A method and system for verification and revision of artificial intelligence generated text

Abstract

This invention discloses a method and system for verifying and revising AI-generated text. The method involves obtaining structured text generated by a large AI model processing original information text; breaking down claims into several independent claims; performing evidence retrieval and alignment operations on each claim within the original information text to obtain and label matching evidence fragments; comparing the consistency between each claim and its corresponding evidence fragment; classifying claims based on the comparison results; and triggering different targeted revision strategies based on the classification results. Finally, a traceable text containing evidence guidance information is generated, including the original text of the evidence fragment corresponding to each claim, the location information of the evidence fragment in the original information text, the consistency comparison results, and a record of the revision operations. This invention significantly improves the accuracy of text generated by existing large AI models in professional application scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, system, and medium for verifying and revising AI-generated text. Background Technology

[0002] With the rapid development of artificial intelligence technology, text generation systems based on Large Language Models (LLMs) have been widely used in text creation, summary writing, and automatic generation of analytical reports. These systems significantly improve the efficiency of information processing, enabling rapid integration and output of content. However, in practical deployments, especially in highly specialized and accurate applications such as industry research, policy analysis, and technical intelligence assessment, existing systems reveal a key flaw: their generated content often suffers from factual hallucination, element drift, and untraceable conclusions. Because the generation process lacks a rigorous evidence anchoring mechanism, the output statements and judgments often cannot be traced back to reliable original information sources, posing a potential risk of misleading users who rely on such content for decision-making.

[0003] Currently, there are two main technical solutions in the industry. The first approach focuses on enhancing the model's knowledge processing and reasoning capabilities. For example, it involves introducing external knowledge graphs to structure multi-source information or using a multi-model fusion framework for comprehensive analysis and reasoning. This method aims to improve the overall consistency and factual accuracy of the generated content, but it essentially remains an "end-to-end generation or reasoning" paradigm. Its final output is a complete, indivisible text, making it difficult for users or systems to verify each independent factual statement (e.g., "Company X's revenue grew by X% in 2023") or key conclusive judgment in a granular manner. There is a lack of explicit, automated correlation between the generated content and the original evidence. The second approach introduces manual intervention at the back end of the generation process. After the model generates the text, the results are verified and corrected through manual review, acceptance, or rule-based scoring. While this method can guarantee quality to some extent, it is highly dependent on manual intervention and difficult to adapt to the needs of large-scale, automated processing. Meanwhile, the manual verification process itself lacks a structured recording and traceability mechanism, and the standards of different auditors may differ, resulting in unstable and unreusable verification results, making it impossible to form an efficient and closed-loop automatic quality control system.

[0004] However, both approaches suffer from the problems of indivisible statement units, lack of evidentiary connections, and crude revision mechanisms. The indivisibility of statement units manifests in the inability to automatically break down the coherent text generated by a large model into the smallest independently verifiable semantic units (such as a single factual claim or judgment). The lack of evidentiary connections manifests in the absence of an automated mechanism to establish a precise and traceable correspondence between each verifiable unit in the generated text and the underlying original text (evidence source), resulting in a broken "conclusion-evidence" chain. The crude revision mechanism manifests in the lack of a refined and controllable automatic revision strategy when verification reveals factual or logical problems in certain content. Current practices typically force a complete rewrite or rely entirely on manual modification; the former may fail to specifically correct errors and easily introduce new biases, while the latter is inefficient and costly.

[0005] Based on the above problems, the applicant proposes the technical solution of this application. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the prior art, the present invention provides a method, system and medium for verifying and revising artificial intelligence-generated text, which can realize the structured decomposition of generated content, automated evidence association and traceability, and precise and controllable revision based on verification feedback, thereby greatly improving the reliability, auditability and credibility of its output while ensuring the efficiency advantage of large model generation.

[0007] To achieve the above objectives, this invention discloses a method for verifying and revising artificial intelligence-generated text, comprising the following steps: Obtain structured text generated by the identification and processing of the original information text by a large artificial intelligence model. The structured text includes at least basic field information, industry element information, and strategic risk information. The structured text is decomposed into several independent claims. For each claim, evidence retrieval and alignment operations are performed in the original information text to obtain evidence fragments that match each claim and mark the evidence fragments. The position information of the evidence fragments in the original information text is recorded. Each claim is compared for consistency with its corresponding evidence fragments. The claims are then categorized based on the comparison results. Different targeted revision strategies are triggered based on the categorization results to revise the claims. Based on the revised results, a traceable text containing evidence guidance information is generated. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the location information of the evidence fragment in the original information text, the consistency comparison results, and the revision operation record.

