Intelligent question answering system based on large-scale pre-training model and optimization method thereof

Abstract

The invention discloses an intelligent question answering system based on a large-scale pre-training model and an optimization method thereof, and particularly relates to the technical field of artificial intelligence question answering, comprising the following steps: building a model resource pool by presetting a plurality of large-scale pre-training models, receiving a question answering request of a user, and generating an answer by a current model; according to the answer confidence degree information, statistical characteristics are extracted to judge whether to trigger dual suitability evaluation or not; the model answer adaptation degree and the interaction adaptation degree are evaluated, and the current model is divided into a high-degree adaptation type or a low-degree adaptation type; if the type is the low-degree adaptation type, selecting an optimization model from the model resource pool to realize dynamic switching in combination with an evaluation result; through double evaluation of the answer adaptation index and the interaction adaptation index, the problem that the answer quality of the model is reduced or the interaction is abnormal is found in real time, the model is switched in time, the stability and reliability of the question answering system are guaranteed, and the risks of model illusion, non-questions or non-compliant answers are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence question - answering, and more specifically, to an intelligent question - answering system based on a large - scale pre - trained model and an optimization method thereof. Background Art

[0002] With the continuous progress of artificial intelligence technology, natural language processing systems based on large - scale pre - trained models have been widely applied to various application scenarios such as intelligent question - answering, virtual assistants, and intelligent customer service. Large - scale pre - trained models rely on huge training corpora and complex deep - learning architectures, and have achieved remarkable results in language understanding and text generation. However, most existing intelligent question - answering systems use a single pre - trained model to provide question - answering services. Although they can achieve a certain degree of language understanding and information retrieval, they have obvious deficiencies in dealing with complex, diverse, and dynamically changing user needs.

[0003] A single model often has specific training biases or domain limitations, and cannot achieve comprehensive coverage when dealing with cross - domain, multi - task, and multi - type question - answering requests. When the user's needs exceed the knowledge scope that the model is good at, problems such as answer deviation, insufficient accuracy, response delay, etc. are likely to occur, which may even lead to a decline in the user experience. Secondly, existing systems lack the ability to real - time evaluate the application status of the model and intelligent scheduling, and cannot dynamically adjust the model selection according to the current adaptation degree, interaction performance, and user feedback of the model, resulting in the possible impact on the service quality due to model inadaptability during long - term interactions.

[0004] Most existing intelligent question - answering systems still rely on manual adjustment or static configuration in model optimization, and it is difficult to achieve adaptive optimization based on user behavior feedback and interaction process data. For question - answering requirements in different scenarios, the system lacks a flexible and efficient model switching and dynamic optimization mechanism, and it is difficult to ensure the continuous and efficient operation of the system in a changing environment. Therefore, the present invention provides an intelligent question - answering system based on a large - scale pre - trained model and an optimization method thereof to solve the above problems. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An intelligent question - answering optimization method based on a large - scale pre - trained model, comprising the following steps:

[0007] Preset a variety of large - scale pre - trained models to form a model resource pool, where each pre - trained model is used for different application scenarios, receive the question - answering request information input by the user for the first time, and generate an answer by the currently applied large - scale pre - trained model;

[0008] Based on the answer confidence information of the output of the large-scale pre-trained model for the current application, within the specified Q&A round window, extract the statistical features of the confidence and determine whether to trigger the dual adaptability evaluation mechanism;

[0009] When the dual adaptability evaluation mechanism is triggered, evaluate the model answer adaptation degree and the model interaction adaptation degree of the current large-scale pre-trained model, and classify the current large-scale pre-trained model into a highly adaptable type or a lowly adaptable type based on the dual evaluation results;

[0010] When the classification result is a lowly adaptable type, search for an optimized model from the model resource pool for switching based on the current large-scale pre-trained model and the dual evaluation results.

[0011] In a preferred embodiment, when receiving the Q&A request information input by the user for the first time, convert it into a request vector, and at the same time convert each large-scale pre-trained model in the model resource pool into an answer vector, select the answer vector with the highest similarity to the request vector, and the corresponding large-scale pre-trained model makes the first answer.

[0012] In a preferred embodiment, the statistical features of the confidence refer to:

[0013] Within the specified Q&A round window, obtain the answer confidence information for each round, and then calculate the confidence entropy value and the confidence skewness value respectively.

[0014] In a preferred embodiment, the logic for triggering the dual adaptability evaluation mechanism is:

[0015] Compare the confidence entropy value with the preset standard chaos degree threshold, and at the same time compare the confidence skewness value with the standard threshold interval composed of the preset negative threshold and positive threshold. If it satisfies that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, then do not trigger the dual adaptability evaluation mechanism. If it does not satisfy that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, then trigger the dual adaptability evaluation mechanism.

[0016] In a preferred embodiment, the evaluation of the model answer adaptation degree and the model interaction adaptation degree refers to:

[0017] Evaluate the model answer adaptation degree of the current large-scale pre-trained model to generate a model answer adaptation index; evaluate the model interaction adaptation degree of the current large-scale pre-trained model to generate a model interaction adaptation index.

[0018] In a preferred embodiment, the logic for obtaining the model answer adaptation index is:

[0019] Obtain the average value of the answer confidence within the specified Q&A round window and denote it as , obtain the entropy value of the answer confidence and denote it as , take the average value of the cosine similarity between the answer vector of each round and the historical dialogue summary vector as the consistency score and denote it as , map the implicit feedback based on user behavior to the average user feedback score within the range of [0, 1] and denote it as ;

[0020] To emphasize the interaction and coupling between indicators, design the interaction term of the confidence factor and the entropy factor :

[0021] ; is the adjustment coefficient, used to strengthen the influence of confidence, and its value range is [1, 3], is the preset entropy sensitivity coefficient, is the preset entropy inflection point threshold, which controls the occurrence of entropy inversion suppression, is the maximum entropy value preset for normalization;

[0022] Design the interaction term of consistency and feedback factor :

[0023] ; is the preset amplification coefficient, used to control the amplification multiple of the consistency score, is the preset user feedback influence factor;

[0024] Design the dynamic penalty term :

[0025] ; and are both preset volatility sensitivity coefficients and their sum is one, is the standard deviation of confidence volatility, is the average value of the confidence decline rate of adjacent round answers;

[0026] The calculation formula for the model answer adaptation index is:

[0027] ; is the model answer adaptation index.

