Large model factual error correction system and method based on multi-source cross validation

Abstract

The invention discloses a large model factual error correction system based on multi-source cross validation. The large model factual error correction system is characterized by comprising a large model generation module, a multi-dimensional fact verification module, a dynamic fact knowledge base updating module, a user feedback closed-loop error correction module and a data interaction bus, the data interaction bus is in communication connection with the large model generation module, the multi-dimensional fact verification module, the dynamic fact knowledge base updating module and the user feedback closed-loop error correction module, and real-time transmission and cooperative work of fact point data, verification results, updated logs and user feedback information among the modules are achieved. The beneficial effects of the invention are that the technical scheme significantly improves the verification accuracy of professional scenes through the cross validation of three layers of data sources and the domain weight algorithm; and the problem of error propagation is solved by event-driven updating and historical backtracking.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a large-scale model factual error correction system and method based on multi-source cross-validation. Background Technology

[0002] Generative artificial intelligence has been widely applied in content creation, professional consulting, and other fields, but frequent factual errors have become a core bottleneck restricting its extension to high-reliability scenarios. Existing technologies suffer from the following key problems: Weak fact verification mechanisms: Most systems rely on a single general data source for verification, lacking domain adaptability, and insufficient cross-validation leads to a high rate of false negatives; The knowledge base is static and lagging: it cannot synchronize with the latest facts released by authoritative institutions in real time, and outdated errors in historically generated content are difficult to correct. Lack of feedback loop: Factual errors reported by users do not form an effective conversion path, making it impossible to quickly optimize the model and causing similar errors to recur. Poor module collaboration: The generation, verification, and update modules operate independently, resulting in low data interaction efficiency and making it difficult to achieve real-time error correction across the entire chain.

[0003] The aforementioned pain points are particularly prominent in professional scenarios such as medical diagnosis and legal compliance, where erroneous facts can easily lead to security risks or misleading decision-making. To address this, this invention constructs an integrated fact-correction system based on multi-source cross-validation technology, an event-driven dynamic knowledge base update mechanism, and a reinforcement learning feedback loop. This system aims to solve the problems of fact inaccuracy, delayed updates, and slow optimization in existing technologies, thereby improving the factual reliability and scenario adaptability of large-scale models. Summary of the Invention

[0004] The main technical problem solved by this invention is to provide a large-scale model factual error correction system and method based on multi-source cross-validation, thereby solving one or more of the above-mentioned prior art problems.

[0005] To address the aforementioned technical problems, the present invention employs the following technical solution: a large-model factual error correction system based on multi-source cross-validation. Its innovation lies in comprising a large-model generation module, a multi-dimensional fact verification module, a dynamic fact knowledge base update module, a user feedback closed-loop error correction module, and a data interaction bus. The data interaction bus is communicatively connected to the large-model generation module, the multi-dimensional fact verification module, the dynamic fact knowledge base update module, and the user feedback closed-loop error correction module, respectively, enabling real-time transmission and collaborative operation of fact point data, verification results, update logs, and user feedback information among the modules. The large-model generation module receives user input generation requests, outputs initial natural language content, and extracts key fact points from it. The key fact points are sent to the multi-dimensional fact verification module, which simultaneously stores the initial natural language content, key fact points, and the final optimized result. The multi-dimensional fact verification module performs multi-source cross-validation on the key fact points, determines the accuracy of the facts, generates fact correction suggestions, and feeds back the fact correction suggestions to the large model generation module. The dynamic fact knowledge base update module updates the fact knowledge base in real time and retrospectively corrects erroneous entries involving updated facts in the historical generated content stored by the large model generation module. The user feedback closed-loop error correction module receives fact error information from users, verifies the authenticity of the feedback, and converts valid feedback into training samples to optimize the large model generation module.

[0006] In some implementations, the large model generation module includes a fact point extraction unit and a generated content caching unit. The fact point extraction unit uses named entity recognition and relation extraction algorithms to automatically identify time, data, concepts, and standard factual information in the initial natural language content to form key fact points. The generated content caching unit is a non-volatile storage unit used to store the initial natural language content, the key fact points output by the fact point extraction unit, and the final result optimized by the large model generation module based on fact correction suggestions. The generated content caching unit is associated with the dynamic fact knowledge base update module and supports the retrieval and backtracking of historical generated content.