[0008] Preferably, the step of performing evidence retrieval and alignment operations in the original information text specifically includes: Using semantic similarity calculation or vector retrieval methods, candidate fragments that are semantically related to each element in the claim are filtered from the original information text; The candidate segments are ranked by relevance, and the segment with the highest similarity is selected as the matching evidence segment. If no fragments meeting the preset similarity threshold are found, the evidence for the claim is marked as empty.

[0009] Preferably, the claims include at least factual claims, judgmental claims, and citation claims; The specific steps for comparing the consistency of each claim with its corresponding evidence are as follows: The process employs natural language reasoning models or rule engines to compare the consistency of objective facts for factual claims, which include at least numerical values, time, and entities. For judgmental claims, it compares the consistency of viewpoints or logical conclusions. For citation-based claims, it compares the consistency of citation content and sources.

[0010] Preferably, the claim classification includes at least supporting claims, contradictory claims, and unknown claims; In the targeted revision strategy, no revision operation is performed for supporting claims; for contradictory claims, a replacement or revision operation based on evidence fragments is adopted to make the revised content consistent with the evidence; for unknown claims, the claim is deleted, uncertain modifiers are added, or it is marked as requiring manual confirmation.

[0011] Preferably, in the targeted revision strategy, for contradictory claims, the prompts are reconstructed based on the annotation of evidence fragments, and a large artificial intelligence model is invoked to regenerate structured text. Then, the subsequent claim decomposition, verification and revision are performed based on the structured text.

[0012] Preferably, acquiring structured text generated by a large artificial intelligence model includes: Obtain multiple structured texts generated by multiple different large artificial intelligence models for the same input; Cross-validation is performed on the multiple structured texts to select claims that appear consistently in multiple models as claims to be verified, and claims that appear only in a single model are marked as low-confidence claims. These claims to be verified are given priority for subsequent verification and revision.

[0013] Preferably, before performing the evidence retrieval and alignment operations, the method further includes: Semantic-level term normalization is performed on the original information text, mapping terms with the same meaning but different expressions in the original information text to a unified standard expression, and constructing a term normalization mapping table; The evidence retrieval and alignment operations are performed based on the normalized original information text, and when outputting evidence fragments, the evidence fragments are restored to their original expressions according to the mapping table.

[0014] Preferably, the semantic-level term normalization process includes at least: Identify proper nouns, abbreviations, synonyms, and polysemous words in the original information text; Term alignment is performed using a domain knowledge base or a pre-trained language model to generate normalized standard text. Record the correspondence between each normalized term and the original term for use in subsequent reverse mapping.

[0015] This invention also discloses a verification and revision system for artificial intelligence-generated text, comprising the following modules: The model input module is used to obtain structured text generated by the artificial intelligence big data model after the original information text has been identified and processed. The structured text includes at least basic field information, industry element information and strategic risk information. The text processing module is used to decompose the structured text into several independent claims. For each claim, it performs evidence retrieval and alignment operations in the original information text to obtain evidence fragments that match each claim and to mark the evidence fragments. It also records the position information of the evidence fragments in the original information text. The verification and revision module is used to compare the consistency of each claim with its corresponding evidence fragment, classify the claim according to the comparison results, and trigger different targeted revision strategies to revise the claim according to the classification results; The result output module is used to generate traceable text containing evidence guidance information based on the revised result. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the position information of the evidence fragment in the original information text, the consistency comparison result, and the revision operation record.

[0016] The present invention also discloses a computer-readable storage medium storing a computer program that, when processed and executed, implements the method described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention significantly reduces the risk of factual illusion by breaking down the text generated by a large AI model into verifiable claim units. Through a structure-preserving targeted revision mechanism, only problematic fields are revised, avoiding the introduction of new errors through a complete rewrite. This forms an automated closed-loop processing flow of "generation-verification-revision-reverification," significantly reducing reliance on manual review. It is particularly suitable for professional application scenarios with high requirements for text reliability, such as industry research, intelligence analysis, and compliance review.

