repeated experiments, and third-party verification. Evidence boundary risks lie in the possibility that a single laboratory performance indicator cannot directly support conclusions for engineering applications.

[0121] For claims of engineering substitution, the system can identify risks related to scale boundaries, time boundaries, stability boundaries, and closed-loop boundaries. Scale boundary risks lie in the issues of size scaling, forming, microstructure homogeneity, defect control, and batch stability between laboratory samples and engineering components. Time boundary risks involve the lack of long-term creep, fatigue, oxidation, corrosion, thermal cycling, and service life data for the material. Stability boundary risks include the lack of descriptions regarding phase stability, interface stability, grain boundary stability, coating compatibility, and service environment suitability. Closed-loop boundary risks involve the lack of a feedback loop for manufacturing, testing, maintenance, failure analysis, repair, recycling, and engineering applications. For public statements advocating "comprehensive substitution," the system can identify risks related to scale boundaries, evidence boundaries, stability boundaries, closed-loop boundaries, and exceeding the boundaries of public statements. These claims extrapolate from local experimental performance to the entire high-temperature alloy system, resulting in a broad scope of conclusions without specifying the exact substitute, applicable temperature, service environment, component type, process route, and industrialization conditions.

[0122] The system identifies breakpoints based on evidence stratification and boundary detection results. In this embodiment, the system can identify the following breakpoints: insufficient scale extrapolation, i.e., lack of engineering scale-up conditions when extrapolating from laboratory sample performance to engineering component replacement; insufficient time verification, i.e., lack of long-term creep, fatigue, oxidation, corrosion, thermal cycling, and service life data; insufficient stability conditions, i.e., lack of descriptions of phase stability, interface stability, grain boundary stability, and service environment adaptation; lack of engineering closed-loop, i.e., lack of manufacturing processes, testing procedures, maintenance cycles, failure analysis, repair paths, and recycling mechanisms; excessively strong conclusions, i.e., expressions such as "complete replacement" significantly exceed the scope of existing data support; and excessive public expression, i.e., the technical route materials describe strategic potential as a definitive replacement conclusion, which may lead to misjudgment of technology maturity.

[0123] The system implements route maturity and risk zoning. In this embodiment, the system can classify existing mature high-temperature alloy routes as main routes because they have a relatively mature engineering application foundation, long-term service data, manufacturing processes, maintenance systems, and industrial closed loops. The system can classify new material routes as strategic routes because these routes show potential in laboratory performance but still require supplementary engineering component verification, long-term life testing, service environment data, manufacturing process verification, and maintenance closed loops. The system can classify expressions such as "completely replacing existing high-temperature alloy systems" as cautious routes or frozen expressions.

[0124] The system generates route fields. Route fields can include route object, route type, zoning basis, maturity, strength of evidence, degree of closure, and public communication risk. The system further generates correction suggestions based on the route fields. In this embodiment, the system can generate the following correction suggestions: supplement service environment data, including temperature range, load conditions, oxidation environment, corrosive media, thermal cycling conditions, and application component type; supplement long-term life testing, including long-term creep, fatigue, thermal fatigue, oxidation, corrosion, microstructure evolution, and failure mode data; supplement engineering component verification, including component size enlargement, processing and forming, microstructure uniformity, defect control, coating matching, and bonding processes; supplement manufacturing and maintenance conditions, including preparation processes, quality inspection, maintenance cycles, failure analysis, repair paths, and recycling mechanisms; reduce the strength of the conclusion, rewriting expressions like "comprehensive replacement" to "has potential application value under specific temperature, load, and service environments, but still requires long-term service data and engineering closure verification"; supplement route zoning descriptions, presenting mature material routes, strategic material routes, cautious expressions, and frozen expressions separately.

[0125] The system generates a structured knowledge audit report. The report includes a summary of the audited objects, a list of key claims, an evidence level table, a boundary detection table, a list of breakpoints, route partitioning results, a list of suggested corrections, and a summary of the audit log. The system saves the audit log, recording the input hash value, document version, rule version, model version, field generation results, report generation time, and manual review records.