[0007] In some implementations, the multi-dimensional fact verification module includes a three-layer data source pool, a domain adaptation unit, and a cross-validation unit. The three-layer data source pool includes a general authoritative database, a domain-specific knowledge base, and a real-time trusted data source. The general authoritative database includes the Britannica database and a national standards database. The domain-specific knowledge base includes a clinical practice guidelines database in the medical field, a legal database, and a historical chronology database. The real-time trusted data source includes government website data interfaces, authoritative media real-time information interfaces, and a database of the latest academic journal articles. The domain adaptation unit is used to obtain the domain tags corresponding to the user-generated requirements and, based on the domain... The tags are matched with corresponding dedicated data sources from the three-layer data source pool; the cross-validation unit is used to perform multi-source comparison between the key fact points and the data in the dedicated data sources matched by the domain adaptation unit, with no less than 2 multi-source comparisons and at least 1 comparison of domain professional knowledge base data; if the multi-source data are consistent, the fact is determined to be accurate; if the multi-source data have deviations, the fact is determined to be incorrect based on the domain weights, where the domain weights are set to 0.5 for the domain professional knowledge base, 0.3 for the real-time trusted data source, and 0.2 for the general authoritative database; when the cross-validation unit determines that the fact is incorrect, it generates fact correction suggestions containing the correct fact content, the data source reference address, and the reason for the error.

[0008] In some implementations, the dynamic fact knowledge base update module includes an event-driven real-time update unit, a historical answer backtracking and correction unit, and an update log unit. The event-driven real-time update unit is configured with an authoritative institution domain name whitelist. It uses web crawling technology to capture real-time fact information within the authoritative institution domain name whitelist and sends the real-time fact information to the multi-dimensional fact verification module for secondary verification. After successful verification, it updates the information to the three-layer data source pool. The historical answer backtracking and correction unit is communicatively connected to the generated content caching unit. When the event-driven real-time update unit completes a fact update, it automatically retrieves historical generated entries related to the updated fact from the generated content caching unit. If a historical generated entry with factual errors is found, a fact correction prompt is generated and pushed to the associated users. The update log unit is used to record the update time, updated fact content, data source address, backtracking and correction historical generated entry ID, and push user list for each fact update.

[0009] In some implementations, the user feedback closed-loop error correction module includes a feedback receiving unit, a feedback verification unit, a reinforcement learning optimization unit, and a feedback incentive unit. The feedback receiving unit provides a visual interactive interface, allowing users to select erroneous factual content through highlighting and simultaneously submit correct factual information and corresponding domain tags. The feedback verification unit invokes the cross-validation function of the multi-dimensional fact verification module to perform multi-source cross-validation on the correct factual information submitted by the user, determining whether the feedback is valid or invalid. The reinforcement learning optimization unit uses the PPO reinforcement learning algorithm to transform the correct factual information in the valid feedback into training samples, optimizing the large model generation module. The feedback incentive unit establishes a points reward mechanism, issuing preset points to users who submit valid feedback; these preset points can be redeemed for service time or access to advanced functions of the large model generation module.

[0010] In some implementations, the data interaction bus supports both synchronous and asynchronous communication modes. Synchronous interaction mode is used when the large model generation module sends key fact points to the multi-dimensional fact verification module, and the multi-dimensional fact verification module provides fact correction suggestions to the large model generation module. Asynchronous interaction mode is used when the dynamic fact knowledge base update module pushes fact correction prompts to the user, and the user feedback closed-loop error correction module transmits domain training samples to the large model generation module.

[0011] In some implementations, the fact point extraction unit uses a BERT pre-trained model combined with named entity recognition algorithm and relation extraction algorithm; the generated content caching unit uses a distributed non-volatile storage cluster with a response time of no more than 100ms.

[0012] In some implementations, the domain adaptation unit uses a dual mechanism of rule matching and semantic understanding to obtain domain labels; when the user's requirements contain explicit domain keywords, the corresponding labels are directly matched; otherwise, the domain labels are automatically labeled by analyzing the semantics of the requirements through the TextCNN text classification model.