[0018] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating one aspect of the AI-generated text verification and revision method of the present invention.

[0020] Figure 2 This is a schematic diagram of the loop judgment of the verification and revision method of the artificial intelligence-generated text of the present invention. Detailed Implementation

[0021] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further illustrated below with reference to specific figures. However, the invention is not limited to the embodiments described below.

[0022] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0023] To address the various problems existing in professional applications of text generated by current large-scale AI models, such as the presence of factual illusions or inconsistencies with the original information, the lack of automated verification methods, the inability to verify individual facts or judgments when using generated content as a whole, and the potential for introducing new biases through complete rewriting or manual revision, this application proposes an automatic verification and targeted revision method for AI-generated text based on an "claim-evidence" framework. Instead of treating the large model output as an indivisible whole, this method breaks it down into multiple independently verifiable claims. It automatically retrieves corresponding evidence from the original text, classifies each claim based on the consistency between the claim and the evidence, and triggers targeted revisions only for the problematic fields. This transforms the "uncontrollable generation" of the large model into "verifiable, repairable, and traceable generation."

[0024] The claim-evidence-based automated verification and targeted revision method proposed in this application is deployed within an automated processing workflow for information indexing and research analysis. This workflow includes generation nodes, verification nodes, rollback revision nodes, and loop control nodes, forming a closed-loop processing structure of "generation-verification-revision-regeneration". The generation node generates structured information text or analysis result text using a large model. The verification node performs claim decomposition, evidence retrieval, alignment, and consistency classification. The rollback revision node triggers targeted regeneration only for the problematic parts based on the classification results. The loop control node controls the repeated execution of the above processes until all claims are deemed supported or a preset convergence condition is met, at which point the final result is output. The workflow in this application is not a one-time process but is repeatedly invoked as the credibility control center for the generated results.

[0025] like Figure 1 As shown, the method for verifying and revising AI-generated text disclosed in this application includes the following steps: Step S1, Model Input.

[0026] The model input step is executed in the workflow generation node to obtain structured text generated by the artificial intelligence big data model after the original information text has been identified and processed. The structured text includes at least basic field information, industry element information and strategic risk information.

[0027] Specifically, the workflow begins with the generation node calling a large artificial intelligence model to generate structured text based on the input raw information text. The output of the structured text includes multiple fields or entries.

[0028] The AI-powered large-scale model automatically identifies and processes the raw information text, extracting key information, primarily including basic field information, industry element information, and strategic risk information. Basic field information refers to the data extracted from the information text, automatically generating fields such as title, source name, source type, publication time, summary, and low-value / low-profile content. Industry element information refers to the data automatically generated by the large-scale model based on the information text, including the sub-industry (aircraft manufacturing, avionics, engines, airline operations, etc.), involved companies (Boeing, GE, Airbus, etc.), involved product models (LEAP-1C, A320, 737MAX, etc.), involved countries / regions (United States, France, China, etc.), and event type (order signing, cooperation, test flight, production launch, etc.). Strategic risk information refers to the data obtained by the large-scale model from the information text, determining the strategic signals indicated by the text and the risks they pose, including strategic signals, strategic inclinations, scope of impact, strategic relevance, and whether risks are involved.

[0029] In one example, when using large models to obtain structured text, multiple structured texts generated by different large AI models for the same input are acquired. These multiple structured texts are then cross-validated. Claims that consistently appear in multiple models are selected as claims to be verified, while claims appearing only in a single model are marked as low-confidence claims and prioritized for further verification and revision. For instance, after receiving the user's raw input, the same generation request is simultaneously sent to at least two pre-configured large language models. Each model interface can operate independently without interference through an asynchronous request mechanism, generating its corresponding structured text. The model identifier corresponding to each output result is recorded for traceability.

[0030] Step S2, text processing.

[0031] The text processing step is executed in the verification node of the workflow. The structured text is decomposed into several independent claims. For each claim, evidence retrieval and alignment operations are performed in the original information text to obtain evidence fragments that match each claim and to mark the evidence fragments. The position information of the evidence fragments in the original information text is recorded.