[0126] Through this embodiment, the present invention can perform hierarchical processing of experimental data, engineering reasoning, and public communication expressions in technical route materials, and identify the scale boundaries, time boundaries, stability boundaries, and closed-loop boundaries between laboratory performance and engineering substitution, thereby reducing the risks of excessive promotion of technical routes, misjudgment of engineering maturity, and overstepping of public expression boundaries.

[0127] Example 4: In this embodiment, the content to be audited is an AI-generated answer. This AI-generated answer includes multiple cited fragments, source links, or search-enhanced results, but it is necessary to determine whether the cited content truly supports the key conclusions in the answer. After receiving the AI-generated answer, citation list, source links, document fragments, or search-enhanced results, the system performs format recognition, text extraction, citation marker recognition, source link recognition, timestamp recognition, and version information recognition on the content to obtain preprocessed text.

[0128] This embodiment does not require specific retrieval strategies, model call suggestions, internal matching thresholds, knowledge base structure details, or undisclosed source library content as necessary conditions. The system can generate structured syntax fields based on AI answers, cited fragments, source links, source versions, retrieval time, and contextual information, and perform claim-citation support relationship auditing accordingly.

[0129] The system identifies the audit target based on the preprocessed text. In this embodiment, the system identifies the audit target as "AI answer credibility audit," the object type as AI-generated content or search-enhanced output, the domain as AI knowledge audit, and the audit objective as determining whether the AI ​​answer has issues such as false citations, weak citations with strong expressions, outdated sources, unknown versions, mismatched evidence, claims without sources, insufficient information boundaries, insufficient evidence boundaries, and insufficient closed-loop boundaries.

[0130] The system generates object fields. Object fields may include object number, object name, object type, domain, audit objective, audit task, object boundaries, excluded content, and contextual summary. Object boundaries can be limited to the core judgments, cited excerpts, source documents, and supporting relationships between the answer's conclusions in the AI ​​response, excluding small talk, formatted expressions, or non-advocative content unrelated to the answer's credibility.

[0131] The system extracts key claims from AI-generated answers. Key claims can include factual statements, definitional statements, technical statements, policy statements, comparative statements, recommendation statements, and conclusion statements. The system associates each key claim with its corresponding paragraph, citation, source link, source document, timestamp, and context.

[0132] In this embodiment, the system can extract key claims of the following types: a technical solution has been proven to operate stably in a production environment; a policy document explicitly permits a certain operation; a research conclusion has been consistently supported by multiple publications. The system generates evidence fields for the above key claims. For each key claim, the system records the citation location, source type, source date, source version, support status, and evidence gap. If a key claim has no supporting citations or source fragments, the system marks it as a claim without a source. If a cited fragment is only relevant to the topic but does not support the conclusion, the system marks it as relevant but not supported. If a cited fragment comes from an older or obsolete version of the material, the system marks it as an outdated source or a version mismatch. If a cited fragment contains contradictory content to the key claim, the system marks it as conflicting evidence.

[0133] In one implementation, the system can generate an assertion number for each key claim, a source number for each source, and a citation number for each cited fragment, and establish associations between assertion numbers, source numbers, citation numbers, version numbers, and contextual text. The system generates the association status between claims and citations based on object consistency, temporal consistency, scope consistency, conclusion consistency, and contextual consistency between the cited fragment and the key claim. Statuses can include direct support, partial support, context-relevant only, weak support, mismatch, conflict, unverifiable, or no citation.

[0134] The system performs evidence stratification. In this embodiment, the system classifies content with a clear source and directly supported by the cited fragment into the fact layer or data layer; content further deduced based on the source fragment into the reasoning layer; content lacking direct support but possessing possibility into the hypothesis layer; simplified explanations for users into the public expression layer; and definitive conclusions with insufficient evidence and high risk into the frozen layer or content requiring manual review.

[0135] The system performs boundary detection. In this embodiment, the system can identify at least the following boundary risks: insufficient information boundary, i.e., some key claims lack sources, or the cited sources do not show the version, time, source level, and validity status; insufficient evidence boundary, i.e., some cited fragments exist, but are only related to the topic and cannot directly support the conclusions in the answer; insufficient time boundary, i.e., some cited materials were published earlier and may not be applicable to the current question or the current version; insufficient closed-loop boundary, i.e., the AI ​​answer does not explain the error feedback, manual review, citation correction, source replacement, or report correction mechanism; insufficient chain of responsibility, i.e., the AI ​​answer does not record the model version, knowledge base version, retrieval time, cited fragment version, manual reviewer, or user confirmation status.