[0013] In some implementations, the event-driven real-time update unit uses Selenium distributed crawler technology, and the fact information update delay does not exceed 1 hour; the historical answer backtracking and correction unit searches for historically generated entries within 12 months prior to the release of the real-time fact information.

[0014] In some implementations, after the reinforcement learning optimization unit accumulates 100 valid feedback training samples, it fine-tunes the large model generation module. During fine-tuning, 60% of the parameters at the bottom layer of the model are frozen, and only the top attention layer and output layer are trained. The preset points redemption rule of the feedback incentive unit is that 100 points can be redeemed for 1 hour of advanced question-and-answer service or corresponding advanced function permissions.

[0015] A method for a large-scale model factual error correction system based on multi-source cross-validation is characterized by: a real-time generation and verification process, a dynamic knowledge base update process, and a user feedback optimization process; the real-time generation and verification process includes: user input to generate requirements, initial content generation and fact point extraction, multi-source cross-validation, and content optimization output; the dynamic knowledge base update process includes: real-time fact capture, secondary verification, data source update, historical entry backtracking, and correction prompt push; the user feedback optimization process includes: feedback submission, feedback verification, effective sample conversion, reinforcement learning model fine-tuning, and points incentive distribution.

[0016] In some implementations, during the real-time generation verification process, the cross-validation unit performs at least two multi-source comparisons, including at least one comparison of domain knowledge base data; the domain weight is set to be greater than the weight of the real-time trusted data source, which is greater than the weight of the general authoritative database.

[0017] In some implementations, in the knowledge base dynamic update process, the event-driven real-time update unit is set to crawl once every 30 minutes, and triggers immediate crawling when sudden important factual information is detected; the retrieval time range of the historical answer backtracking and correction unit is set to historical generated entries within 12 months before the release of real-time factual information.

[0018] In some implementations, during the user feedback optimization process, the reinforcement learning optimization unit fine-tunes the large model generation module after accumulating 100 valid feedback training samples. During fine-tuning, 60% of the parameters at the bottom layer of the model are frozen, and only the top attention layer and output layer are trained. The preset points redemption rule of the feedback incentive unit is that 100 points can be redeemed for 1 hour of advanced Q&A service or corresponding advanced function permissions.

[0019] The beneficial effects of this invention are: technical breakthrough: cross-validation of three data sources plus domain weighting algorithm significantly improves the accuracy of validation in professional scenarios; event-driven update (delay not exceeding 1 hour) plus historical backtracking (12-month range) solves the error propagation problem; Scenario adaptation: Covering high-requirement fields such as medical and legal, we offer customized verification solutions to ensure compliance; User closed loop: Feedback is verified and used for model optimization, and points incentives increase participation, forming a virtuous cycle of users helping the model and the model feeding back to users; Industry empowerment: Full-process logs support compliance auditing, reduce medical follow-up visits and financial misrepresentation risks, and promote the extension of large models to professional production scenarios. Attached Figure Description

[0020] 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 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, wherein: Figure 1 This is an overall architecture diagram of a large-scale model factual error correction system based on multi-source cross-validation according to the present invention.

[0021] Figure 2 This is a flowchart of the real-time generation and verification process of a large-scale model factual error correction system based on multi-source cross-validation, as described in this invention.

[0022] Figure 3 This is a user feedback closed-loop flowchart of a large-scale model factual error correction system based on multi-source cross-validation according to the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, the embodiments of the present invention include: This embodiment is based on a large model factual error correction system based on multi-source cross-validation, and describes in detail the implementation details, data interaction process and key technical parameters of each module, aiming to clearly demonstrate the implementation logic and operation mechanism of the system.

[0025] 1. Large model generation module Fact Point Extraction Unit: Using a BERT pre-trained model combined with Named Entity Recognition (NER) and Relation Extraction (RE) algorithms, it automatically identifies factual information in the initial content, such as time (e.g., May 2024), data (e.g., blood pressure 130 / 80 mmHg), concepts (e.g., stage 1 hypertension), and standards (e.g., Article 1043 of the Civil Code), and outputs structured key fact points (JSON format: {Entity 1: stage 1 hypertension; Relation: diagnostic criteria; Entity 2: systolic blood pressure 130-139 mmHg / diastolic blood pressure 80-89 mmHg; Domain tag: medicine-cardiovascular disease}).

