Data verification and screening method and system for output results of multi-source large model

Abstract

The invention provides a data verification and screening method and a data verification and screening system for output results of a multi-source large model. The method comprises a multi-source heterogeneous model parallel calling and extracting step, a first-level cross validation and anomaly elimination step, a second-level scene weight distribution step, a third-level conflict arbitration step and a log tracing and output generation step. The system comprises an input layer, a model layer, an extractor, a calibrator, a weighting device, an arbiter, an output layer and the like. According to the method, the heterogeneous large model is dynamically called, a multi-stage verification mechanism is constructed, and the problems of high illusion risk, insufficient robustness, poor scene adaptation and the like of a single model are solved, so that high-credibility result output is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically to a data verification and filtering method and system for the output results of multi-source large model, and more particularly to an algorithm and system for cross-validation, weighted fusion and conflict arbitration of inference results of multiple large language models (LLM) in the context of life services (O2O) on Internet platforms. Background Technology

[0002] With the application of large-scale model technology in the life services sector, intelligent agents have been widely used in complex tasks such as food delivery ordering, e-commerce price comparison, and logistics tracking. However, existing single large-scale model architectures suffer from the following technical bottlenecks: Hallucination risk: When large models process numerical values ​​(such as order amounts) and address entities, they may generate erroneous results that appear to be "completely nonsensical," leading to financial losses for users. Insufficient robustness: When relying on a single model (such as using only GPT-4), if the model service crashes or a logical dead end occurs, the entire business process will be interrupted; Poor scenario adaptability: Existing multi-model integration methods mostly adopt the "majority voting method", ignoring the differences in advantages of different models in different scenarios (such as DeepSeek's strong understanding of Chinese context and Qwen's superior e-commerce knowledge base).

[0003] Therefore, there is an urgent need for an algorithm that can simultaneously call multiple heterogeneous large models and ensure the absolute reliability of the output results through a three-level mechanism of "cross-validation + scenario weighting + rule arbitration". Summary of the Invention

[0004] This invention provides a data verification and filtering method and system for the output results of multi-source large models. By dynamically calling heterogeneous large models and constructing a multi-level verification mechanism, it solves the problems of high illusion risk, insufficient robustness, and poor scene adaptability of single models, thereby achieving the output of highly reliable results.

[0005] The present invention achieves the above objectives through the following technical solutions: A method for data validation and filtering of output results from a multi-source large model, comprising: Parallel invocation and extraction of multi-source heterogeneous models: Asynchronously invoke at least two pre-defined large language models and extract keyword sets from the natural language responses of each model; First-level cross-validation and anomaly removal: Construct a consistency matrix, check the similarity of the output keywords of any two models, calculate the total number of conflicts between each model and other models, and if the total number of conflicts of a model exceeds a preset threshold, the model is determined to be an anomaly model and removed. Secondary scene weight allocation: Based on the identified scene type, a weight vector is dynamically loaded, and a weighted vote is performed on the effective model set after the primary cross-validation to calculate the weighted score of the candidate results; Three-level conflict arbitration: When the calculated highest score is tied, or when there is a highest score but there is a conflict in key fields, the conflict arbitration mechanism is triggered. For numerical conflicts, a third-party API is called for verification, and for address / semantic conflicts, manual confirmation is triggered. The model output is updated according to the arbitration result. Log traceability and output generation: Generate structured logs to record model output, verification process and final results, and output unique and highly reliable results.

[0006] According to the data verification and filtering method of the multi-source large model output results provided by the present invention, in the parallel calling and extraction steps of the multi-source heterogeneous model, a standardized prompt word request is sent to the preset large language model simultaneously through an asynchronous input and output mechanism. The large language model includes: DeepSeek model, used for deep reasoning and complex Chinese semantic understanding; Qwen model, used for e-commerce domain knowledge and product parameter parsing; and Gemini model, used for multi-language processing and general common sense. For each large language model's returned natural language response, a keyword set is extracted using regular expression matching or a JSON parser. Where Amount is the amount, Address is the address, ProductName is the product name, and Time is the time.

[0007] According to the data validation and filtering method for the output results of a multi-source large model provided by the present invention, in the first-level cross-validation and outlier removal steps, a preset model set is used. M ={ m 1, m 2, m 3}, for any two models m i , m j Output keyword set K i , K j Perform Jaccard similarity calculation or full match verification; Define the conflict function Conflict( m i , m j ),when m i and m j Returns 1 if the key field values ​​in the keyword set are not equal, otherwise returns 0; Calculate each model m i Total number of conflicts with other models in the model set If a certain model m x Total number of conflicts C x >2, then determine m x The model is considered an anomaly and is removed from the weighted voting process so that it does not participate in the result fusion.

[0008] According to the data verification and filtering method for the output results of a multi-source large model provided by the present invention, the analytic hierarchy process (AHP) is introduced in the second-level scene weight allocation step to dynamically load the corresponding weight vector based on the identified scene type. W ; Set up a model m i The output result is R i , R i The scene weight in the current scenario is: w i Regarding the candidate results R Its overall confidence score (Score) R Calculated using the following formula:

[0009] Where I is the indicator function, when the model m i Output results R i With candidate results R If the values ​​match, the value is 1; otherwise, the value is 0. The number of models participating in the vote.