[0136] The system identifies breakpoints based on evidence stratification and boundary detection results. In this embodiment, the system can identify the following breakpoints: citations do not support the conclusion, meaning that although the cited fragment exists, its content cannot directly prove the key claim in the AI ​​response; the source is outdated, meaning the AI ​​response uses old versions, expired materials, or materials without a valid status to support the current conclusion; there is no version description, meaning the AI ​​response does not record the source document version, knowledge base version, model version, or retrieval time; the level of evidence is insufficient, meaning the AI ​​response presents weakly supported, indirectly supported, or related but unsupported material as strong evidence; the conclusion is too strong, meaning the AI ​​response uses strong conclusion expressions such as "already proven," "explicitly permitted," or "consistently supported," but the evidence is insufficient; the write-back chain is missing, meaning the AI ​​response does not provide error feedback, citation correction, manual review, or version update mechanisms.

[0137] System execution route maturity and risk zoning. In this embodiment, the system can classify "AI-assisted answers with sources" as the main route enhancement; "response generation enhanced by traceable retrieval and verified by citation support" as the main route or strategic route; "AI-generated answers without sources" as the cautious route; and "responses that have citations but do not support the conclusion and still output a strong conclusion" as frozen expression or high-risk output.

[0138] The system generates correction suggestions based on the breakpoints and route partitioning results. In this embodiment, the system can generate the following correction suggestions: supplement genuine sources, providing source documents or citation fragments that directly support each key claim; delete false citations, removing citation relationships that exist but do not support the conclusion; replace outdated sources, replacing old versions, obsolete versions, or sources with unclear status with currently valid versions; reduce the strength of the conclusion, changing "has been proven" to "shows potential support in some materials", "explicitly allows" to "requires further confirmation in conjunction with currently valid documents", and "consistently supports" to "partially supported by sources"; supplement version information, including document version, knowledge base version, model version, retrieval time, and report generation time; add manual review, initiating manual review processes for high-risk claims, source conflicts, unclear versions, or citations that do not support the conclusion; establish an error write-back mechanism, writing discovered citation mismatches, outdated sources, claims without sources, or weak support for strong statements into the knowledge base maintenance queue, retrieval strategy adjustment queue, or rule update record.

[0139] The system generates a structured knowledge audit report. The report may include object cards, evidence cards, boundary cards, break cards, route cards, correction cards, and an audit log summary. Object cards record the audit object name, object type, domain, audit objective, and object boundaries. Evidence cards record the source, citation, version, evidence level, and support status of each key claim. Boundary cards record risks related to information boundaries, evidence boundaries, time boundaries, and closed-loop boundaries. Break cards record citations that do not support the conclusion, outdated sources, lack of version information, insufficient evidence level, and missing write-back chains. Route cards record route partitions for AI responses, retrieval-enhanced responses, responses generated without sources, and pseudo-citation outputs. Correction cards provide suggestions such as supplementing with genuine sources, deleting pseudo-citations, reducing conclusion strength, supplementing version information, and initiating manual review. The audit log summary records the input hash value, model version, rule version, cited version, report generation time, user modification records, and manual review records.

[0140] Through this embodiment, the present invention can perform structured auditing of the supporting relationships between key claims and cited fragments in AI-generated answers, identify problems such as citations that do not support conclusions, outdated sources, lack of version information, insufficient evidence boundaries, and insufficient closed-loop boundaries, and output a verifiable structured knowledge audit report, thereby reducing the risks of false citations, evidence mismatches, over-conclusions, and irresponsible outputs in AI answers.