[0026] Content generation caching unit: It adopts a distributed non-volatile storage cluster (such as MongoDB) to store the initial content, key fact points, final results and unique generation ID; it supports retrieval by domain tags, generation time and fact keywords, and the response time is no more than 100ms.

[0027] 2. Multi-dimensional fact verification module (1) Three-tier data source pool: 1) Universal authoritative databases: Integrating Britannica encyclopedia, China National Standards System, and ISO database; 2) Domain-specific knowledge bases: The medical field includes the "Chinese Clinical Practice Guidelines" and the "PMC Academic Paper Database"; the legal field includes the Chinese laws and regulations database and the Supreme People's Court case database; and the historical field includes the "General History of China" chronology database. 3) Real-time trusted data source: API interface integration.

[0028] (2) Domain Adaptation Unit: A dual mechanism of rule matching and TextCNN text classification is adopted: if the user's requirements contain explicit keywords (such as: Civil Code), the tags are directly matched; if there are no keywords, the semantics are analyzed by TextCNN and the tags "medical insurance - employee reimbursement" are automatically labeled.

[0029] (3) Cross-validation unit: Select no less than two data sources for comparison for the same fact point, and vote according to the weight of domain knowledge base (0.5), real-time trusted data source (0.3), and general authoritative database (0.2); when there is an error, generate a correction suggestion containing the erroneous content, the correct fact, the data source URL, and the reason for the error.

[0030] 3. Dynamic Fact Knowledge Base Update Module Event-driven real-time update unit: Uses Selenium distributed crawler to crawl real-time information of authoritative domain whitelist (such as: gov.cn, nature.com), and updates it to the data source pool after secondary verification, with an update delay of no more than 1 hour.

[0031] Historical Answer Retrospective Correction Unit: Retrieves historical entries from the 12 months prior to the publication of real-time facts, generates prompts containing incorrect content, correct facts, and reasons for updates, and pushes them to associated users.

[0032] Update log unit: Records update time, content, data source URL, backtracking entry ID, and list of push users to form an auditable log.

[0033] 4. User feedback closed-loop error correction module Feedback receiving unit: Provides a visual interface for web / app applications, allowing users to highlight and mark erroneous content and submit correct facts and domain tags.

[0034] Feedback Verification Unit: Invokes cross-validation function to verify the facts submitted by the user; valid feedback must be consistent with authoritative data from multiple sources, while invalid feedback will provide the reason and correct reference.

[0035] Reinforcement learning optimization unit: The PPO algorithm is used to convert effective feedback into training samples (format: {input: demand plus false facts; output: correct facts; domain label: XX; reward value: 1.0}). After accumulating 100 samples, the model is fine-tuned, 60% of the bottom layer parameters are frozen, and only the top attention layer and output layer are trained.

[0036] Feedback Incentive Unit: Valid feedback earns 100 points, which can be redeemed for 1 hour of premium service or professional function access.

[0037] 5. Data Interaction Bus Synchronous interaction: used for large models; fact point transmission in the verification module; feedback of correction suggestions for large models; Asynchronous interaction: used for dynamically updating module push correction prompts and user feedback module transmit training samples.

[0038] Example: Taking the medical diagnosis of hypertension as an example User input: Please explain the diagnostic criteria for stage 1 hypertension in the 2024 edition; The large model generates initial content: systolic blood pressure 140-159 mmHg, diastolic blood pressure 90-99 mmHg, extracts fact points and caches them; Domain-adaptive units are labeled for medical and cardiovascular diseases, and matched with dedicated data sources (Clinical Practice Guidelines plus PMC papers). Cross-validation revealed an error (the correct standard is 130-139 / 80-89 mmHg), and correction suggestions were generated. The large model is optimized to output the final result and stored. If the guidelines are updated in the future, the dynamic module will capture and retrospectively correct historical entries and push notifications to users. When users report errors, the data is verified and converted into training samples to fine-tune the model and reward points.