[0028] In a preferred embodiment, the acquisition logic of the model interaction adaptation index is:

[0029] Within the specified Q&A round window, divide the total number of model inputs and outputs by the window duration to obtain the interaction activity , obtain the average value of the number of times each round answer references historical information to obtain the context retention degree , obtain the ratio of the number of topic changes to the total number of inputs for all rounds of inputs to obtain the change adaptability , obtain the time taken for each round of answering and calculate the average value to obtain the interaction response , the calculation formula for the model interaction adaptation index is:

[0030] ;

[0031] is a preset interaction activity adjustment factor, is a preset gating steepness adjustment factor, is a preset adaptive gating threshold, is a preset power exponent adjustment factor, is a preset response time adjustment factor, is an overall gain adjustment factor, is the comprehensive term of interaction complexity, which is calculated by the following formula: ; is the model interaction adaptation index.

[0032] In a preferred embodiment, based on the pre-trained convolutional neural network model, analyze the overall adaptation degree of the current large-scale pre-trained model based on the model answer adaptation index and the model interaction adaptation index, and divide it into a highly adapted type or a lowly adapted type.

[0033] In a preferred embodiment, when the division result is a lowly adapted type, obtain the number, the model answer adaptation index, and the model interaction adaptation index of the current large-scale pre-trained model, and use them as input variables of fuzzy logic. Use the numbers corresponding to the remaining models in the model resource pool as output variables of fuzzy logic. Perform fuzzy processing on the input variables to convert the values of the input variables into fuzzy sets. Perform fuzzy processing on the output variables to convert the output variables into fuzzy sets. Formulate fuzzy rules to describe the adaptation degrees of the remaining models in the model resource pool under different combinations of data types. Infer the fuzzy input variables through the fuzzy rules to obtain the number corresponding to the optimized model in the model resource pool.

[0034] In a preferred embodiment, an intelligent question-answering system based on a large-scale pre-trained model includes:

[0035] A model resource pool that stores multiple large-scale pre-trained models;

[0036] A confidence monitoring module that monitors the answer confidence information of the answers output by the large-scale pre-trained model currently in use;

[0037] The interaction anomaly detection module extracts the statistical features of confidence levels within the specified Q&A round window and determines whether to trigger the dual adaptability evaluation mechanism;

[0038] The adaptability evaluation module is used to evaluate the adaptability of the current large-scale pre-trained model in terms of model answer adaptation and model interaction adaptation when triggering the dual adaptability evaluation mechanism;

[0039] The type classification module classifies the current large-scale pre-trained model into a highly adaptable type or a lowly adaptable type based on the dual evaluation results;

[0040] The optimization switching module, when the classification result is a lowly adaptable type, searches for an optimized model from the model resource pool based on the current large-scale pre-trained model and the dual evaluation results for switching;

[0041] The context migration module is used to transfer the user conversation context during model switching to achieve seamless migration;

[0042] The Q&A execution module is used to perform intelligent Q&A interactions based on the current model.

[0043] The technical effects and advantages of the present invention:

[0044] In the present invention, multiple large-scale pre-trained models are preset in the model resource pool, and each model is specialized for different fields and task scenarios. The dynamic switching between professional models and general models is realized. When the user's needs involve specific industry knowledge, the system can intelligently identify and have the professional model provide the first-round answer, improving the accuracy and professionalism of the answer and avoiding the problem of insufficient professionalism caused by the general model due to its wide knowledge scope. If the adaptability of the professional model decreases due to reasons such as user topic switching, it can be optimized to the general model to provide a wide range of knowledge coverage and association for answering.

[0045] In the present invention, through the dual adaptability evaluation mechanism, the answer quality and interaction performance of the model are dynamically evaluated to achieve real-time monitoring of the model application status. The system automatically identifies the adaptability level of the current model according to the evaluation results, and combines with the fuzzy logic inference system to realize model switching, improving the system's adaptive ability to complex and changing user needs and ensuring efficient response in different user scenarios.

[0046] In the present invention, through the dynamic evaluation and intelligent scheduling mechanism, different large-scale pre-trained models are called on demand to avoid resource waste caused by a single model running at a high load for a long time. The professional model is put into use in professional scenarios, and the general model is used for general needs, realizing the reasonable allocation and scheduling of model resources, reducing the computing cost and energy consumption, and improving the overall operation efficiency of the system.

[0047] Through the dual evaluation of the answer adaptation index and the interaction adaptation index, the present invention can timely discover the problems of the decline in the model answer quality or abnormal interaction in real time, and timely switch to a better model to ensure the stability and reliability of the question-answering system.

[0048] The present invention can automatically identify the risks of model output, switch to an optimized model through dynamic adaptation evaluation, effectively reduce the risks of model hallucinations, irrelevant answers or non-compliant answers, and improve the overall reliability and business compliance of the question-answering system. Through a fuzzy logic inference system, the complex model adaptability evaluation data is intelligently fused with the resource pool model status information to realize the automation and intelligence of model switching decisions. The system does not need to rely on manual intervention or manual configuration, improves the intelligence level of the system, and reduces the operation and maintenance difficulty. Brief Description of the Drawings

[0049] For the convenience of those skilled in the art to understand, the present invention will be further described below in conjunction with the accompanying drawings;

[0050] Figure 1 It is the schematic diagram of the intelligent question-answering optimization method based on the large-scale pre-trained model in the present invention.

[0051] Figure 2 It is the schematic diagram of the intelligent question-answering system based on the large-scale pre-trained model in the present invention. Detailed Embodiments

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] Refer to Figure 1 - Figure 2 The following embodiments are obtained:

[0054] Embodiment 1:

[0055] With the rapid development of artificial intelligence technology, large-scale pre-trained models have made remarkable breakthroughs in the field of natural language processing and are widely used in intelligent question-answering systems. Existing large-scale pre-trained models have strong language understanding and generation capabilities through deep learning and big data training. However, the current intelligent question-answering systems generally rely on a single model to perform question-answering tasks, and the problems of insufficient model adaptability and poor flexibility are becoming increasingly prominent.