[0141] Examples two through four illustrate the application of this invention in knowledge auditing of published manuscripts, technical route auditing, and AI response auditing, respectively. While the audit targets differ across the three examples, all utilize the same structured technical process to complete the knowledge audit: receiving the content to be audited, preprocessing the content, identifying the audit target, extracting key claims, generating structured syntax fields, performing evidence layering, performing boundary detection, identifying breakpoints, performing route maturity and risk partitioning, generating correction suggestions, outputting a structured knowledge audit report, and saving the audit log. In the published manuscript knowledge auditing scenario, this invention can identify strong conclusions, insufficient evidence, leaps in reasoning, and risks of public expression exceeding boundaries in publicly disseminated texts. In the technical route auditing scenario, this invention can identify risks related to scale boundaries, time boundaries, stability boundaries, and closed-loop boundaries between laboratory data and engineering alternatives. In the AI ​​response auditing scenario, this invention can identify risks such as citations not supporting conclusions, outdated sources, unclear versions, insufficient evidence boundaries, and missing write-back chains.

[0142] Therefore, this invention is not limited to a specific field, a specific document format, a specific model, a specific prompt template, or a specific internal rule. Instead, it transforms different types of knowledge materials into auditable data objects through a set of structural grammar fields. Furthermore, it improves the traceability, auditability, and correctability of AI-generated content, knowledge base output, published texts, research materials, and technical route materials through a unified process of field generation, evidence layering, boundary detection, breakpoint identification, route partitioning, and correction suggestion generation.

[0143] Reference manual attached Figure 7 The diagram shows the overall architecture of an AI knowledge auditing system based on structured syntax fields provided by this invention.

[0144] This invention also provides an AI knowledge auditing system based on structured syntax fields, comprising: The user interaction interface module provides a human-computer interaction interface for user input, audit task initiation, audit report display, user confirmation, and variable update confirmation.

[0145] The input receiving module is used to receive user input.

[0146] The preprocessing module is used to preprocess user input to obtain preprocessed text.

[0147] The object recognition module is used to identify audit objects in preprocessed text and generate object fields.

[0148] The key claim extraction module is used to extract key claims from preprocessed text and generate claim fields.

[0149] The structure syntax field generation module is used to generate a set of structure syntax fields based on the source, reference, data, version information or context corresponding to the object field, claim field, and key claim; wherein the set of structure syntax fields includes at least one or more of the following: object field, claim field, evidence field, boundary field, break field, route field, correction suggestion field, report field or log field.

[0150] The evidence stratification module is used to stratify key claims based on preset evidence stratification rules and obtain evidence stratification results.

[0151] The boundary detection module is used to perform boundary detection on key claims and audit objects based on a preset set of boundary types, and generate boundary fields.

[0152] The breakpoint identification module is used to identify breakpoints based on object fields, evidence fields, and boundary fields, and generate breakpoint fields. The evidence fields include at least one or more of the following: source information, citation information, data information, version information, evidence level information, and support status information corresponding to key claims.

[0153] The route partitioning module is used to partition relevant models, technical routes, project plans, or expression content based on evidence stratification results, boundary fields, and fracture fields to determine route maturity and risk, and generate route fields.

[0154] The correction suggestion generation module is used to generate correction suggestion fields based on the break field and the route field.

[0155] The report generation module is used to generate structured knowledge audit reports based on object fields, evidence fields, boundary fields, break fields, route fields, and correction suggestion fields.

[0156] The audit log module is used to record one or more of the following: audit object, key claims, set of structured syntax fields, evidence stratification results, boundary detection results, breakpoint identification results, route maturity and risk zoning results, remedial suggestions, audit report version, generation time, user confirmation status, and system processing status, forming a traceable audit log.

[0157] The optional variable update module is used to generate variable update items based on the audit report, break field, route field and correction suggestion field after user confirmation, and write the variable update items into the user variable library, project variable library or historical audit record.

[0158] The AI ​​knowledge auditing system based on structured syntax fields provided by this invention can execute the above-mentioned AI knowledge auditing method based on structured syntax fields and achieve the same or similar technical effects. To avoid duplication, this invention will not elaborate further.