[0010] According to the present invention, a data verification and filtering method for the output results of a multi-source large model is provided, which dynamically loads the corresponding weight vector. W At that time, by parsing the standardized prompts or task context input by the user, the scenario type S to which the current task belongs can be identified, where S ∈{food delivery scenario, e-commerce scenario}; Build a scenario-weight vector mapping library, and define weight vectors for food delivery and e-commerce scenarios respectively. W food and W ecom ; Weight vector W food The configuration principle is as follows: Wfood =[ w DeepSeek , w Qwen , w Gemini ], and satisfy w DeepSeek ≥ w Qwen > w Gemini ; Weight vector W ecom The configuration principle is as follows: W ecom =[ w DeepSeek ′, w Qwen ′, w Gemini ′], and satisfy w Qwen ′> w DeepSeek ′≥ w Gemini ′; Based on the scene type output by the scene recognition module S Match the corresponding weight vector from the scene-weight vector mapping library. W : like S =In the food delivery scenario, then load W = W food ; like S =In e-commerce scenarios, then load W = W ecom .

[0011] According to the data verification and filtering method for the output results of a multi-source large model provided by the present invention, in the three-level conflict arbitration step, the conflict arbitration mechanism is a three-level conflict arbitration process, including: Numerical conflict handling: For conflicts in key fields such as amount or inventory in candidate results, a trusted source verification mechanism is executed: the current inference process is paused, the real-time data of the corresponding key field is directly called through the third-party API interface, and the real-time data returned by the third-party API interface is used as the final result, forcibly overriding the output result of the large model; Address / Semantic Conflict Handling: For conflicts in address or semantic key fields in candidate results, a conservative strategy is implemented: Since the authenticity of the conflicting content cannot be verified through a third-party API, a manual confirmation request with prompts is generated, and a prompt window pops up on the user's terminal. The content of the prompt window is used to prompt the user to confirm the correct value of the ambiguous key field. After waiting for the user's confirmation information, the final result is output based on the user's confirmation result. Dynamic updates to the arbitration rule base: Configure scheduled tasks to retrieve the latest compliance rule information from the regulatory platform or business back-end system according to preset time cycles, and use the latest compliance rule information to update the rule entries in the arbitration logic base.

[0012] According to the data verification and filtering method for multi-source large model output results provided by the present invention, in the log tracing and output generation steps, a uniquely determined high-confidence result R is output. final Generate structured logs, represented as:

[0013] Wherein, TraceID is the trace identifier, Inputs is the user input information, Outputs{m1,m2,m3} is the output result of each preset large language model, Weights is the scene weight vector, and Decision_Path is the decision path of the fusion verification process.

[0014] According to the data validation and filtering method for multi-source large model output results provided by the present invention, when performing Jaccard similarity calculation, from K i and K j Extract a set of keywords, where each keyword corresponds to a unique value, forming a set of keyword-value pairs; right K i and K j Perform consistency checks on the keyword-value pair set: calculate K i and K j Jaccard similarity J ( K i , K j The formula is:

[0015] in, express K i and Kj A set whose keyword-value pairs are completely identical. express K i and K j The keyword-value pair union, |.| represents the number of elements in the set; Compare one by one K i and K j The value of each keyword is considered. If all the keyword values ​​are the same, it is considered a complete match; otherwise, it is considered a mismatch. Define a conflict function Conflict( based on the consistency check result). m i , m j If Jaccard similarity J ( K i , K j If the similarity is less than the preset similarity threshold, or if the exact match check determines that it is not a match, then Conflict( m i , m j ) = 1; otherwise Conflict( m i , m j )=0.

[0016] According to the data verification and filtering method for the output results of a multi-source large model provided by the present invention, after the three-level conflict arbitration, a dynamic feedback correction module is added to perform statistical analysis on the deviation between the model output results and the final arbitration results, generate a model performance evaluation report, and dynamically adjust the scene weight vector W or the conflict arbitration rule base. Record the difference between the output of each model in historical tasks and the arbitration result, and calculate the model's performance in the scenario. S Confidence score (under) m k , S ), represented as:

[0017] in, For the model m k In the scene S The total number of historical tasks under the current system The output should show the number of tasks whose results match the arbitration result. Adjust the scene weight vector based on confidence scoreW , represented as:

[0018] in, For the model m k In the scene S The updated weights below α This is the attenuation coefficient, used to balance historical weights with current performance. For the model m k In the scene S The historical weighting, m k For the first k A large language model.

[0019] A data verification and filtering system for the output results of a multi-source large model is provided. This system is applied to the aforementioned data verification and filtering method for the output results of a multi-source large model. The system includes: The input layer is used to receive standardized prompt word requests from the user and forward them to the model layer; The model layer includes at least two pre-defined large language models for parallel processing of standardized prompt word requests and generating natural language responses. Extractor, used to extract a set of keywords from the natural language responses of the model layer; The validator performs first-level cross-validation and outlier removal steps to remove outlier models. The weighter performs the secondary scene weight allocation step and calculates the weighted score of the candidate results; The arbitrator performs a three-level conflict arbitration process, handles key field conflicts, and generates the final arbitration result. The output layer includes an external knowledge base to support the arbitrator's decisions, and a fusion engine to integrate the outputs of each module, generate structured logs, and output the final results.