[0039] The working principle of this technical solution is as follows: the system achieves end-to-end fact-checking through the linkage of three core processes: 1) Real-time generation of verification process: user requirements, initial generation, fact point extraction, multi-source cross-validation, and output correction; 2) Dynamic knowledge base update process: real-time fact capture, secondary verification, data source update, historical entry review, and push correction prompts; 3) User feedback optimization process: feedback submission, verification, conversion of valid samples, reinforcement learning fine-tuning of the model, and points incentive.

[0040] The three processes work together through a data interaction bus to form a closed loop of generation, verification, updating and optimization, continuously improving the factual accuracy of the large model.

[0041] The advantages of this technical solution are: Technological breakthroughs: Three-layer data source cross-validation plus domain weighting algorithm significantly improves the accuracy of validation in professional scenarios; event-driven updates (delay of no more than 1 hour) plus historical backtracking (within 12 months) solve the problem of error propagation; Scenario adaptation: Covering high-requirement fields such as medical and legal, we offer customized verification solutions to ensure compliance; User closed loop: Feedback is verified and used for model optimization, and points incentives increase participation, forming a virtuous cycle of users helping the model and the model feeding back to users; Industry empowerment: Full-process logs support compliance auditing, reduce medical follow-up visits and financial misrepresentation risks, and promote the extension of large models to professional production scenarios.

[0042] This system effectively overcomes the pain point of inaccurate facts in large models, providing a reliable fact assurance solution for demanding scenarios.

[0043] The specific implementation of a large-scale model factual error correction method based on multi-source cross-validation is as follows.

[0044] (a) Real-time generation of verification process The real-time generation and verification process is the core of the method, responsible for ensuring factual accuracy across the entire chain from user input to final accurate content output. The specific steps are as follows: Step S1: Demand Input and Fact Point Extraction Operation details: Users input their generation requirements (e.g., 2024 diagnostic criteria for stage 1 hypertension) through the Web / APP interface, and the large model generation module outputs the initial natural language content (e.g., systolic blood pressure 140-159 mmHg, diastolic blood pressure 90-99 mmHg). Technical Implementation: The fact point extraction unit uses a BERT pre-trained model combined with named entity recognition and relation extraction algorithms to automatically identify factual information such as time (e.g., 2024), data (e.g., 140-159 mmHg), and concepts (e.g., stage 1 hypertension) in the initial content, generating structured key fact points (JSON format: {Entity 1: Stage 1 hypertension; Relationship: Diagnostic criteria; Entity 2: Systolic blood pressure 140-159 mmHg / Diastolic blood pressure 90-99 mmHg; Domain tag: Medicine - Cardiovascular disease}). Data interaction: Key fact points are sent to the multi-dimensional fact verification module via the data interaction bus through synchronous interaction mode (delay not exceeding 500ms).

[0045] Step S2: Domain Adaptation and Data Source Matching Operational details: The domain adaptation unit obtains domain tags through a dual mechanism of rule matching and semantic understanding. If the request contains specific keywords (e.g., hypertension), directly match the medical-cardiovascular disease tag; If there are no explicit keywords, the semantics are analyzed using the TextCNN text classification model, and the medical insurance-employee reimbursement tag is automatically labeled. Technical parameters: Domain tag coverage: Medicine - Cardiovascular diseases, Law - Civil Code, History - Modern and contemporary events, etc. Data interaction: The domain adaptation unit selects exclusive data sources from the three-layer data source pool (e.g., the Medical-Cardiovascular Disease corresponding to the Clinical Practice Guidelines database plus the PMC academic paper database) and synchronizes the data source list to the cross-validation unit.

[0046] Step S3: Multi-source cross-validation Operational details: The cross-validation unit uses a weighted voting method for multi-source comparison. Number of comparisons: For each key fact point, select no less than 2 multi-source comparisons, and include at least 1 comparison with the domain knowledge base (e.g., medical fact points need to be compared with both the Clinical Practice Guidelines and PMC papers). Weighting rules: The weight of the domain knowledge base (0.5) is greater than the weight of the real-time trusted data source (0.3), which is greater than the weight of the general authoritative database (0.2). Result determination: If the data from multiple sources are consistent, the fact is determined to be accurate; if there is a deviation, the sum of weights is greater than 0.5, then it is accurate, otherwise it is incorrect. Technical output: When an error occurs, generate XML-formatted correction suggestions (including: error content, correct facts, data source URL, and error reason), and synchronously feed them back to the large model generation module.

