Interactive question answering system and question answering method based on multi-model parallel inference
An interactive question-answering system using multi-model parallel reasoning combines user preference-adjusted weight calculation and semantic similarity calculation to generate high-confidence answers and provide visual traceability. This solves the problems of unstable output from large models and difficulty in user trust, thus improving user experience and efficiency.
Patent Information
- Application Number
- PCT/CN2025/098748
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2025-06-03
- Publication Date
- 2026-11-12
AI Technical Summary
Existing large models exhibit significant output fluctuations across different application scenarios, making them difficult for users to trust and time-consuming to evaluate. Current integration methods cannot effectively combine user needs for answer integration and tracing.
An interactive question-answering system based on multi-model parallel reasoning is adopted. Combining pre-reasoning and post-reasoning integration methods, the system generates the final answer through an interaction module, a model database, and an answer integration module, using weight calculation and semantic similarity calculation, and provides a visual traceability function.
It improves the confidence of answers and user experience, reduces time costs, enhances the match between answers and user needs, and facilitates user tracing and evaluation.
Smart Images

Figure CN2025098748_12112026_PF_FP_ABST
Abstract
Description
An interactive question and answer system and question and answer method based on multi-model parallel inference TECHNICAL FIELD
[0001] The application relates to the field of large model applications, and particularly relates to an interactive question and answer system and question and answer method based on multi-model parallel inference. BACKGROUND
[0002] The release of GPT-3 has made the society and market realize the value and potential of large language models (hereinafter referred to as large models). In the past few years, companies, laboratories and research institutions around the world have independently or cooperatively released several hundred large model products. The initially released large models are usually trained based on massive multi-source heterogeneous data, and the data sources are extensive, such as news reports, social media, encyclopedias and the like. The trained model focuses on obtaining extensive knowledge and language patterns, and this kind of model becomes a general-purpose model. On the basis of the general-purpose model, the training data is concentrated in a specific field, for example, the training data of a medical professional model may come from medical records, medical research papers, clinical images and the like, so that the performance of the trained model in the professional field will be significantly higher than that of the general-purpose model, which is called a professional model. For ordinary users, large models can provide a lot of help in work, study and life. However, at present, there are many types of large models, and the application scenarios are different. Even if it is a high-performance model, the output fluctuation in some application scenarios is still large, which makes it difficult for users to believe. And the user asks a large number of large models one by one, and then collects and evaluates the answers, which consumes a high time cost, seriously affecting the production and learning efficiency of the user.
[0003] In order to enable users to quickly obtain answers with high confidence, the current method is to use multiple large models to cooperate, and the implementation ways are fusion (Merge), integration (Ensemble) and cooperation (Cooperate). Among them, the integration of large models in the integration way is divided into three kinds: integration before inference, integration during inference and integration after inference.
[0004] (1) Integration before inference
[0005] The integration before inference is to integrate the large language model before the inference operation, and to select the most suitable large model for the current task or input from a large number of large models. The specific implementation way is to train an external router, and the role of the router is to select a suitable large model for this inference task according to certain standards or rules before inference.
[0006] (2) Integration during inference
[0007] The integration during inference is to combine the outputs of multiple large models during the inference process, that is, in the decoding step.
[0008] (3) Post-inference integration method
[0009] The post-inference ensemble approach involves having multiple large models perform inference separately, generating their own outputs, and then integrating these outputs after inference is complete. Specifically, multiple large models are sorted according to performance metrics (such as number of parameters, BERTScore, BLEURT, BARTScore, etc.) and then integrated through cascading, selection, or fusion of models.
[0010] The existing integration methods have the following shortcomings:
[0011] 1) While pre-inference ensemble can screen and ensemble large language models in advance to avoid poor performance due to the use of unsuitable large language models during inference, the selected single large model may not provide the most comprehensive and accurate answer. Furthermore, the selection of the large model is greatly affected by the router. When the router performance is poor, the quality of the final output answer will be poor.