[0020] Therefore, compared with existing technologies, the data verification and filtering method and system for multi-source large model output results proposed in this invention significantly improves the reliability, adaptability, and maintainability of large model applications in life service scenarios, and has the following beneficial effects: 1. This invention uses a three-level verification mechanism (cross-validation to eliminate illusory models, scenario-weighted selection of advantageous results, and conflict arbitration as a fallback) to eliminate risks such as numerical errors and entity contradictions in the model output, ensuring that the final result meets business logic and user needs, and avoiding economic losses or operational accidents caused by model errors.

[0021] 2. This invention can dynamically load scenario weight vectors and arbitration rule bases according to the characteristics of different scenarios such as food delivery and e-commerce, giving full play to the advantages of each model in different fields (such as Chinese contextual understanding and e-commerce knowledge base matching), avoiding the limitations of a single model or fixed strategy, and achieving efficient and accurate processing of complex tasks.

[0022] 3. This invention records key information such as decision-making path, model output, and weight allocation in detail through structured logs, forming a complete traceability chain. Through backtracking analysis, it can quickly locate model error patterns, providing data support for weight adjustment, rule updates, and model iteration, forming a closed loop of "verification-feedback-optimization" to continuously improve system performance and user experience.

[0023] In summary, this invention solves the core problems of low reliability and poor adaptability of a single large model in life service scenarios through the collaborative innovation of three-level verification, dynamic adaptation and traceable design, and provides a safe, efficient and evolvable solution for intelligent applications in high-risk and multi-scenario applications.

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0025] Figure 1 This is a flowchart of an embodiment of a data verification and filtering method for the output results of a multi-source large model according to the present invention.

[0026] Figure 2 This is a schematic diagram of the first step in an embodiment of a data verification and filtering method for the output results of a multi-source large model according to the present invention.

[0027] Figure 3 This is a schematic diagram of the second step in an embodiment of a data verification and filtering method for the output results of a multi-source large model according to the present invention.

[0028] Figure 4 This is a schematic diagram of the first-level cross-validation and keyword consistency detection matrix in an embodiment of the data verification and filtering method for the output results of a multi-source large model according to the present invention.

[0029] Figure 5 This is a decision logic tree diagram regarding three-level conflict arbitration in an embodiment of a data verification and filtering method for the output results of a multi-source large model according to the present invention.

[0030] Figure 6 This is a modular architecture diagram of an embodiment of a data verification and filtering system for multi-source large model output results according to the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0033] An Example of a Data Validation and Filtering Method for Output Results of a Multi-Source Large Model See Figures 1 to 5 This embodiment provides a method for data verification and filtering of output results from a multi-source large model, including: Parallel invocation and extraction of multi-source heterogeneous models: Asynchronously invoke at least two pre-defined large language models and extract keyword sets from the natural language responses of each model; First-level cross-validation and anomaly removal: Construct a consistency matrix, check the similarity of the output keywords of any two models, calculate the total number of conflicts between each model and other models, and if the total number of conflicts of a model exceeds a preset threshold, the model is determined to be an anomaly model and removed. Secondary scene weight allocation: Based on the identified scene type, a weight vector is dynamically loaded, and a weighted vote is performed on the effective model set after the primary cross-validation to calculate the weighted score of the candidate results; Three-level conflict arbitration: When the calculated highest score is tied, or when there is a highest score but there is a conflict in key fields, the conflict arbitration mechanism is triggered. For numerical conflicts, a third-party API is called for verification, and for address / semantic conflicts, manual confirmation is triggered. The model output is updated according to the arbitration result. Log traceability and output generation: Generate structured logs to record model output, verification process and final results, and output unique and highly reliable results.

[0034] In the parallel invocation and extraction steps of multi-source heterogeneous models, standardized prompt requests are sent to the preset large language models simultaneously through an asynchronous input-output mechanism. The large language models include: DeepSeek model, used for deep reasoning and complex Chinese semantic understanding; Qwen model, used for e-commerce domain knowledge and product parameter parsing; and Gemini model, used for multi-language processing and general common sense. For each large language model's returned natural language response, a keyword set is extracted using regular expression matching or a JSON parser. Where Amount is the amount, Address is the address, ProductName is the product name, and Time is the time.

[0035] As can be seen, the above steps improve system processing efficiency and result diversity by calling multiple large models in parallel and extracting keywords, thus avoiding the performance bottleneck of a single model. For different tasks (such as Chinese semantics and e-commerce parameters), appropriate models are selected, and standardized extraction (regular expressions / JSON) is used to ensure the structure of keywords and improve the accuracy of verification.

[0036] In the first-level cross-validation and outlier removal step, for the preset model set... M ={ m 1, m 2, m 3}, for any two models m i , m j Output keyword set K i , K j Perform Jaccard similarity calculation or full match verification; Define the conflict function Conflict( m i , m j ),when m i and m j Returns 1 if the values ​​of the key fields (amount, address, product name, time) in the keyword set are not equal, otherwise returns 0; Calculate each model m i Total number of conflicts with other models in the model set If a certain model m x Total number of conflicts C x If the value is >2 (meaning it doesn't match any other model), then it is determined that...m x The model is an abnormal model (hallucination model), and it is removed from the weighted voting so that it does not participate in the result fusion.

[0037] As can be seen, the above steps quickly identify and eliminate "illusion models" through consistency matrix and conflict threshold, reduce interference from low-quality results, and improve system robustness.