[0047] Step S4: Content Optimization and Storage Operation details: After receiving correction suggestions, the large model generation module optimizes the initial content (e.g., correcting 140-159mmHg to 130-139mmHg) and outputs the final result; Data storage: The generated content caching unit stores the initial content, key fact points, and final results, and the associated dynamic knowledge base update module supports historical backtracking.

[0048] (II) Knowledge Base Dynamic Update Process This process enables real-time updates of the fact knowledge base and correction of historical errors. The specific steps are as follows: Step S5: Real-time Fact Capture and Verification Operational details: The event-driven real-time update unit uses Selenium distributed crawlers to retrieve real-time information from the authoritative domain whitelist; Technical parameters: Regular crawling frequency: once every 30 minutes; Sudden trigger mechanism: When keywords such as policy documents and major scientific discoveries are detected, they are immediately captured; Verification mechanism: After the crawled content is cleaned, it is sent to the multi-dimensional fact verification module for secondary verification (comparing with the domain knowledge base) to ensure accuracy.

[0049] Step S6: Knowledge Base Update Operational details: Verified real-time facts are categorized by domain tags and updated to the three-tier data source pool; Technical parameters: Update delay not exceeding 1 hour.

[0050] Step S7: Historical entry backtracking and correction Operational details: The historical answer backtracking and correction unit retrieves historically generated entries within the 12 months prior to the release of real-time facts (e.g., after a policy update, retrieve the Q&A records related to the policy from the past 12 months). Technical output: Generate correction prompts (including: error content, correct facts, and reason for update) for erroneous entries and push them to associated users through asynchronous interaction mode.

[0051] Step S8: Log Recording Operational details: The update log unit records the update time, content, data source URL, retrospective entry ID, and list of push users, forming an auditable log.

[0052] (III) User Feedback Optimization Process This process enables continuous model optimization driven by user feedback, and the specific steps are as follows: Step S9: Submit Feedback Operation details: Users can highlight incorrect content through the Web / APP interface and submit correct facts and domain tags (e.g., medicine - cardiovascular disease - medication). Interaction design: The interface provides drop-down options and custom label functionality, allowing users to quickly annotate.

[0053] Step S10: Feedback Verification Operational details: The feedback verification unit calls the cross-validation function of the multi-dimensional fact verification module to compare with the dedicated data source (e.g., medical drug use facts compared with the "Clinical Drug Use Guidelines" and drug instructions). Result determination: If the result is consistent with authoritative data from multiple sources, it is considered valid feedback; otherwise, it is considered invalid feedback.

[0054] Step S11: Model Optimization Operation details: Sample Transformation: The reinforcement learning optimization unit uses the PPO algorithm to transform effective feedback into training samples (format: {Input: demand plus incorrect facts; Output: correct facts; Domain label: XX; Reward value: 1.0}). Fine-tuning trigger: Fine-tuning is initiated after accumulating 100 valid samples; Parameter update: During fine-tuning, 60% of the model's bottom layer parameters are frozen, and only the top attention layer and output layer are trained. The training cycle is 5 epochs to ensure efficiency and stability. Data interaction: Training samples are transmitted to the large model generation module through asynchronous interaction mode to update model parameters.

[0055] Step S12: Feedback and Incentives Operational details: The feedback incentive unit awards 100 points to users who provide valid feedback; Redemption rules: 100 points can be redeemed for 1 hour of premium Q&A service or advanced feature permissions (such as fact-finding query). Points are stored in the user account system and can be redeemed in real time.