[0056] In practical applications, the Q&A requirements of different users are complex and variable, involving knowledge backgrounds in multiple fields and across scenarios. Due to the limitations of training data and the constraints of model structures, a single large-scale pre-trained model is difficult to comprehensively cover the knowledge systems of different application scenarios. When the model processes tasks beyond its advantageous fields, the accuracy, reliability, and interaction experience of the answers often cannot be effectively guaranteed, and even situations such as model hallucinations or irrelevant answers may occur. In addition, most existing systems lack a real-time dynamic evaluation and optimization mechanism for model adaptability, resulting in low utilization rate of model resources and the inability to adaptively optimize the overall performance of the system.

[0057] There are also problems in the current technology of insufficient evaluation of model adaptability and interaction ability, lacking effective evaluation indicators and switching mechanisms, and being unable to dynamically determine and adjust the usage status of large-scale pre-trained models, which further limits the stability and accuracy of intelligent Q&A systems.

[0058] The present invention aims to solve the technical problems of the single large-scale pre-trained model, insufficient adaptability, and lack of dynamic optimization and scheduling mechanism in existing intelligent Q&A systems, and improve the flexibility and intelligence level of intelligent Q&A systems in multiple scenarios.

[0059] Through the intelligent Q&A optimization method based on large-scale pre-trained models proposed by the present invention, the system presets multiple large-scale pre-trained models to form a model resource pool, and selects the optimal model to execute Q&A tasks for different application scenarios. When the system receives the Q&A request information input by the user for the first time, it adopts a similarity matching mechanism between the request vector and the answer vectors of each large-scale pre-trained model in the model resource pool to achieve the optimal selection of the first answer model, and improve the accuracy and efficiency of the first Q&A.

[0060] The present invention further introduces statistical feature analysis based on answer confidence information, and adopts the joint determination of confidence entropy value and confidence skewness value to dynamically judge whether it is necessary to trigger a dual adaptability evaluation mechanism. Under the dual adaptability evaluation mechanism, the system evaluates the model answer adaptation degree and the model interaction adaptation degree of the current model, and generates a model answer adaptation index and a model interaction adaptation index respectively, providing refined and comprehensive data support for model adaptability analysis.

[0061] Through a pre-trained convolutional neural network model, comprehensively analyze the model answer adaptation index and the model interaction adaptation index to judge the overall adaptation degree of the current large-scale pre-trained model, and dynamically divide it into a highly adapted type or a lowly adapted type. For models of the lowly adapted type, the system, based on a fuzzy logic reasoning mechanism, combines the current model number, the dual evaluation results, and the model resource pool information to accurately reason and switch to the optimal model, achieving seamless and dynamic optimization switching, and improving the overall response efficiency and interaction experience of the system. Through a dynamic evaluation mechanism and an intelligent switching method, the present invention significantly improves the adaptability, interactivity, and self-optimization ability of large-scale pre-trained models in intelligent question-and-answer systems, enhances the comprehensive performance of the system in processing question-and-answer requirements in multiple scenarios and domains, and has important application value and industrial promotion prospects.

[0062] Specifically, the present invention proposes an intelligent question-and-answer optimization method based on a large-scale pre-trained model, including the following steps:

[0063] Preset a variety of large-scale pre-trained models to form a model resource pool. Each pre-trained model is used for different application scenarios. Receive the question-and-answer request information input by the user for the first time, and generate an answer by the large-scale pre-trained model currently in use. In this step, by preloading and deploying a variety of large-scale pre-trained models, a model resource pool is constructed, and dedicated models are configured for different application fields or scenarios (such as general question-and-answer, etc.). Each model forms a highly targeted and domain-adapted intelligent question-and-answer ability according to its training data source and algorithm optimization strategy.

[0064] When receiving the question-and-answer request information input by the user for the first time, first preprocess the request information, and use natural language understanding (NLU) technology to convert the text input by the user into a request vector. At the same time, convert each pre-trained model in the model resource pool into an answer vector, and compare the request vector with the answer vector based on a semantic similarity algorithm (such as cosine similarity, vector distance measure, etc.) to select the optimal matching model to achieve the first question-and-answer response. The selected large-scale pre-trained model generates and outputs an answer immediately without relying on additional evaluation, ensuring the real-time and high efficiency of the user's question-and-answer experience.

[0065] By presetting multiple models, the limitation of traditional intelligent question-and-answer systems relying on a single model is broken, and the coverage breadth and domain accuracy of the overall question-and-answer ability of the system are significantly improved. The use of the request vector and answer vector matching mechanism makes the model selection real-time and intelligent, ensuring that the system can achieve high-quality answers in the first interaction stage, enhancing the user experience. It lays a foundation for subsequent model dynamic optimization and switching, and improves the flexibility and scalability of the system.

[0066] Based on the answer confidence information output by the large-scale pre-trained model of the current application, within the specified Q&A round window, extract the statistical features of the confidence and determine whether to trigger the dual adaptability evaluation mechanism; while the large-scale pre-trained model of the current application generates an answer, the system obtains the answer confidence information output by the model, and the confidence information can be obtained based on data such as the model output probability distribution, classification decision boundary, or token probability weights of the generative model. The system continuously collects the answer confidence information within the specified Q&A round window (such as the most recent N rounds), and extracts key statistical features from it, including but not limited to the confidence entropy value and the confidence skewness value. The system compares the confidence entropy value of the current window with the preset standard chaos degree threshold, and at the same time compares the confidence skewness value with the preset positive and negative threshold intervals to form a dual judgment mechanism. If the confidence entropy value is less than or equal to the standard chaos degree threshold, and the confidence skewness value falls within the standard threshold interval, it is considered that the current model adaptability is normal and the evaluation mechanism is not triggered; otherwise, the subsequent dual adaptability evaluation process is triggered.

[0067] By dynamically monitoring the change of model confidence through a dynamic window, capturing potential problems of model stability and adaptability in real time, adopting a joint determination method of confidence entropy and skewness, improving the accuracy and robustness of judgment, avoiding misjudgment by a single index, constructing a dynamic adaptability evaluation trigger mechanism, enabling the system to have self-monitoring and adaptive optimization capabilities, and reducing the risk of the system misusing an inadaptable model.