[0038] In the second-level scenario weight allocation step, the Analytic Hierarchy Process (AHP) is introduced to dynamically load the corresponding weight vector based on the identified scenario type (food delivery scenario or e-commerce scenario). W ; Set up a model m i The output result is R i , R i The scene weight in the current scenario is: w i Regarding the candidate results R Its overall confidence score (Score) R Calculated using the following formula:

[0039] Where I is the indicator function, when the model m i Output results R i With candidate results R If the values ​​match, the value is 1; otherwise, the value is 0. The number of models participating in the vote; Dynamically load the corresponding weight vector W At that time, by parsing the standardized prompts or task context input by the user, the scenario type S to which the current task belongs can be identified, where S ∈{food delivery scenario, e-commerce scenario}; Build a scenario-weight vector mapping library, and define weight vectors for food delivery and e-commerce scenarios respectively. W food and W ecom ; Weight vector W food The configuration principle is to assign higher weights to models with strong deep reasoning capabilities and complex Chinese semantic understanding capabilities, for example... W food =[ w DeepSeek , w Qwen , wGemini ], and satisfy w DeepSeek ≥ w Qwen > w Gemini ; Weight vector W ecom The configuration principle is to assign higher weights to models with strong e-commerce domain knowledge and product parameter analysis capabilities, for example... W ecom =[ w DeepSeek ′, w Qwen ′, w Gemini ′], and satisfy w Qwen ′> w DeepSeek ′≥ w Gemini ′; Based on the scene type output by the scene recognition module S Match the corresponding weight vector from the scene-weight vector mapping library. W : like S =In the food delivery scenario, then load W = W food ; like S =In e-commerce scenarios, then load W = W ecom ; For example, the corresponding configuration of scene type and weight vector is as follows: In the context of food delivery, weight vector W food ={DeepSeek model: 0.4, Qwen model: 0.4, Gemini model: 0.2}, DeepSeek and Qwen outperform Gemini in understanding local life services; In e-commerce scenarios, weight vectors W ecom ={Qwen model: 0.4, DeepSeek model: 0.3, Gemini model: 0.3}, Qwen is fully trained on Alibaba's e-commerce corpus and is given the highest weight.

[0040] As can be seen, the above steps, by dynamically loading scenario-based weight vectors, make the model voting results more in line with actual needs (such as address for food delivery and product database for e-commerce), thus optimizing scenario adaptability; and clarify the weight allocation logic for food delivery / e-commerce scenarios (such as Qwen having the highest weight in e-commerce scenarios), ensuring that the model's advantages are accurately matched with scenario requirements, thereby improving the credibility of the results.

[0041] In the three-tier conflict arbitration process, the conflict arbitration mechanism is a three-tier conflict arbitration procedure, including: Numerical conflict handling: For conflicts in key fields such as amount or inventory in candidate results, a trusted source verification mechanism is executed: the current inference process is paused, and a third-party API interface (including but not limited to Meituan Open Platform interface, Taobao Open Platform interface, etc.) is directly called to query the real-time data of the corresponding key field. The real-time data returned by the third-party API interface is used as the final result, forcibly overriding the output result of the large model. Address / Semantic Conflict Handling: For conflicts in address or semantic key fields in candidate results, a conservative strategy is implemented: Since the authenticity of the conflicting content cannot be verified through a third-party API, a manual confirmation request with prompts is generated, and a prompt window pops up on the user's terminal. The content of the prompt window is used to prompt the user to confirm the correct value of the ambiguous key field (e.g., "An ambiguous address has been detected. Please confirm whether it is Building A or Building B?"). After waiting for the user's confirmation, the final result is output based on the user's confirmation. Dynamic updates to the arbitration rule base: Configure a scheduled task (Cron Job) to retrieve the latest compliance rule information (including but not limited to the latest list of prohibited drugs and business compliance requirements) from the regulatory platform or business back-end system according to a preset time period (e.g., every 7 days), and use the latest compliance rule information to update the rule entries in the arbitration logic base to ensure that the arbitration logic base is consistent with the current regulatory requirements or business rules and prevent the model from outputting content with compliance risks.

[0042] As can be seen, the above steps resolve key field conflicts through numerical API validation (such as real-time price queries) and semantic manual confirmation, ensuring the absolute reliability of the final result and avoiding economic losses.

[0043] In the log tracing and output generation steps, a unique and highly reliable result R is output. final Generate structured logs:

[0044] Where TraceID is the tracking identifier, Inputs is the user input information, Outputs{m1,m2,m3} is the output result of each preset large language model, Weights is the scene weight vector, and Decision_Path is the decision path of the fusion verification process; Structured logs Used for bad case analysis, it identifies the pattern of incorrect outputs by the model in the corresponding scenario by tracing back the decision path, model output results and scenario weight information in the log.

[0045] As can be seen, the above steps, by recording the entire decision-making process and model output, support subsequent auditing and bad case analysis, thereby improving system traceability and optimization efficiency.