[0159] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An AI knowledge auditing method based on structured syntax fields, characterized in that, The method is executed by at least one processor and includes: Receive user input; The user input is preprocessed to obtain preprocessed text; The preprocessed text is subjected to audit object identification to generate object fields; Key claims are extracted from the preprocessed text to generate claim fields for subsequent evidence field generation, evidence layering, boundary detection, breakpoint identification, route partitioning, and structured knowledge audit report generation. Evidence fields are also generated based on the source, data, citation, or context corresponding to the key claims. Based on preset evidence hierarchy rules, the key claims are layered into evidence layers to obtain evidence layering results; Boundary detection is performed on the key claims and the audit objects based on a preset set of boundary types to generate boundary fields; Breakpoint identification is performed based on the object field, the evidence field, and the boundary field to generate a break field; Based on the evidence stratification results, the boundary fields, and the fracture fields, the relevant models, technical routes, project plans, or expression contents are partitioned by route maturity and risk, and route fields are generated. Based on the fracture field and the route field, a correction suggestion field is generated; A structured knowledge audit report is generated based on the object field, the evidence field, the boundary field, the break field, the route field, and the correction suggestion field.

2. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The object fields must include at least the object name, object type, domain, audit target, and object boundaries; The object types include at least one of the following: theoretical models, technical routes, engineering systems, medical texts or health information content, published texts, AI answers, knowledge base entries, platform mechanisms, project plans, and public communication texts.

3. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The key claim extraction specifically includes: Identify at least one of the following in the user input content: factual judgment, data judgment, model judgment, causal judgment, route judgment, risk judgment, conclusion judgment, and public communication expression; Each of the key claims is associated with the text paragraph, source, citation, timestamp, and context in which the key claim is located.

4. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The evidence stratification results include evidence stratification markers generated for each of the key claims; The evidence hierarchy markers include at least two of the following: fact layer markers, data layer markers, model layer markers, reasoning layer markers, hypothesis layer markers, metaphor layer markers, public expression layer markers, and frozen layer markers; The evidence field includes support status information, which includes at least one of the following: strong support, weak support, indirect support, relevant but not supported, conflicting, outdated, without source, and requiring manual review.

5. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The boundary detection includes: Identify scale boundaries based on the scale range, applicable objects, sample range, or application scope in the key claims; Identify time boundaries based on the time range, data timeliness, life cycle, disease course, engineering life, or feedback cycle in the key claims mentioned above; Identify information boundaries based on the information sources, data gaps, unobservable areas, source contamination, or version status corresponding to the key claims; Identify the boundaries of evidence based on the source type, citation matching degree, evidence level, and conclusion strength of the key claims; Identify stability boundaries based on the long-term operating conditions, material conditions, system conditions, physical conditions, platform conditions, or institutional conditions of the audited entity; Identify solvable boundaries based on the computational difficulty, proof status, observation conditions, engineering implementation conditions, or resource constraints of the problem; Identify the closed-loop boundary based on whether the audited object has feedback, maintenance, versioning, accountability, write-back, rollback, or long-term correction mechanisms.

6. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The breakpoints include at least one of the following: insufficient object definition, insufficient conditions for establishment, unstable structural constraints, incomplete generation chain, insufficient verification chain, missing write-back chain, unclear chain of responsibility, insufficient level of evidence, excessive strength of conclusion, and overreach of public expression.

7. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The route maturity and risk zoning includes classifying relevant models, technical routes, research routes, project plans, or expression content into at least one of the following: main route, strategic route, cautious route, and frozen expression.

8. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The suggested correction fields include at least one of the following: supplementing sources, supplementing data, supplementing model conditions, supplementing experiments or observations, supplementing long-term records, supplementing engineering closure loops, supplementing the chain of responsibility, reducing the strength of conclusions, labeling the hypothesis layer, adjusting public expression, and entering manual review.

9. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The structured knowledge audit report includes at least two of the following: audit object summary, object definition results, evidence level table, boundary detection table, breakpoint list, route partitioning results, correction suggestion list, and audit log summary.

10. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, The structured knowledge audit report is output in at least one of the following formats: webpage, PDF, Word, Markdown, table, JSON, or API response.

11. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, Also includes: Save audit logs, which include at least one of the following: input content, input content hash value, document version, model version, rule version, field generation result, boundary detection result, report generation time, user modification record, and manual review record; Audit logs can be used to relocate affected key claims, evidence fields, boundary fields, break fields, and route fields when document versions, model versions, rule versions, knowledge base versions, or user variables change, and generate re-audit, error write-back, or rollback tasks.