[0068] When the dual adaptability evaluation mechanism is triggered, evaluate the model answer adaptation degree and the model interaction adaptation degree of the current large-scale pre-trained model, and classify the current large-scale pre-trained model into a highly adaptable type or a lowly adaptable type based on the dual evaluation results; when the dual adaptability evaluation mechanism is triggered, the system respectively evaluates the model answer adaptation degree and the model interaction adaptation degree of the current application model.

[0069] The evaluation of the model answer adaptation degree is mainly based on the model output data within the specified Q&A instance window, comprehensively analyzing information such as the mean value of the model answer confidence, the confidence entropy value, the consistency score (such as the average cosine similarity of the answer vector and the historical dialogue summary vector), and the implicit feedback score of user behavior, and using a non-linear interaction calculation model to calculate the model answer adaptation index.

[0070] The evaluation of the model interaction adaptation degree calculates the model interaction adaptation index by analyzing the interaction activity (total number of model input and output times / window duration), context retention degree (average value of the number of times of answering and citing historical information), change adaptability (ratio of the number of topic changes to the total number of inputs), and interaction response degree (average answering time). The system comprehensively analyzes the model answer adaptation index and the model interaction adaptation index through a convolutional neural network model to judge the overall adaptation degree of the current large-scale pre-trained model, and classifies it into a "highly adapted type" or a "lowly adapted type".

[0071] The dual evaluation mechanism comprehensively measures the static answer ability and dynamic interaction adaptation ability of the model, improves the comprehensiveness and scientificity of model selection, integrates the convolutional neural network model to realize intelligent classification of evaluation results, improves the evaluation accuracy and the intelligent level of model switching decision-making, quickly identifies the insufficient model adaptability, and ensures the stability and reliability of the system interaction experience.

[0072] When the classification result is the lowly adapted type, an optimized model is searched for and switched from the model resource pool based on the current large-scale pre-trained model and the dual evaluation results. When the current large-scale pre-trained model is determined to be of the lowly adapted type, the system takes the model number, model answer adaptation index, model interaction adaptation index and other evaluation data as input variables and introduces them into the fuzzy logic inference system, and combines the adaptation status numbers of the remaining models in the model resource pool as output variables to complete the fuzzification process of input and output.

[0073] The system infers the fuzzy set according to the preset fuzzy rules (describing the adaptation degree and switching strategies of each model under different input data combinations) to obtain the number corresponding to the optimal optimized model. Subsequently, the system switches from the model resource pool to the optimized model and realizes the seamless migration of user dialogue data and context information to ensure that the user experience is not interrupted.

[0074] Through the fuzzy logic inference model, the fuzzy processing and flexible control of the model optimization switching decision are realized, the flexibility and adaptability of the switching strategy are improved, the switching process is based on dynamic evaluation data, the current model running state and system requirements are fully considered, the rationality and efficiency of the switching behavior are ensured, and the seamless switching technology ensures the continuity of the user question-answering process and avoids the problems of lost dialogue context or system response interruption caused by model switching.

[0075] In the present invention, the system presets a variety of large-scale pre-trained models based on the requirements of different application scenarios to form a diversified model resource pool. Each large-scale pre-trained model in the model resource pool has been subjected to targeted training and optimization and is respectively used to handle user question-answering requests in different fields and of different types.

[0076] For example: General domain models such as GPT series (GPT-3, GPT-4) are mainly applicable to open-domain question answering, encyclopedia knowledge-based question answering, daily conversation generation, etc., and have a large-scale knowledge graph and strong generalization ability.

[0077] Medical domain models such as BioGPT, Med-PaLM (Google), ChatDoctor, etc. are mainly applied to medical literature retrieval, clinical diagnosis assistance, health consultation, etc., and can understand and generate professional text content involving medical terms, clinical pathways, etc. When the user input involves symptom description, medical terms or health consultation, the medical domain model is preferentially matched to generate a professional answer.

[0078] Legal domain models such as LawGPT, ChatLaw, etc. are used in scenarios such as legal regulations query, legal consultation, contract review, etc., and have the ability of legal text understanding, article citation and legal principle analysis. When the user's question involves contracts, legal consultation, case description, the system will preferentially dispatch the legal domain model to answer.

[0079] Financial domain models such as FinGPT, BloombergGPT, etc. are used in scenarios such as financial news interpretation, investment advice, financial analysis, etc., and have the ability to analyze financial data, generate reports and predict market trends. When the user asks questions related to the stock market, financial management, and financial policies, the financial domain model is automatically matched to provide a professional answer.

[0080] Educational and learning domain models such as EduGPT, ChatEDU, etc. are used in scenarios such as intelligent education tutoring, question explanation, learning plan formulation, etc., and can achieve multi-disciplinary knowledge explanation and adapt to the educational needs of different grades. When facing the user's homework tutoring and course explanation needs, the educational domain model is preferentially enabled.

[0081] When receiving the first input of the question-and-answer request information from the user, it is converted into a request vector. At the same time, each large-scale pre-trained model in the model resource pool is converted into an answer vector, and the large-scale pre-trained model corresponding to the answer vector with the highest similarity to the request vector is selected for the first answer. After the user first inputs the question-and-answer request information, the system performs semantic vectorization processing on the input content through natural language understanding (NLU) and intent recognition to form a request vector.

[0082] Each large-scale pre-trained model in the model resource pool is pre-set with its answer vector, which represents the model's domain coverage ability and answer characteristics. By comparing the request vector with the answer vector (such as calculating the cosine similarity), the system filters out the most matching model, ensuring that the most domain-appropriate model is used in the first round of Q&A, thereby improving the accuracy and professionalism of the answer. For example: When the user inputs: "I've had a cold recently and keep coughing. What medicine should I take?", the system recognizes that it involves the medical and health field. After vector matching, it automatically selects ChatDoctor as the large-scale pre-trained model for the current application to generate a professional health guidance answer. When the user inputs: "What legal norms should the articles of association follow?", the system recognizes it as a legal consultation question and matches it to the LawGPT model for the first round of Q&A to generate relevant legal provisions and explanations.

[0083] The statistical features of confidence refer to: within the specified Q&A round window, obtain the confidence information of each round of answers, and then calculate the confidence entropy value and the confidence skewness value respectively. The logic for triggering the dual adaptability evaluation mechanism is: compare the confidence entropy value with the preset standard chaos degree threshold, and at the same time compare the confidence skewness value with the standard threshold interval composed of the preset negative threshold and positive threshold. If it satisfies that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, the dual adaptability evaluation mechanism is not triggered. If it does not satisfy that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, the dual adaptability evaluation mechanism is triggered.