[0046] When performing Jaccard similarity calculation, from K i and K j Extract a set of keywords, including Amount, Address, Product Name, and Time. Each keyword corresponds to a unique value, forming a set of keyword-value pairs. right K i and K j Perform consistency checks on the keyword-value pair set: calculate K i and K j Jaccard similarity J ( K i , K j The formula is:

[0047] in, express K i and K j A set whose keyword-value pairs are completely identical. express K i and K j The keyword-value pair union, |.| represents the number of elements in the set; Compare one by one K i and K j The value of each keyword is considered. If all the keyword values ​​are the same, it is considered a complete match; otherwise, it is considered a mismatch. Define a conflict function Conflict( based on the consistency check result). mi , m j If Jaccard similarity J ( K i , K j If the similarity is less than the preset similarity threshold, or if the exact match check determines that it is not a match, then Conflict( m i , m j ) = 1; otherwise Conflict( m i , m j )=0.

[0048] As can be seen, the above steps, which combine similarity calculation and strict matching, improve the sensitivity of conflict detection and avoid incorrect rejection or omission of abnormal models due to inconsistencies in some fields.

[0049] After the three-level conflict arbitration, a dynamic feedback correction module is added to statistically analyze the deviation between the model output and the final arbitration result, generate a model performance evaluation report, and dynamically adjust the scenario weight vector W or the conflict arbitration rule base. Record the differences between the output of each model in historical tasks and the arbitration results (such as amount deviation rate, address error rate), and calculate the model's performance in the scenario. S Confidence score (under) m k , S ), represented as:

[0050] in, For the model m k In the scene S The total number of historical tasks under the current system The number of tasks whose output matches the arbitration result. Adjust the scene weight vector based on confidence score W ,For example:

[0051] in, For the model m k In the scene S The updated weights below α This is a decay factor (e.g., 0.9), used to balance historical weights with current performance. For the model m k In the sceneS The historical weighting, m k For the first k A large language model; If a certain type of conflict (such as "address ambiguity") frequently triggers manual confirmation, the threshold for semantic conflicts will be automatically lowered to reduce manual intervention.

[0052] As can be seen, the above steps form a "verification-feedback-optimization" closed loop through model performance evaluation and dynamic weight adjustment, continuously improving the long-term accuracy and adaptability of the system.

[0053] In the log traceability module, a data anonymization submodule is added to encrypt and store sensitive information (such as user addresses and phone numbers); and a compliance verification layer is added before conflict arbitration to filter out illegal content output by the model; and sensitive fields in core keywords (such as addresses) are encrypted using hash encryption (such as SHA-256) or partially masked (such as "Beijing ** District"). The structured logs store only anonymized data, while the original data is encrypted and stored in a separate database. A predefined library of prohibited keywords is used to detect prohibited content in the model output through string matching or semantic similarity calculation (such as the BERT model). If a violation is detected, manual confirmation or rejection of the output is triggered directly, and the compliance event is recorded.

[0054] In this embodiment, the data anonymization submodule includes: a sensitive information identification unit, which automatically identifies sensitive information fields in structured logs using a predefined sensitive field rule base (such as regular expression templates for user addresses, phone numbers, ID card numbers, bank card numbers, etc.); a dynamic anonymization processing unit, which uses irreversible encryption algorithms (such as SHA-256) or partial masking (such as the phone number "138****1234") to process the identified sensitive fields and generate anonymized log data; and a dynamic anonymization rule update unit, which synchronizes the latest sensitive information types (such as adding a "biometric information" category) from the regulatory platform or business backend through a scheduled task (Cron Job) and updates the sensitive field rule base in real time to ensure that the anonymization scope covers the latest compliance requirements.

[0055] In this embodiment, the compliance verification layer includes: a violation content detection unit, which calls a pre-trained compliance detection model (such as a BERT-based text classification model) or a rule engine (such as keyword blacklist matching) to identify violation information (such as pornography, violence, false advertising, prohibited product names, etc.) in the output content before the model output enters the three-level conflict arbitration; a dynamic rule base loading unit, which synchronizes the latest compliance rules (such as "the latest list of prohibited drugs" and "the prohibited word library of the Advertising Law") from the regulatory platform (such as the Market Supervision Administration API) or the business backend, and dynamically updates the rule base of the violation content detection unit to ensure that the verification logic is consistent with the latest regulations; and an output interception and alarm unit, which immediately intercepts the model output if violation content is detected, does not enter the subsequent arbitration process, sends alarm information (such as email or SMS) to the system administrator, and records the violation type, model name, and timestamp to the structured log.

[0056] Specifically, the model output first passes through the compliance verification layer. If the verification passes, it enters the log tracing module to generate the original log. The original log is then processed by the data anonymization submodule and stored in the database or output to an external analysis system. If the compliance verification fails, an anonymized log containing the "violation interception" mark is directly generated, and the subsequent arbitration process is skipped. The anonymized log retains non-sensitive metadata such as model output time, model name, and conflict type, and supports association with the original request context through a unique transaction ID to meet regulatory audit requirements.

[0057] It is evident that dynamic desensitization ensures that sensitive information is irreversibly encrypted during log storage, transmission, and analysis, thus avoiding the risk of user privacy leaks. A real-time updated compliance verification layer filters out illegal content output by the model, preventing the system from becoming a channel for the spread of illegal information and reducing legal risks. Desensitized logs retain key metadata, supporting regulatory authorities in conducting compliance reviews of the model's decision-making process while protecting user privacy.