12. The AI ​​knowledge auditing method based on structured syntax fields according to claim 1, characterized in that, Also includes: When changes are detected in object definitions, evidence status, breakpoints, route partitions, or output boundaries in a user's long-term project, a variable update item to be confirmed is generated. The variable update item to be confirmed includes the variable name, original variable value, new variable value, reason for change, associated document, associated audit report, associated breakpoint, associated route partition, and user confirmation status. After receiving user confirmation, the variable update item to be confirmed is written into the user's project variable library.

13. An AI knowledge auditing system based on structured syntax fields, characterized in that, include: The user interaction interface module is used to provide human-computer interaction interfaces for user input, audit task initiation, audit report display, user confirmation, and variable update confirmation. The input receiving module is used to receive user input. The preprocessing module is used to preprocess the user input to obtain preprocessed text; The object recognition module is used to identify audit objects in the preprocessed text and generate object fields; The key claim extraction module is used to extract key claims from the preprocessed text and generate claim fields; The structure syntax field generation module is used to generate a set of structure syntax fields based on the source, reference, data, version information or context corresponding to the object field, claim field, and key claim; wherein the set of structure syntax fields includes at least one or more of the following: object field, claim field, evidence field, boundary field, break field, route field, correction suggestion field, report field or log field; The evidence stratification module is used to stratify the key claims based on preset evidence stratification rules to obtain evidence stratification results. The boundary detection module is used to perform boundary detection on the key claims and the audit object based on a preset set of boundary types, and generate boundary fields; The breakpoint identification module is used to identify breakpoints based on the object field, the evidence field, and the boundary field, and generate breakpoint fields. The evidence field includes at least one or more of the following: source information, citation information, data information, version information, evidence level information, and support status information corresponding to the key claim. The route partitioning module is used to partition relevant models, technical routes, project plans, or expression content based on the evidence stratification results, the boundary fields, and the breakpoint fields to determine the route maturity and risk, and generate route fields. The correction suggestion generation module is used to generate correction suggestion fields based on the break field and the route field; The report generation module is used to generate a structured knowledge audit report based on the object field, the evidence field, the boundary field, the break field, the route field, and the correction suggestion field; The audit log module is used to record one or more of the following: audit object, key claims, set of structured syntax fields, evidence stratification results, boundary detection results, breakpoint identification results, route maturity and risk zoning results, remedial suggestions, audit report version, generation time, user confirmation status, and system processing status, forming a traceable audit log; The optional variable update module is used to generate variable update items based on the audit report, break field, route field and correction suggestion field after user confirmation, and write the variable update items into the user variable library, project variable library or historical audit record.

14. The AI ​​knowledge auditing system based on structured syntax fields according to claim 13, characterized in that, The audit log module is also used to record at least one of the following: input content, input content hash value, document version, model version, rule version, field generation result, boundary detection result, report version, report generation time, user modification record, and manual review record.

15. The AI ​​knowledge auditing system based on structured syntax fields according to claim 13, characterized in that, The optional variable update module is also used to generate a variable update item to be confirmed when changes are detected in the object definition, evidence status, breakpoint, route partition or output boundary of the user's long-term project, and write it into the user project variable library after user confirmation; the variable update item to be confirmed includes at least one of the following: variable name, original variable value, new variable value, reason for change, associated document, associated audit report, associated breakpoint, associated route partition and user confirmation status.

16. The AI ​​knowledge auditing system based on structured syntax fields according to claim 13, characterized in that, The boundary detection module includes at least two of the following: scale boundary detection unit, time boundary detection unit, information boundary detection unit, evidence boundary detection unit, stable boundary detection unit, solvable boundary detection unit, and closed-loop boundary detection unit.

17. The AI ​​knowledge auditing system based on structured syntax fields according to claim 13, characterized in that, The route partitioning module can be configured to divide relevant models, technical routes, research routes, project plans, or expression content into main routes, strategic routes, cautious routes, or frozen expressions.

18. An AI knowledge auditing device based on structured syntax fields, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the AI ​​knowledge auditing method based on structured syntax fields as described in any one of claims 1 to 12.

19. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the AI ​​knowledge auditing method based on structured syntax fields as described in any one of claims 1 to 12.

20. A computer program product, characterized in that: The computer program product includes a computer program that, when executed by a processor, implements the AI ​​knowledge auditing method based on structured syntax fields as described in any one of claims 1 to 12.

Images