[0012] 2) While the final answer obtained through post-inference integration integrates answers from different large models, users cannot intervene in the integration process according to their own needs. This may result in the loss of some information that is important to the user. Furthermore, users cannot trace the source of the final answer, making it difficult to ask follow-up questions. Summary of the Invention
[0013] To address the shortcomings of existing technologies, this invention proposes an interactive question-answering system and method based on multi-model parallel reasoning, combining pre-inference and post-inference integration approaches. The specific technical solution is as follows:
[0014] An interactive question-answering system based on multi-model parallel reasoning includes an interaction module, a model database, and an answer integration module;
[0015] The interaction module is used to receive user input, send it to the answer integration module, and display the information.
[0016] The model database is used to store information sets of n large models;
[0017] The answer integration module includes a weight calculation submodule, an answer integration submodule, and a statistical information calculation and visualization submodule; wherein...
[0018] The weight calculation submodule has a built-in model weighting algorithm that is adjusted by user preferences. Based on the question raised by the user and the information of each large model, it calculates the weight of each large model for the question raised by the user.
[0019] The answer integration submodule has a built-in answer integration algorithm that is adjusted by user preferences. Based on the answers given by each large model to the user's question, the weights of each model passed by the weight calculation submodule, and the user's preference settings, the final integrated answer is calculated.
[0020] The statistical information calculation and visualization submodule calculates the statistical information of each word in the final integrated answer in all model answers based on the final integrated answer and the answers of each model to the user's question, and then visualizes and displays it in the interactive module.
[0021] Furthermore, the specific execution process of the weight calculation submodule is as follows:
[0022] S1.1: Initialize the user ratings for all large models to unbiased ratings, and adjust the ratings for the corresponding large models to the user input values based on the user input.
[0023] S1.2: Using machine learning methods, each large model label t in the set of tags consisting of all large model labels in the model database is selected. i Convert to vector representation v i Construct a model tag vector set (tags) v ; where tags = {t1, t2, ..., t i ,…,t x}, tags v ={v1,v2,…,v i ,…,v x Additionally, for each large model M in the model database... i Construct a model label matrix for all corresponding labels. Where 'a' represents the total number of tags;
[0024] S1.3: Extract keywords from user questions using neural network-based methods in natural language processing. Provide a classification label for the question based on a set of tags composed of all large model labels. Select the k tags with the highest probabilities from the algorithm results and construct a question label matrix. And record the probability value of each label.
[0025] S1.4: Match the extracted user question classification labels against all labels owned by the large models in the model database, and calculate the similarity matrix S. i The successfully matched models are then sorted from highest to lowest matching score, and the matching scores are normalized to obtain matching weight values.
[0026] S1.5: Based on user selection, use either publicly available ratings or user-defined ratings for the large models on the network to construct an aggregate of all large model ratings and normalize them. The model rating is denoted as "score". norm ={s1,…,s i ,…,s n} Calculate the weight w of the i-th large model for the question raised by the current user. i And calculate the weight w i The result is passed to the answer integration submodule for subsequent calculations.
[0027] Furthermore, in step S1.4, the specific matching process for matching the extracted user question classification labels against all the labels owned by the large models in the model database is as follows:
[0028] The problem label vector V q With each large model M in the model database i Model label matrix Calculate the weighted cosine vector similarity matrix S i If S i Each row in the table has a similarity threshold h that is not less than the user-defined threshold. s If the value is , then the model M is considered to be . i Match successful, and S recorded. i In each row, the value is not less than h s The number of values is used as the model matching degree.
[0029] Furthermore, in S1.5, the weights w of each large model i The calculation formula is as follows:
[0030] Where z is the user-set preference coefficient, used to indicate the overall control system's tendency to make decisions.
[0031] Furthermore, the process by which the answer integration submodule calculates the final integrated answer is as follows:
[0032] S2.1: First, read the answers output by each major model for this problem from the model database, and then standardize them one by one;
[0033] S2.2: Map the answer output by each large model to a high-dimensional semantic space to generate the corresponding semantic embedding vector; then, calculate the semantic similarity between all answer pairs based on cosine similarity and construct a symmetric similarity matrix, denoted as Sim(i,j), where i and j represent different answers, and Sim(i,j) represents the semantic similarity between answer i and answer j.