[0058] An embodiment of a data verification and filtering system for the output results of a multi-source large model like Figure 6 As shown, this embodiment provides a data verification and filtering system for the output results of a multi-source large model. This system is applied to a data verification and filtering method for the output results of a multi-source large model as described in the above embodiment. The system includes: The input layer is used to receive standardized prompt word requests from the user and forward them to the model layer; The model layer includes at least two pre-defined large language models for parallel processing of standardized prompt word requests and generating natural language responses. Extractor, used to extract a set of keywords from the natural language responses of the model layer; The validator performs first-level cross-validation and outlier removal steps to remove outlier models. The weighter performs the secondary scene weight allocation step and calculates the weighted score of the candidate results; The arbitrator performs a three-level conflict arbitration process, handles key field conflicts, and generates the final arbitration result. The output layer includes an external knowledge base (External KB) to support the arbitrator's decision-making, and a fusion engine to integrate the outputs of each module, generate structured logs, and output the final results.

[0059] In this embodiment, the arbitrator further includes: The numerical conflict handling module is used to call third-party APIs to verify numerical conflicts and to override the model results with the API return values. The semantic conflict handling module is used to generate a manual confirmation request when the API cannot be called to verify semantic conflicts, and to prompt the user to confirm the ambiguous content through a pop-up window.

[0060] The input layer deploys Nginx as a reverse proxy to receive user HTTP / HTTPS requests. It uses FastAPI to build a gateway service responsible for request authentication (JWT), rate limiting (Redis token bucket algorithm), and standardization cleaning. It transforms user natural language (e.g., "Buy me a milk tea") into structured Prompts, including Task_Type, Context, and Constraints. The core components of the input layer include the API Gateway and the PromptTemplate Engine.

[0061] The model layer deploys inference services (vLLM or TGI frameworks) for models such as DeepSeek, Qwen, and Gemini using Docker containers, exposing gRPC or HTTP interfaces. It uses Python asyncio or Celery task queues to enable simultaneous non-blocking calls from multiple models, setting a uniform timeout (e.g., 3 seconds). If a model experiences excessive concurrency, Kubernetes automatically scales the system. The core components of the model layer include Model Worker (inference container) and Async Scheduler (asynchronous scheduler).

[0062] The extractor utilizes a rule engine to develop a regular expression library for fields such as amount, time, and address. For unstructured responses, it employs a lightweight BERT model or a low-parameter LLM for slot filling, forcing the output to be in JSON format: { "Amount": "20", "Address": "..."}. Furthermore, it removes Markdown tags and irrelevant text from the model's response, retaining only the core entities. The core components of the extractor include RegexMatcher, JSON Parser, and Slot Filling Model.

[0063] The validator constructs a consistency matrix by building a two-dimensional table in memory, where rows represent models and columns represent keywords (amount, address, etc.). It uses a conflict calculation engine to implement the Jaccard similarity algorithm and hash comparison algorithm, and configures a rule engine (Drools). If Conflict > 2, the model ID is marked as INVALID and removed from subsequent processes. The core components of the validator include the Consistency Matrix and the Conflict Calculator.

[0064] The weighting unit identifies scenarios S using a scene recognizer based on keyword matching (e.g., e.g., "product ID" indicates e-commerce, "delivered" indicates food delivery) or a text classification model; it stores Scene_Weights (e.g., Food_Scene: {DeepSeek: 0.4, Qwen: 0.4, Gemini: 0.2}) using a Redis Hash structure through a weight library; and it calculates the score (Score(R)) of each candidate result based on the retained set of valid models and the weight vector using a voting calculator. The core components of the weighting unit include Scene Classifier, Redis Config Store, and Voting Engine.

[0065] The arbitrator determines the conflict type through rule routing. If the conflict is numerical (amount / inventory), it calls an external API Client; if it is semantic (address), it triggers a manual confirmation task. It establishes an API connection pool with a third-party open platform to enable real-time data queries. A WebSocket service is built through a human-computer interaction interface to push pop-up requests to the front end and block subsequent processes until user feedback is received. A rule updater scheduled task (Crontab) pulls rules from the regulatory platform and hot-updates the local arbitration logic. The core components of the arbitrator include an API Connector, a WebSocket Server, and a Rule Engine.

[0066] The output layer aggregates arbitration results, original model outputs, weights, and decision paths through a fusion engine. The log system uses the ELK Stack (Elasticsearch, Logstash, Kibana) or Loki, employing asynchronous writing to record structured logs. Before writing to disk, the data masking module processes sensitive fields (phone numbers, addresses) using AES encryption or SHA-256 masking. Before being returned to the user, compliance filtering performs a final content scan using a trie tree or a large security model. The core components of the output layer include the Log Aggregator, Data Masking Module, and ComplianceChecker.

[0067] In practical applications, assuming a user initiates a "place a food delivery order" request, the internal system flow is as follows: Step 1: Request Access and Distribution The user sent a request: "Please order a 20 yuan braised chicken for me on the 3rd floor of Building A".

[0068] The input layer cleans the request, identifies the scenario as food delivery, extracts the intent parameters, and asynchronously distributes them to DeepSeek, Qwen, and Gemini in the model layer.