[0032] Specifically, the structured text generated by the large model is decomposed into claims, forming multiple claims. For each claim, evidence is searched in the original information text input to the large model, and the evidence is labeled one by one.

[0033] When performing evidence retrieval and alignment operations in the original information text, the candidate segments that are semantically related to each element in the claim are filtered in the original information text using semantic similarity calculation or vector retrieval methods; the candidate segments are sorted by relevance, and the segment with the highest similarity is selected as the matching evidence segment; if no segment that meets the preset similarity threshold is found, the evidence for the claim is marked as empty.

[0034] In one example, before evidence retrieval and alignment operations, the original information text is preprocessed to improve subsequent retrieval efficiency and accuracy. For instance, semantic-level terminology normalization is performed on the original information text, mapping terms with the same meaning but different expressions to a unified standard representation, and constructing a terminology normalization mapping table. Subsequent evidence retrieval and alignment operations are performed based on the normalized original information text, and when outputting evidence fragments, the evidence fragments are restored to their original representations according to the mapping table. Semantic-level terminology normalization specifically involves identifying proper nouns, abbreviations, synonyms, and polysemous words in the original information text; performing terminology alignment using a domain knowledge base or pre-trained language model to generate normalized standard text; and recording the correspondence between each normalized term and the original term for subsequent reverse mapping. During terminology recognition, a named entity recognition strategy based on a pre-trained language model combined with a domain dictionary matching strategy is used to automatically identify and extract candidate terms requiring normalization from the original information text. For each identified candidate term, semantic alignment is performed using a pre-constructed domain knowledge graph or a pre-trained YuYan model, mapping to form a unified standard representation. During the normalization process, a bidirectional mapping table is simultaneously constructed and stored, recording the correspondence between each original term and its standard representation. The mapping table includes at least the following fields: the original text of the term, its paragraph and offset, and the normalized standard representation. This mapping table is stored in key-value pair format, supporting both forward and reverse queries. All identified terms in the original information text are replaced with the standard representation according to the mapping relationship, generating a "normalized version" of the original text. All subsequent evidence retrieval and alignment operations are performed on this normalized text to ensure semantic equivalence between terms in different representations.

[0035] Before claim decomposition, the structured text can be cleaned and sentence boundary detected. Text cleaning removes irrelevant formatting tags, redundant statements, and repetitive expressions, retaining the core claims. Sentence boundary detection segments the cleaned text into semantically complete sentences. Then, for each preprocessed sentence or clause, it is determined whether it contains multiple independently verifiable atomic claims; if so, atomic decomposition is performed.

[0036] After the claims are broken down, evidence retrieval and alignment are performed on each claim within the original information text. The location of the evidence is recorded in the original information text to ensure traceability. The retrieval uses a core semantic query approach, combining the structured fields of the claims to generate standardized query statements for execution.

[0037] In one example, claims are first categorized into claim types: factual claims, judgmental claims, and citation claims. Then, the query strategy is dynamically adjusted based on the claim type. For factual claims, the weight of numerical values ​​and entities is strengthened; for judgmental claims, the weight of relationships and evaluations is strengthened; and for citation claims, the weight of source metadata is strengthened.

[0038] Step S3, Verification and Revision.

[0039] The verification and revision steps are executed in the rollback revision node, and are controlled by the loop control node during specific revisions. Each claim is compared for consistency with its corresponding evidence fragment. Based on the comparison results, the claims are categorized, and different targeted revision strategies are triggered to revise the claims according to the categorization results.

[0040] Specifically, each claim is compared for consistency with its corresponding evidence. This is done using a natural language reasoning model or rule engine. For factual claims, the comparison is made to determine whether the objective facts are consistent, including at least numerical values, time, and entities. For judgmental claims, the comparison is made to determine whether the viewpoints or logical conclusions are consistent. For citation-based claims, the comparison is made to determine whether the content and source of the citations are consistent.

[0041] By comparing the consistency between claims and evidence and annotating each piece of evidence, the attributes of each piece of evidence can be determined, and the corresponding claims can be classified. Evidence attributes include three types: supported, contradictory to the original text, and not found in the original text. Claims corresponding to supported claims are identified as supporting claims, claims corresponding to contradictory claims are identified as contradictory claims, and claims not found in the original text are identified as unknown claims. By determining which category the evidence attribute belongs to, the authenticity and accuracy of the content in the generated structured text can be judged. For example, in the generated summary, there is the statement "Delta has received 47 Airbus aircraft produced in Canada, Germany, and France, making it Airbus's second-largest delivery customer in 2024." After verification, evidence for this claim can be found in the original text: "47 aircraft were received in the past two years, and it is the second-largest customer in 2024." Therefore, this claim is determined to be a supported claim.