[0084] The present invention realizes the dynamic evaluation of the answer stability and confidence distribution state of the current large-scale pre-trained model by extracting the answer confidence information output by the model within the specified Q&A round window and calculating the confidence entropy value and the confidence skewness value respectively. The confidence entropy value reflects the chaos degree of the model output result and can effectively measure the answer uncertainty of the model within the current interaction window. The confidence skewness value reveals the skewness of the confidence distribution and identifies the answer tendency and consistency level of the model. Through the combined analysis of the two indicators, it provides the system with a fine-grained model state perception ability and enhances the real-time early warning ability for potential model failure risks.

[0085] By jointly comparing the confidence entropy value with the preset standard chaos degree threshold and the confidence skewness value with the standard threshold interval, a dynamic monitoring and triggering mechanism in the intelligent Q&A system is constructed. When it is detected that the confidence characteristics of the model output do not meet the preset stability requirements, the system automatically triggers the dual adaptability evaluation mechanism, thereby comprehensively evaluating the answer adaptability and interaction adaptability of the current model. This intelligent determination logic based on statistical features effectively avoids the model being in a low adaptability state for a long time, improves the intelligent level of system model switching and optimization, and ensures the stability of the Q&A process and the coherence of the user interaction experience.

[0086] Performing model answer adaptation degree evaluation and model interaction adaptation degree evaluation refers to:

[0087] Performing model answer adaptation degree evaluation on the current large-scale pre-trained model to generate a model answer adaptation index; performing model interaction adaptation degree evaluation on the current large-scale pre-trained model to generate a model interaction adaptation index.

[0088] The acquisition logic of the model answer adaptation index is as follows:

[0089] Obtain the mean answer confidence within the specified Q&A round window and denote it as , reflecting the average confidence level of the model's answers; obtain the answer confidence entropy value and denote it as , measuring the degree of chaos in the confidence distribution of the model's answers; the average value of the cosine similarity between the answer vector of each round and the historical dialogue summary vector is used as the consistency score and denoted as , measuring whether the answer is consistent with the historical context; the average score of the implicit feedback based on user behavior mapped to the range [0, 1] is denoted as , reflecting the implicit satisfaction or acceptance of the user with the model's answers, which can be obtained by taking the average value after assigning values according to various user behavior indicators.

[0090] To emphasize the interaction and coupling between indicators, design the interaction term of the confidence factor and the entropy factor :

[0091] ; is a regulation coefficient used to strengthen the influence of confidence, with a value range of [1, 3]. The larger the value, the stronger the influence of confidence; is a preset entropy sensitivity coefficient, which determines the influence intensity of the answer confidence entropy value on the denominator exponential function. The larger the value, the more sensitive to the change of entropy. is a preset entropy inflection point threshold, which determines when the entropy value starts to produce an obvious inhibitory effect and controls the occurrence time of the adaptation inhibition of the model when the entropy value is high. is a preset maximum entropy value for normalization; controls the enhancement effect of high confidence, enabling the model to give a positive evaluation to high-confidence answers. The denominator adopts the Sigmoid function form to dynamically adjust the inhibitory effect of the entropy value on the overall score. The higher the entropy value, the more dispersed the answer distribution and the greater the uncertainty, resulting in a decrease in adaptability. Through and flexibly set the influence interval and intensity of the entropy value, so that when the entropy value is too large, the model adaptability is significantly reduced, reasonably suppressing the "hallucinations" or low-confidence answers that the model may generate.

[0092] Design the interaction term of consistency and feedback factor :

[0093] ; is a preset amplification factor, which is used to control the amplification multiple of the consistency score, adjust the gain intensity of the consistency score in the overall score, and control the influence of consistency on the model fitness. is a preset user feedback influence factor, which adjusts the weight of the average user feedback score on the overall score and controls the adjustment effect of user feedback on the model fitness; Amplify the influence of consistency on model evaluation to ensure that the model answers have good context coherence. Introduce and is controlled by to control its influence intensity, reflecting the key role of user behavior feedback in model evaluation. Use a logarithmic function to balance the overall output value, avoid "overflow" of evaluation caused by too high consistency or feedback score, and ensure the stability and rationality of system scoring.

[0094] Design a dynamic penalty term , which is used to reflect the negative impact of model stability problems:

[0095] ; and are both preset volatility sensitivity coefficients and the sum of the two is one. Adjust 's contribution to the dynamic penalty term, determining the degree of influence of volatility on the overall fitness index. Adjust 's contribution to the dynamic penalty term, determining the inhibitory strength of the sudden drop in confidence on the fitness index. is the standard deviation of confidence volatility, which measures the stability of the model answer confidence. The larger the value, the greater the answer volatility. is the average value of the confidence decline rate of adjacent round answers, which measures the continuity of the model confidence. A large value indicates that the model performance is unstable and prone to a sudden drop in confidence; the dynamic penalty term directly reflects the volatility and continuity problems of the model answer confidence.

[0096] The combined effect of volatility and decline rate: Simultaneously examine the volatility of confidence and the average decline rate to achieve double punishment for the "jitter" and "collapse" performance of the model. Through and control the ratio of the two, flexibly adjust the negative impact on the system score, and prevent system risks brought by poor model stability.

[0097] The calculation formula for the model answer fitness index is:

[0098] ; It is the Model Answer Adaptation Index. The larger the MAAI, the better the performance of the current large-scale pre-trained model in aspects such as answer confidence, confidence distribution stability, answer consistency, and user feedback satisfaction. It also has high stability, good answer continuity, and no obvious confidence fluctuations or sudden drops. The current answer ability of the model highly matches the interaction requirements, has good question understanding ability and context continuity, positive user feedback, high satisfaction shown by implicit or explicit user behaviors, strong interaction fluency, and the system can continue to maintain the current model to process Q&A requests without switching for optimization.