[0069] Step 2: Parallel Invocation and Extraction of Multi-Source Heterogeneous Models The model layer generates responses simultaneously: DeepSeek: "The address is 3rd floor of Building A, and the amount is 20 yuan." Qwen: "Address: Room 301, Building A; Amount: 25 yuan" Gemini: "Address: 1st Floor, Building B; Amount: 20 yuan" The extractor parsed three sets of JSON: D: {Addr:"Building A, 3rd Floor", Amt:"20"} Q: {Addr:"Building A, Room 301", Amt:"25"} G: {Addr:"Building B, 1st Floor", Amt:"20"} Step 3: First-level cross-validation and outlier removal The checker calculates the collision matrix: D vs Q: Address conflict (1) + Amount conflict (1) = 2 points D vs G: Address conflict (1) + Amount match (0) = 1 point Q vs G: Address conflict (1) + Amount conflict (1) = 2 points Total number of conflicts: D(3), Q(4), G(3).

[0070] Judgment: Assuming a threshold of 3, if the number of Q-collisions is greater than 3, Qwen is determined to be an abnormal model and is removed. Remaining valid models: DeepSeek, Gemini.

[0071] Step 4: Secondary Scene Weight Allocation and Voting The weighted calculator identifies the scenario as food delivery and loads weight W. food = [DeepSeek:0.4, Gemini:0.2] (Qwen has been removed, and its weights have been reset to zero or normalized).

[0072] Candidate results: Result A (D): {Addr:"Building A, 3rd Floor", Amt:"20"} Result B (G): {Addr:"Building B, 1st Floor", Amt:"20"} Calculate the score: Score(A) = 0.4 (D supported) + 0 (G not supported) = 0.4 Score(B) = 0 (D not supported) + 0.2 (G supported) = 0.2 Conclusion: Result A scored the highest (0.4), but the inspection revealed a conflict in the address field (addresses D and G are different), triggering a three-level arbitration.

[0073] Step 5: Three-level conflict arbitration The arbitrator detected a semantic conflict in the "address" field, but the amounts were the same.

[0074] Decision: Unable to verify the authenticity of the address via API, triggering manual confirmation.

[0075] Interaction: The system pops up a window to the front end: "Address ambiguity detected. DeepSeek thinks it is the 3rd floor of Building A, while Gemini thinks it is the 1st floor of Building B. Please confirm."

[0076] User feedback: The user clicked on "Building A, 3rd Floor".

[0077] Arbitration result: The user's confirmed "3rd floor of Building A" is adopted, and the amount remains "20 yuan".

[0078] Step 6: Log recording and output The output layer generates the final result: "A 20 yuan braised chicken has been ordered for you on the 3rd floor of Building A".

[0079] Log tracing: Generate structured logs (including TraceID), recording: Original input, output of each model, conflict calculation process, reasons for Qwen being removed, and user selection during arbitration.

[0080] Anonymized storage: The specific house number "3rd floor" in the log is masked and then stored in the database.

[0081] Model optimization: The dynamic feedback module records the error in Qwen and lowers its confidence score in the food delivery scenario, and automatically reduces its weight next time.

[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for data verification and filtering of output results from a multi-source large model, characterized in that, include: Parallel invocation and extraction of multi-source heterogeneous models: Asynchronously invoke at least two pre-defined large language models and extract keyword sets from the natural language responses of each model; First-level cross-validation and anomaly removal: Construct a consistency matrix, check the similarity of the output keywords of any two models, calculate the total number of conflicts between each model and other models, and if the total number of conflicts of a model exceeds a preset threshold, the model is determined to be an anomaly model and removed. Secondary scene weight allocation: Based on the identified scene type, a weight vector is dynamically loaded, and a weighted vote is performed on the effective model set after the primary cross-validation to calculate the weighted score of the candidate results; Three-level conflict arbitration: When the calculated highest score is tied, or when there is a highest score but there is a conflict in key fields, the conflict arbitration mechanism is triggered. For numerical conflicts, a third-party API is called for verification, and for address / semantic conflicts, manual confirmation is triggered. The model output is updated according to the arbitration result. Log traceability and output generation: Generate structured logs to record model output, verification process and final results, and output unique and highly reliable results.

2. The method according to claim 1, characterized in that: In the parallel invocation and extraction steps of multi-source heterogeneous models, standardized prompt word requests are sent to the preset large language model simultaneously through an asynchronous input-output mechanism. The large language model includes: DeepSeek model, used for deep reasoning and complex Chinese semantic understanding; Qwen model, used for e-commerce domain knowledge and product parameter parsing; and Gemini model, used for multi-language processing and general common sense. For each large language model's returned natural language response, a keyword set is extracted using regular expression matching or a JSON parser. Where Amount is the amount, Address is the address, ProductName is the product name, and Time is the time.

3. The method according to claim 1, characterized in that: In the first-level cross-validation and outlier removal step, for the preset model set... M ={ m 1, m 2, m 3}, for any two models m i , m j Output keyword set K i , K j Perform Jaccard similarity calculation or full match verification; Define the conflict function Conflict( m i , m j ),when m i and m j Returns 1 if the key field values ​​in the keyword set are not equal, otherwise returns 0; Calculate each model m i Total number of conflicts with other models in the model set If a certain model m x Total number of conflicts C x >2, then determine m x The model is considered an anomaly and is removed from the weighted voting process so that it does not participate in the result fusion.