[0042] In the targeted revision strategy, no revision is performed on supporting claims; for contradictory claims, replacement or revision based on evidence fragments is used to ensure the revised content is consistent with the evidence; and for unknown claims, the claim is deleted, uncertain modifiers are added, or it is marked as requiring manual confirmation. For example... Figure 2As shown, the revision does not involve a complete regeneration of the entire output of the large model. Instead, for claims deemed supported, the corresponding fields are directly used for subsequent output and do not participate in subsequent loops. For claims deemed contradictory or undiscovered, their corresponding field paths are marked as requiring revision, and the process reverts to the corresponding sub-generation node. During regeneration, the reverted generation node only reconstructs the hints for the marked fields and calls the large model, leaving the remaining fields unchanged. This achieves targeted revision while preserving the structure, avoiding the introduction of new uncertainties through a complete rewrite. For example, "warning tariffs may force" is changed to "warning tariff policies may force it" to be closer to the original wording; "Delta hopes its 50 A321neo aircraft will be delivered from Mobile's third final assembly line" is moved to the low-heat, high-value information section to maintain the summary.

[0043] In one example, for contradictory claims, the prompts are reconstructed based on the annotation of evidence fragments, and a large artificial intelligence model is invoked to regenerate structured text. Then, based on the structured text, subsequent claim decomposition, verification, and revision are performed.

[0044] Step S4, output the result.

[0045] Based on the revised results, a traceable text containing evidence guidance information is generated. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the location information of the evidence fragment in the original information text, the consistency comparison results, and the revision operation record.

[0046] This embodiment forms a closed-loop processing structure through a workflow. By breaking down the text generated by the large AI model into verifiable claim units, the risk of factual illusion is significantly reduced. Through a structure-preserving targeted revision mechanism, only problematic fields are revised, avoiding the introduction of new errors through a complete rewrite. This forms an automated closed-loop processing flow of "generation-verification-revision-reverification," significantly reducing reliance on manual review. It is particularly suitable for professional application scenarios with high requirements for text reliability, such as industry research, intelligence analysis, and compliance review.

[0047] The second embodiment of the present invention discloses a verification and revision system for artificial intelligence-generated text, including a model input module, a text processing module, a verification and revision module, and a result output module.

[0048] The model input module is used to obtain structured text generated by the artificial intelligence big data model after the original information text has been identified and processed. The structured text includes at least basic field information, industry element information and strategic risk information.

[0049] The text processing module is used to decompose the structured text into several independent claims. For each claim, it performs evidence retrieval and alignment operations in the original information text to obtain evidence fragments that match each claim and to mark the evidence fragments. It also records the position information of the evidence fragments in the original information text.

[0050] The verification and revision module is used to compare the consistency of each claim with its corresponding evidence fragment, classify the claim according to the comparison results, and trigger different targeted revision strategies to revise the claim according to the classification results.

[0051] The result output module is used to generate traceable text containing evidence guidance information based on the revised result. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the position information of the evidence fragment in the original information text, the consistency comparison result, and the revision operation record.

[0052] Since the first embodiment corresponds to this embodiment, this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment remain valid in this embodiment, and the technical effects achievable in the first embodiment can also be achieved in this embodiment. To reduce repetition, they will not be repeated here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0053] The third embodiment of the present invention relates to a computer-readable storage medium having a computer program / instructions stored thereon, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the method in the first embodiment.

[0054] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for verifying and revising artificial intelligence-generated text, characterized in that, Includes the following steps: Obtain structured text generated by the identification and processing of the original information text by a large artificial intelligence model. The structured text includes at least basic field information, industry element information, and strategic risk information. The structured text is decomposed into several independent claims. For each claim, evidence retrieval and alignment operations are performed in the original information text to obtain evidence fragments that match each claim and mark the evidence fragments. The position information of the evidence fragments in the original information text is recorded. Each claim is compared for consistency with its corresponding evidence fragments. The claims are then categorized based on the comparison results. Different targeted revision strategies are triggered based on the categorization results to revise the claims. Based on the revised results, a traceable text containing evidence guidance information is generated. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the location information of the evidence fragment in the original information text, the consistency comparison results, and the revision operation record.