[0099] The acquisition logic of the model interaction adaptation index is as follows:

[0100] Within the specified Q&A round window, divide the total number of model inputs and outputs by the window duration to obtain the interaction activity. , obtain the average value of the number of times of citing historical information in each round of answers to get the context retention. , obtain the ratio of the number of topic changes in all rounds of inputs to the total number of inputs to get the change adaptability. , obtain the average value of the time taken for each round of answers to get the interaction responsiveness. , the calculation formula of the model interaction adaptation index is:

[0101] ;

[0102] is the preset interaction activity adjustment factor, is the preset gating steepness adjustment factor, is the preset adaptability gating threshold, is the preset power index adjustment factor, is the preset response time adjustment factor, is the overall gain adjustment factor, is the comprehensive item of interaction complexity, which is calculated by the following formula: ; is the model interaction adaptation index. The calculation formula structure of the model interaction adaptation index is fractional nesting + power adjustment + logarithmic gain, which reflects the comprehensive evaluation and dynamic regulation ability of multi-dimensional indicators. It comprehensively combines key indicators such as interaction activity (IA), context retention (CR), change adaptability (AD), and interaction responsiveness (RS) in a non-linear way, not only measuring the overall interaction performance of the model but also achieving dynamic adjustment and balance.

[0103] In the first part, the power adjustment of IA: Strengthen the contribution of the power of IA to the model's interaction adaptability, indicating that frequent interaction between the user and the model is one of the signs of the model's good state. The square root of CR: Smooth the context retention degree to prevent it from affecting the overall score due to excessive fluctuations, ensuring that the model has continuous context understanding and response capabilities. The exponential amplification of AD: The change adaptability determines the model's response ability to topic switching and changes in user intentions. Through the exponential function, the sensitivity to dialogue changes is enhanced rapidly. The Sigmoid gating mechanism (denominator part) : Control the score suppression or excitation of the model at different stages of change adaptability. When AD exceeds the adaptability gating threshold , the suppression effect on the denominator is significantly weakened, encouraging well-adapted models to obtain higher scores. The overall power adjustment: Control the weight of the overall score, enabling the system to flexibly adjust the intensity of adaptability evaluation in different scenarios.

[0104] In the second part, the comprehensive item of interaction complexity: Measure the overall interaction complexity and system load status, integrating three core indicators: activity, context retention, and adaptability, improving the comprehensiveness and accuracy of evaluation. The interaction response degree regulation item : The shorter the response time, the smaller RS, the smaller the denominator, and the greater the overall gain, reflecting the advantage of the model's rapid response. The logarithmic function smooth growth: Avoid exponential explosion of scores, keep the scores growing smoothly, reflecting the improvement of adaptability brought by the optimization of the model's interaction performance.

[0105] The adjustment mechanism: Different adjustment parameters such as the interaction activity adjustment factor and the gating steepness adjustment factor respectively adjust the sensitivity, gating threshold, and score gain of different sub-items and the overall score, making the model evaluation adjustable, flexible, and adaptable, and adapting to the needs of different types of question-and-answer systems and business scenarios. The larger the MIAI, the higher the interaction adaptability of the current model in this application scenario. The model has strong long-term interaction stability, the ability to adapt to the conversation context, and brings a good user experience, which can avoid user loss or dialogue interruption caused by insufficient model adaptability and improve the overall service quality of the system. The evaluation results can be directly used for dynamic decision-making to determine whether to continue the interaction of the current model or switch to a better model.

[0106] Through the pre-trained convolutional neural network model, analyze the overall adaptation degree of the current large-scale pre-trained model based on the model answer adaptation index and the model interaction adaptation index, and classify it into a highly adapted type or a lowly adapted type.

[0107] Through a pre-trained convolutional neural network model, the overall adaptation degree of the current large-scale pre-trained model is intelligently analyzed. This convolutional neural network model takes the model answer adaptation index and the model interaction adaptation index as input features, and through multi-layer convolution and feature extraction operations, automatically identifies and analyzes the non-linear relationship and coupling characteristics between different indexes. Local features are extracted through the convolutional layer, and the feature dimension is reduced by combining with the pooling layer. Further, feature fusion and comprehensive evaluation are realized in the fully connected layer, and finally, the comprehensive adaptability determination result of the current large-scale pre-trained model is output.

[0108] Based on the evaluation results of the convolutional neural network model, the system classifies the current large-scale pre-trained model into a highly adapted type or a lowly adapted type. The highly adapted type indicates that the current model meets the system application requirements in terms of answer ability, interaction performance, and overall stability, and is suitable for continuing to maintain the question-and-answer interaction; the lowly adapted type indicates that there is a lack of adaptability in the current model, and an optimized model needs to be selected for switching according to the evaluation results. Through this method, accurate classification and dynamic optimization of model adaptability are achieved, improving the overall question-and-answer quality of the system and the user interaction experience.

[0109] When the classification result is the lowly adapted type, obtain the number of the current large-scale pre-trained model, the model answer adaptation index, and the model interaction adaptation index, and use them together as the input variables of the fuzzy logic. Use the numbers corresponding to the remaining models in the model resource pool as the output variables of the fuzzy logic. Perform fuzzy processing on the input variables to convert the values of the input variables into fuzzy sets, perform fuzzy processing on the output variables to convert the output variables into fuzzy sets, formulate fuzzy rules to describe the adaptation degrees of the remaining models in the model resource pool under different combinations of data types, and infer the fuzzy input variables through the fuzzy rules to obtain the number corresponding to the optimized model in the model resource pool.

[0110] In the present invention, the input variables include the number of the current large-scale pre-trained model, the model answer adaptation index, and the model interaction adaptation index. The system first needs to fuzzify these input values for subsequent fuzzy inference processing. Fuzzification is to convert the exact numerical values into linguistic values (such as "high", "medium", "low"), and define the corresponding membership functions for each linguistic value. The membership function is used to determine the membership degree of the input value belonging to a certain linguistic value. Commonly used membership functions include trigonometric functions, trapezoidal functions, and Gaussian functions, etc. The present invention preferably uses the triangular membership function because of its simple structure, high calculation efficiency, and suitability for practical applications.

[0111] Example of triangular membership function design: Taking the model answer adaptation index as an example, it is divided into three linguistic values: "low", "medium", and "high". The triangular membership function can be defined as follows: The peak of the "low" membership function is located at 20%, and the bottom range is from 0 to 40%. The peak of the "medium" membership function is located at 50%, and the bottom range is from 20% to 80%. The peak of the "high" membership function is located at 80%, and the bottom range is from 50% to 100%. Through the membership function, the system can determine the membership degree of a certain actual input value under different linguistic values. For example, if the model answer adaptation index is 55%, the membership degree under the "medium" linguistic value is 0.8, and the membership degree under the "high" linguistic value is 0.2.