4. The method according to claim 1, characterized in that: In the second-level scene weight allocation step, the concept of the analytic hierarchy process is introduced to dynamically load the corresponding weight vector based on the identified scene type. W ; Set up a model m i The output result is R i , R i The scene weight in the current scenario is: w i Regarding the candidate results R Its overall confidence score (Score) R Calculated using the following formula: Where I is the indicator function, when the model m i Output results R i With candidate results R If the values ​​match, the value is 1; otherwise, the value is 0. The number of models participating in the vote.

5. The method according to claim 4, characterized in that: Dynamically load the corresponding weight vector W At that time, by parsing the standardized prompts or task context input by the user, the scenario type S to which the current task belongs can be identified, where S ∈{food delivery scenario, e-commerce scenario}; Build a scenario-weight vector mapping library, and define weight vectors for food delivery and e-commerce scenarios respectively. W food and W ecom ; Weight vector W food The configuration principle is as follows: W food =[ w DeepSeek , w Qwen , w Gemini ], and satisfy w DeepSeek ≥ w Qwen > w Gemini ; Weight vector W ecom The configuration principle is as follows: W ecom =[ w DeepSeek ′, w Qwen ′, w Gemini ′], and satisfy w Qwen ′> w DeepSeek ′≥ w Gemini ′; Based on the scene type output by the scene recognition module S Match the corresponding weight vector from the scene-weight vector mapping library. W : like S =In the food delivery scenario, then load W = W food ; like S =In e-commerce scenarios, then load W = W ecom .

6. The method according to claim 1, characterized in that: In the three-tier conflict arbitration process, the conflict arbitration mechanism is a three-tier conflict arbitration procedure, including: Numerical conflict handling: For conflicts in key fields such as amount or inventory in candidate results, a trusted source verification mechanism is executed: the current inference process is paused, the real-time data of the corresponding key field is directly called through the third-party API interface, and the real-time data returned by the third-party API interface is used as the final result, forcibly overriding the output result of the large model; Address / Semantic Conflict Handling: For conflicts in address or semantic key fields in candidate results, a conservative strategy is implemented: Since the authenticity of the conflicting content cannot be verified through a third-party API, a manual confirmation request with prompts is generated, and a prompt window pops up on the user's terminal. The content of the prompt window is used to prompt the user to confirm the correct value of the ambiguous key field. After waiting for the user's confirmation information, the final result is output based on the user's confirmation result. Dynamic updates to the arbitration rule base: Configure scheduled tasks to retrieve the latest compliance rule information from the regulatory platform or business back-end system according to preset time cycles, and use the latest compliance rule information to update the rule entries in the arbitration logic base.

7. The method according to claim 1, characterized in that: In the log tracing and output generation steps, a unique and highly reliable result R is output. final Generate structured logs, represented as: Wherein, TraceID is the trace identifier, Inputs is the user input information, Outputs{m1,m2,m3} is the output result of each preset large language model, Weights is the scene weight vector, and Decision_Path is the decision path of the fusion verification process.

8. The method according to claim 3, characterized in that: When performing Jaccard similarity calculation, from K i and K j Extract a set of keywords, where each keyword corresponds to a unique value, forming a set of keyword-value pairs; right K i and K j Perform consistency checks on the keyword-value pair set: calculate K i and K j Jaccard similarity J ( K i , K j The formula is: in, express K i and K j A set whose keyword-value pairs are completely identical. express K i and K j The keyword-value pair union, |.| represents the number of elements in the set; Compare one by one K i and K j The value of each keyword is considered. If all the keyword values ​​are the same, it is considered a complete match; otherwise, it is considered a mismatch. Define a conflict function Conflict( based on the consistency check result). m i , m j If Jaccard similarity J ( K i , K j If the similarity is less than the preset similarity threshold, or if the exact match check determines that it is not a match, then Conflict( m i , m j ) = 1; otherwise Conflict( m i , m j )=0.

9. The method according to any one of claims 1 to 8, characterized in that: After the three-level conflict arbitration, a dynamic feedback correction module is added to statistically analyze the deviation between the model output and the final arbitration result, generate a model performance evaluation report, and dynamically adjust the scenario weight vector W or the conflict arbitration rule base. Record the difference between the output of each model in historical tasks and the arbitration result, and calculate the model's performance in the scenario. S Confidence score (under) m k , S ), represented as: in, For the model m k In the scene S The total number of historical tasks under the current system The output should show the number of tasks whose results match the arbitration result. Adjust the scene weight vector based on confidence score W , represented as: in, For the model m k In the scene S The updated weights below α This is the attenuation coefficient, used to balance historical weights with current performance. For the model m k In the scene S The historical weighting, m k For the first k A large language model.

10. A data verification and filtering system for the output results of a multi-source large model, characterized in that, This system is applied to a data verification and filtering method for the output results of a multi-source large model as described in any one of claims 1 to 9, and the system includes: The input layer is used to receive standardized prompt word requests from the user and forward them to the model layer; The model layer includes at least two pre-defined large language models for parallel processing of standardized prompt word requests and generating natural language responses. Extractor, used to extract a set of keywords from the natural language responses of the model layer; The validator performs first-level cross-validation and outlier removal steps to remove outlier models. The weighter performs the secondary scene weight allocation step and calculates the weighted score of the candidate results; The arbitrator performs a three-level conflict arbitration process, handles key field conflicts, and generates the final arbitration result. The output layer includes an external knowledge base to support the arbitrator's decisions, and a fusion engine to integrate the outputs of each module, generate structured logs, and output the final results.

Images