[0056] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A large-scale model factual error correction system based on multi-source cross-validation, characterized in that: The system includes a large model generation module, a multi-dimensional fact verification module, a dynamic fact knowledge base update module, a user feedback closed-loop error correction module, and a data interaction bus. The data interaction bus is communicatively connected to the large model generation module, the multi-dimensional fact verification module, the dynamic fact knowledge base update module, and the user feedback closed-loop error correction module, enabling real-time transmission and collaborative operation of fact point data, verification results, update logs, and user feedback information between the modules. The large model generation module receives user input generation requests, outputs initial natural language content, extracts key fact points, sends these key fact points to the multi-dimensional fact verification module, and simultaneously stores them. The system comprises: initial natural language content, key fact points, and the final result after subsequent optimization; a multi-dimensional fact verification module for performing multi-source cross-validation on the key fact points, determining fact accuracy, generating fact correction suggestions, and feeding these suggestions back to the large model generation module; a dynamic fact knowledge base update module for updating the fact knowledge base in real time and retrospectively correcting erroneous entries related to fact updates in the historical generated content stored by the large model generation module; and a user feedback closed-loop error correction module for receiving fact error information from users, verifying the authenticity of the feedback, converting valid feedback into training samples, and optimizing the large model generation module.

2. The large-scale model factual error correction system based on multi-source cross-validation according to claim 1, characterized in that: The large model generation module includes a fact point extraction unit and a generated content caching unit. The fact point extraction unit uses named entity recognition and relation extraction algorithms to automatically identify time, data, concepts, and standard factual information in the initial natural language content to form key fact points. The generated content caching unit is a non-volatile storage unit used to store the initial natural language content, the key fact points output by the fact point extraction unit, and the final result optimized by the large model generation module based on fact correction suggestions. The generated content caching unit is associated with the dynamic fact knowledge base update module and supports the retrieval and retrospection of historical generated content. The fact point extraction unit uses a BERT pre-trained model combined with named entity recognition algorithm and relation extraction algorithm; the content generation caching unit uses a distributed non-volatile storage cluster with a response time of no more than 100ms.

3. The large-scale model factual error correction system based on multi-source cross-validation according to claim 1, characterized in that: The multi-dimensional fact verification module includes a three-layer data source pool, a domain adaptation unit, and a cross-validation unit. The three-layer data source pool includes a general authoritative database, a domain-specific knowledge base, and a real-time trusted data source. The general authoritative database includes the Britannica database and a national standards database. The domain-specific knowledge base includes a clinical practice guidelines database in the medical field, a legal regulations database in the legal field, and a chronological database in the historical field. The real-time trusted data source includes government website data interfaces, authoritative media real-time information interfaces, and a database of the latest academic journal papers. The domain adaptation unit is used to obtain the domain tags corresponding to the user-generated requirements and, based on the domain tags, [the system]... The three-layer data source pool matches the corresponding dedicated data source; the cross-validation unit is used to perform multi-source comparison between the key fact points and the data in the dedicated data source matched by the domain adaptation unit, with no less than 2 multi-source comparisons and at least 1 comparison of domain professional knowledge base data; if the multi-source data are consistent, the fact is determined to be accurate; if the multi-source data have deviations, the fact is determined to be incorrect based on the domain weight, which is set to 0.5 for the domain professional knowledge base, 0.3 for the real-time trusted data source, and 0.2 for the general authoritative database; when the cross-validation unit determines that the fact is incorrect, it generates fact correction suggestions containing the correct fact content, the data source reference address, and the reason for the error; The domain adaptation unit uses a dual mechanism of rule matching and semantic understanding to obtain domain labels; when the user's requirements contain explicit domain keywords, the corresponding labels are directly matched; otherwise, the domain labels are automatically labeled by analyzing the semantics of the requirements through the TextCNN text classification model.

4. The large-scale model factual error correction system based on multi-source cross-validation according to claim 1, characterized in that: The dynamic fact knowledge base update module includes an event-driven real-time update unit, a historical answer backtracking and correction unit, and an update log unit. The event-driven real-time update unit is configured with an authoritative institution domain name whitelist. It uses web crawling technology to capture real-time fact information in the authoritative institution domain name whitelist, sends the real-time fact information to the multi-dimensional fact verification module for secondary verification, and updates it to the three-layer data source pool after the verification is passed. The historical answer backtracking and correction unit is communicatively connected to the generated content caching unit. After the event-driven real-time update unit completes the fact update, it automatically retrieves the historical generated entries related to the updated fact in the generated content caching unit. If a historical generated entry with factual errors is found, a fact correction prompt is generated and pushed to the associated user. The update log unit is used to record the update time, updated fact content, data source address, backtracking and correction historical generated entry ID, and push user list for each fact update. The event-driven real-time update unit uses Selenium distributed crawler technology, and the update delay of fact information does not exceed 1 hour; the retrieval time range of the historical answer backtracking and correction unit is the historical generated entries within 12 months before the release of real-time fact information.