[0112] After the input variables are fuzzified, a fuzzy set is formed as the input condition for the subsequent fuzzy inference engine. The output variable is the association information between the numbers of candidate models in the model resource pool and the adaptation level. These numbers also need to be transformed into fuzzy sets through the membership function. Common linguistic values include "preferred recommendation", "considerable", "not recommended for the time being", etc. For each model, the membership degree is preset according to indicators such as its historical performance, domain adaptation ability, and interaction responsiveness. For example, a certain model has performed well in multiple historical applications, and the system sets a high membership degree of "preferred recommendation" for it.

[0113] Fuzzy rules embody the empirical knowledge and decision-making logic of the system. For example: Rule 1: "If the model answer adaptation index is low and the model interaction adaptation index is low, then switch to the preferred recommendation model." Rule 2: "If the model answer adaptation index is medium and the model interaction adaptation index is high, then keep the current model or consider the preferred recommendation model." Rule 3: "If the model answer adaptation index is high and the model interaction adaptation index is high, then keep the current model." The rule base can be expanded and optimized according to actual application requirements to improve the inference ability of the system through continuous training.

[0114] The fuzzy inference engine makes a comprehensive judgment on the input fuzzy set based on the fuzzy rule base, using common maximum-minimum inference method or weighted average method. In the present invention, the maximum-minimum inference method is preferably used. Its basic principle is to determine the rule activation strength according to the matching degree between the input condition and the premise part of the rule, and then generate the output fuzzy set according to the conclusion part of the rule. The system synthesizes the output results of all activated rules to obtain the fuzzy set of the output variable, representing the adaptation level and recommendation degree of all candidate models.

[0115] Defuzzification is to convert the output fuzzy set into a specific decision value, that is, to determine which optimization model should be finally switched to. Common methods include the centroid method, the maximum membership degree method, etc. The present invention preferably adopts the maximum membership degree method, that is, selects the candidate model with the highest membership degree as the final output. For example, if the membership degree of "Model Three" in the "Preferred Recommendation" set is the largest, the system finally selects "Model Three" as the current optimization model to complete the model switching operation.

[0116] Example 1: Low answer adaptation and low interaction adaptation, recommend high adaptation model:

[0117] Scenario description: The current large-scale pre-trained model is numbered "Model One". The system detects that the model answer adaptation index is 35% and the model interaction adaptation index is 40%. According to the preset membership function: the answer adaptation index of 35% corresponds to the "low" level with a membership degree of 0.8. The interaction adaptation index of 40% corresponds to the "low" level with a membership degree of 0.7. After the input variables are fuzzified, they form a "low-low" state. The system searches for rules that meet the input conditions: If the model answer adaptation index is low and the model interaction adaptation index is low, then switch to the "Preferred Recommendation" model. The model resource pool contains four models, namely "Model Two", "Model Three", "Model Four", and "Model Five". The membership degree of "Model Three" in the "Preferred Recommendation" state is 0.9, and the membership degrees of the other models are all lower than 0.5. According to the maximum membership degree method, select the "Model Three" with the highest membership degree. The system finally selects "Model Three" as the number of the optimization model to complete the switching.

[0118] Example 2: Medium answer adaptation and low interaction adaptation, recommend comprehensive balance model:

[0119] Scenario description: The current large-scale pre-trained model is numbered "Model Four". The system detects that the model answer adaptation index is 60% and the model interaction adaptation index is 35%. According to the membership function mapping: the answer adaptation index of 60% corresponds to the "medium" level with a membership degree of 0.6. The interaction adaptation index of 35% corresponds to the "low" level with a membership degree of 0.75. The system searches for rules: If the model answer adaptation index is medium and the model interaction adaptation index is low, then switch to the "Considerable" model or the "Preferred Recommendation" model. The model status in the resource pool is as follows: The membership degree of "Model Two" in the "Considerable" state is 0.85, and the membership degree of "Model Three" in the "Preferred Recommendation" state is 0.7. According to the rule, preferentially match the "Considerable" model and select the "Model Two" with the highest membership degree. The system determines that "Model Two" is the number of the optimization model.

[0120] Embodiment 2: An intelligent question-answering system based on a large-scale pre-trained model, including:

[0121] A model resource pool that stores multiple large-scale pre-trained models;

[0122] A confidence monitoring module that monitors the confidence information of the answers output by the large-scale pre-trained model currently in use;

[0123] An interaction anomaly detection module that extracts statistical features of the confidence level within a specified Q&A round window and determines whether to trigger a dual adaptability evaluation mechanism;

[0124] An adaptability evaluation module that, when the dual adaptability evaluation mechanism is triggered, evaluates the adaptability of the model's answers and the adaptability of the model's interaction for the current large-scale pre-trained model;

[0125] A type classification module that classifies the current large-scale pre-trained model into a highly adaptable type or a lowly adaptable type based on the dual evaluation results;

[0126] An optimization switching module that, when the classification result is a lowly adaptable type, searches for an optimized model from the model resource pool for switching based on the current large-scale pre-trained model and the dual evaluation results;

[0127] A context transfer module that is used to transfer the user's conversation context during model switching to achieve seamless migration;

[0128] A Q&A execution module that is used to perform intelligent Q&A interactions based on the current model.

[0129] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] It should be understood that in various embodiments of the present application, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0131] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0132] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0133] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or replacements, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

Claims

1. An intelligent question-answering optimization method based on a large-scale pre-trained model, characterized in that, It includes the following steps: Preset a variety of large-scale pre-trained models to form a model resource pool. Each pre-trained model is used for different application scenarios. Receive the question-and-answer request information input by the user for the first time, and generate an answer by the large-scale pre-trained model of the current application; Based on the answer confidence information of the answer output by the large-scale pre-trained model of the current application, within the specified question-and-answer round window, extract the statistical features of the confidence and judge whether to trigger the dual adaptability evaluation mechanism; When the dual adaptability evaluation mechanism is triggered, evaluate the model answer adaptability degree and the model interaction adaptability degree of the current large-scale pre-trained model, and classify the current large-scale pre-trained model into a highly adaptable type or a lowly adaptable type based on the dual evaluation results; When the classification result is a lowly adaptable type, search for an optimized model from the model resource pool for switching based on the current large-scale pre-trained model and the dual evaluation results.