2. The method for verifying and revising AI-generated text according to claim 1, characterized in that, The evidence retrieval and alignment operation performed in the original information text specifically includes: Using semantic similarity calculation or vector retrieval methods, candidate fragments that are semantically related to each element in the claim are filtered from the original information text; The candidate segments are ranked by relevance, and the segment with the highest similarity is selected as the matching evidence segment. If no fragments meeting the preset similarity threshold are found, the evidence for the claim is marked as empty.

3. The method for verifying and revising AI-generated text according to claim 1, characterized in that, The claims include at least factual claims, judgmental claims, and citation claims; The specific steps for comparing the consistency of each claim with its corresponding evidence are as follows: The process employs natural language reasoning models or rule engines to compare the consistency of objective facts for factual claims, which include at least numerical values, time, and entities. For judgmental claims, it compares the consistency of viewpoints or logical conclusions. For citation-based claims, it compares the consistency of citation content and sources.

4. The method for verifying and revising AI-generated text according to claim 1, characterized in that, The claim classification includes at least supporting claims, contradictory claims, and unknown claims; In the targeted revision strategy, no revision is performed on supporting claims; for contradictory claims, a replacement or revision operation based on evidence fragments is adopted to make the revised content consistent with the evidence; and for unknown claims, the claim is deleted, uncertain modifiers are added, or it is marked as requiring manual confirmation.

5. The method for verifying and revising AI-generated text according to claim 4, characterized in that, In the targeted revision strategy, for contradictory claims, prompts are reconstructed based on the annotation of evidence fragments, and a large artificial intelligence model is invoked to regenerate structured text. Then, based on the structured text, subsequent claim decomposition, verification and revision are performed.

6. The method for verifying and revising AI-generated text according to claim 1, characterized in that, When acquiring structured text generated by large artificial intelligence models, the following are included: Obtain multiple structured texts generated by multiple different large artificial intelligence models for the same input; Cross-validation is performed on the multiple structured texts to select claims that appear consistently in multiple models as claims to be verified, and claims that appear only in a single model are marked as low-confidence claims. These claims to be verified are given priority for subsequent verification and revision.

7. The method for verifying and revising AI-generated text according to claim 1, characterized in that, Before performing the evidence retrieval and alignment operations, the method further includes: Semantic-level term normalization is performed on the original information text, mapping terms with the same meaning but different expressions in the original information text to a unified standard expression, and constructing a term normalization mapping table; The evidence retrieval and alignment operations are performed based on the normalized original information text, and when outputting evidence fragments, the evidence fragments are restored to their original expressions according to the mapping table.

8. The method for verifying and revising AI-generated text according to claim 7, characterized in that, The semantic-level term normalization process includes at least the following: Identify proper nouns, abbreviations, synonyms, and polysemous words in the original information text; Term alignment is performed using a domain knowledge base or a pre-trained language model to generate normalized standard text. Record the correspondence between each normalized term and the original term for use in subsequent reverse mapping.

9. A verification and revision system for artificial intelligence-generated text, characterized in that, Includes the following modules: The model input module is used to obtain structured text generated by the artificial intelligence big data model after the original information text has been identified and processed. The structured text includes at least basic field information, industry element information and strategic risk information. The text processing module is used to decompose the structured text into several independent claims. For each claim, it performs evidence retrieval and alignment operations in the original information text to obtain evidence fragments that match each claim and to mark the evidence fragments. It also records the position information of the evidence fragments in the original information text. The verification and revision module is used to compare the consistency of each claim with its corresponding evidence fragment, classify the claim according to the comparison results, and trigger different targeted revision strategies to revise the claim according to the classification results; The result output module is used to generate traceable text containing evidence guidance information based on the revised result. The evidence guidance information includes the original text of the evidence fragment corresponding to each claim, the position information of the evidence fragment in the original information text, the consistency comparison result, and the revision operation record.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when processed and executed, implements the method as described in any one of claims 1 to 8.

Images