5. The large-scale model factual error correction system based on multi-source cross-validation according to claim 1, characterized in that: The user feedback closed-loop error correction module includes a feedback receiving unit, a feedback verification unit, a reinforcement learning optimization unit, and a feedback incentive unit. The feedback receiving unit provides a visual interactive interface, allowing users to select erroneous factual content through highlighting and simultaneously submit correct factual information and corresponding domain tags. The feedback verification unit invokes the cross-validation function of the multi-dimensional fact verification module to perform multi-source cross-validation on the correct factual information submitted by the user, determining whether the feedback is valid or invalid. The reinforcement learning optimization unit uses the PPO reinforcement learning algorithm to transform the correct factual information in valid feedback into training samples, optimizing the large model generation module. The feedback incentive unit establishes a points reward mechanism, issuing preset points to users who submit valid feedback; these preset points can be redeemed for service time or access to advanced functions of the large model generation module. After accumulating 100 valid feedback training samples, the reinforcement learning optimization unit fine-tunes the large model generation module. During fine-tuning, 60% of the parameters at the bottom layer of the model are frozen, and only the top attention layer and output layer are trained. The preset points redemption rule of the feedback incentive unit is that 100 points can be redeemed for 1 hour of advanced question-and-answer service or corresponding advanced function permissions.

6. The large-scale model factual error correction system based on multi-source cross-validation according to claim 1, characterized in that: The data interaction bus supports both synchronous and asynchronous communication modes. Synchronous interaction mode is used when the large model generation module sends key fact points to the multi-dimensional fact verification module, and the multi-dimensional fact verification module provides fact correction suggestions to the large model generation module. Asynchronous interaction mode is used when the dynamic fact knowledge base update module pushes fact correction prompts to the user, and the user feedback closed-loop error correction module transmits domain training samples to the large model generation module.

7. A method for a large-scale model factual error correction system based on multi-source cross-validation as described in any one of claims 1-6, characterized in that: This includes a real-time generation and verification process, a dynamic knowledge base update process, and a user feedback optimization process; The real-time generation and verification process includes: user input of generation requirements, initial content generation and fact point extraction, multi-source cross-validation, and content optimization output; the knowledge base dynamic update process includes: real-time fact capture, secondary verification, data source update, historical entry backtracking, and correction prompt push; the user feedback optimization process includes: feedback submission, feedback verification, effective sample conversion, reinforcement learning model fine-tuning, and points incentive distribution.

8. The method for a large-scale model factual error correction system based on multi-source cross-validation according to claim 7, characterized in that: In the real-time generation and verification process, the cross-validation unit performs at least two multi-source comparisons, including at least one comparison of domain knowledge base data; the domain weight is set to be greater than the weight of the real-time trusted data source, which is greater than the weight of the general authoritative database.

9. The method for a large-scale model factual error correction system based on multi-source cross-validation according to claim 7, characterized in that: In the dynamic update process of the knowledge base, the crawling frequency of the event-driven real-time update unit is set to once every 30 minutes, and instant crawling is triggered when sudden important factual information is detected. The retrieval time range of the historical answer retrospective correction unit is set to historical generated entries within 12 months prior to the release of real-time factual information.

10. The method for a large-scale model factual error correction system based on multi-source cross-validation according to claim 7, characterized in that: In the user feedback optimization process, after the reinforcement learning optimization unit accumulates 100 valid feedback training samples, it fine-tunes the large model generation module. During fine-tuning, 60% of the parameters at the bottom layer of the model are frozen, and only the top attention layer and output layer are trained. The preset points redemption rule of the feedback incentive unit is that 100 points can be redeemed for 1 hour of advanced Q&A service or corresponding advanced function permissions.

Images