2. The intelligent question-answering optimization method based on a large-scale pre-trained model according to claim 1, wherein, When receiving the question-and-answer request information input by the user for the first time, convert it into a request vector. At the same time, convert each large-scale pre-trained model in the model resource pool into an answer vector, select the answer vector with the highest similarity to the request vector, and the corresponding large-scale pre-trained model makes the first answer.

3. The intelligent question-answering optimization method based on a large-scale pre-trained model according to claim 2, wherein The statistical features of the confidence refer to: Within the specified question-and-answer round window, obtain the answer confidence information of each round, and then calculate the confidence entropy value and the confidence skewness value respectively.

4. The intelligent question-answering optimization method based on a large-scale pre-trained model according to claim 3, characterized in that The logic for triggering the dual adaptability evaluation mechanism is: Compare the confidence entropy value with the preset standard chaos degree threshold, and at the same time compare the confidence skewness value with the standard threshold interval composed of the preset negative threshold and positive threshold. If it satisfies that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, then the dual adaptability evaluation mechanism is not triggered. If it does not satisfy that the confidence entropy value is less than or equal to the preset standard chaos degree threshold and the confidence skewness value falls within the standard threshold interval, then the dual adaptability evaluation mechanism is triggered.

5. The intelligent question and answer optimization method based on a large-scale pre-trained model according to claim 4, wherein Conducting the model answer adaptability degree evaluation and the model interaction adaptability degree evaluation refers to: Evaluate the model answer adaptability degree of the current large-scale pre-trained model to generate a model answer adaptability index; evaluate the model interaction adaptability degree of the current large-scale pre-trained model to generate a model interaction adaptability index.

6. The intelligent question and answer optimization method based on a large-scale pre-trained model according to claim 5, wherein, The obtaining logic of the model answer adaptability index is: Obtain the average value of the answer confidence within the specified Q&A round window and denote it as , obtain the entropy value of the answer confidence and denote it as , use the average value of the cosine similarity between the answer vector of each round and the historical dialogue summary vector as the consistency score and denote it as , map the implicit feedback based on user behavior to the average user feedback score within the range of [0, 1] and denote it as ; To emphasize the interaction and coupling between indicators, an interaction term of the confidence factor and the entropy factor is designed : ; is a regulation coefficient, used to strengthen the influence of confidence, with a value range of [1, 3], is a preset entropy sensitivity coefficient, is a preset entropy inflection point threshold, which controls the occurrence of entropy inversion inhibition, is the maximum entropy value preset for normalization; Design consistency and feedback factor interaction term : ; is a preset amplification factor used to control the amplification multiple of the consistency score, is a preset user feedback influence factor; Design dynamic penalty term : ; and are both preset volatility sensitivity coefficients and the sum of the two is one, is the confidence volatility standard deviation, is the average value of the confidence decline rate of adjacent round responses; The formula for calculating the model answer adaptability index is: ; is the adaptation index for model answers.

7. The intelligent question-answering optimization method based on a large-scale pre-trained model according to claim 6, characterized in that, The obtaining logic of the model interaction adaptability index is: Within the specified Q&A round window, divide the total number of times the model inputs and outputs by the window duration to obtain the interaction activity level , obtain the number of times the historical information of each round of answers is cited and calculate the average value to obtain the context retention degree , obtain the ratio of the number of topic changes in all rounds of inputs to the total number of inputs to obtain the change adaptability , obtain the time taken for each round of answers and calculate the average value to obtain the interaction response degree , the calculation formula for the model interaction adaptation index is: ； is a preset interaction activity adjustment factor, is a preset gating steepness adjustment factor, is a preset adaptive gating threshold, is a preset power exponent adjustment factor, is a preset response time adjustment factor, is an overall gain adjustment factor, is a comprehensive term of interaction complexity, calculated by the following formula: ; is the model interaction adaptation index.

8. The intelligent question-answering optimization method based on a large-scale pre-trained model according to claim 7, characterized in that Through the pre-trained convolutional neural network model, analyze the overall adaptability degree of the current large-scale pre-trained model based on the model answer adaptability index and the model interaction adaptability index, and classify it into a highly adaptable type or a lowly adaptable type.

9. The intelligent question and answer optimization method based on a large-scale pre-trained model according to claim 8, characterized in that, When the classification result is of the low - adaptation type, obtain the number of the current large - scale pre - trained model, the model answer adaptation index, and the model interaction adaptation index, and use them together as the input variables of fuzzy logic. Use the numbers corresponding to the remaining models in the model resource pool as the output variables of fuzzy logic. Perform fuzzy processing on the input variables, convert the values of the input variables into fuzzy sets, perform fuzzy processing on the output variables, and convert the output variables into fuzzy sets. Formulate fuzzy rules to describe the adaptation degrees of the remaining models in the model resource pool under different combinations of data types. Infer the numbers corresponding to the optimized models in the model resource pool by applying the fuzzy rules to the fuzzified input variables.

10. An intelligent question-answering system based on a large-scale pre-trained model, which is used to implement the intelligent question-answering optimization method based on a large-scale pre-trained model according to any one of claims 1-9, characterized in that, Including: A model resource pool that stores multiple large - scale pre - trained models; A confidence monitoring module that monitors the answer confidence information of the answers output by the currently applied large - scale pre - trained model; An interaction anomaly detection module that extracts the statistical features of confidence within a specified Q&A round window and determines whether to trigger the dual - adaptation evaluation mechanism; An adaptation evaluation module that, when triggering the dual - adaptation evaluation mechanism, evaluates the adaptation degree of the model answer and the adaptation degree of model interaction of the current large - scale pre - trained model; A type classification module that classifies the current large - scale pre - trained model into a high - adaptation type or a low - adaptation type based on the dual - evaluation results; An optimization switching module that, when the classification result is of the low - adaptation type, searches for an optimized model from the model resource pool for switching based on the current large - scale pre - trained model and the dual - evaluation results; A context migration module that is used to transfer the user dialogue context during model switching to achieve seamless migration; A Q&A execution module that is used to perform intelligent Q&A interactions based on the current model.

Images