[0034] S2.3: Combining the weights w of various major modelsi Based on semantic similarity, calculate the final integrated weight for each answer:
[0035] Where, α i denoted by , where represents the integration weight of the i-th answer, n represents the number of answers, λ is a hyperparameter that adjusts the influence of semantic similarity, with a value range of [0,1]; ∈ is the semantic similarity threshold; f(Sim(i,j),∈) is a piecewise function;
[0036] S2.4: Select a large model with contextual understanding, reasoning, and text generation capabilities as the final answer integration model. By inputting the original answers of each large model and the final integration weight, generate a logically coherent and grammatically fluent answer. Receive user input of keywords related to the question and semantic similarity thresholds to adjust the semantic tendency of the final integrated answer. The answer integration submodule also maps each word in the final integrated answer back to the original answer.
[0037] Furthermore, the interaction module specifically includes:
[0038] Initialization phase: Receive questions from users, default model data selected by users, or custom model data, and display model information in categories;
[0039] The interactive stage of answer integration: receiving keywords related to the user's question, semantic similarity threshold, modified model information, and preference coefficients; displaying the integrated answer obtained from the answer integration module; and displaying the user's question parameters and information on the various major models used in the integrated answer.
[0040] Once the final integrated answer is determined: the user's word highlighting operation is received, and the source tracing and statistical information related to the final integrated answer is displayed.
[0041] Furthermore, the information set of each large model in the model database is represented by a tuple: M i =(name i ,source i ,parameters i ,score_original i score_client i ,domain_orginal i ,domain_client i tags_original i ,tags_client i weight i description i ,answeri ,…)
[0042] Among them, name i Represents the name of the i-th large model; source i Indicates the source of the i-th large model (which company, institution, or university released it); parameters i Indicates the number of parameters in the i-th large model; score_original i Represents the public score of the i-th large model on the network; score_client i This represents the user's rating of the i-th large model; domain_orginal i The domain_client represents the application domain targeted by the publisher of the i-th large model. i It refers to the application area that the user deems the i-th large model suitable based on their own needs; tags_original i Tags_client represents the tags assigned to the publisher of the i-th large model. i These are the labels that users assign to the i-th large model based on their own preferences; weight i The weight score of the i-th large model is derived from a model weighting algorithm adjusted by user preferences; description i This represents the publisher's description of the i-th large model, which users can also modify; answer i Used to store the answer obtained by the i-th large model in the reasoning of the user's question.
[0043] Furthermore, the initial value of the semantic similarity threshold ∈ is 0.5. Increasing the value of ∈ will filter out answers with higher relevance during the answer integration process, while decreasing the value of ∈ will introduce more potential related answers during the answer integration process.
[0044] Furthermore, the standardization process specifically includes unifying capitalization, removing punctuation, and removing stop words.
[0045] An interactive question-answering method based on multi-model parallel reasoning is disclosed. This method is implemented by an interactive question-answering system based on multi-model parallel reasoning and includes the following steps:
[0046] Step 1: The interaction module sets model information, loads default values, selects a model group, receives user questions, and sends user questions to all large models in the model cluster for answers;
[0047] Step 2: The weight calculation submodule in the answer integration module classifies and locates user questions, and calculates the weight of each major model in the model database for the user's question;
[0048] Step 3: The answer integration submodule in the answer integration module outputs the integrated answer based on the answers given by each model to the user's question, the weights of each model passed by the weight calculation submodule, and the user's preference settings.
[0049] Step 4: The integrated answer is displayed to the user in real time through the interaction module, and the user's instructions are received to confirm whether it is the final integrated answer; if yes, proceed to step 5; if no, the user's modified preference settings are received, and the answer integration submodule re-outputs the integrated answer.
[0050] Step 5: The statistical information calculation and visualization submodule in the answer integration module calculates the statistical information of each word in the final integrated answer in all model answers based on the final integrated answer and the answers of each major model. After visualization processing, the data is displayed by the interactive module.
[0051] The beneficial effects of this invention are as follows:
[0052] (1) In the system and method of the present invention, after each user asks a question, each model will receive the user's question and provide a reasoning answer. The system assigns weights to all large models based on the user's question and the various parameter indicators of the models in the system. These weights are only for the question asked by the user in this instance. The present invention can consider and retain valuable answers provided by some models with insufficient overall scores.
[0053] (2) The system and method of the present invention enable users to personalize the answer integration process according to their own needs (such as setting keywords, setting word frequency thresholds, etc.), and the integrated answer will change with the user's intervention, so that the final answer matches the user's needs more closely.
[0054] (3) The system and method of the present invention can identify each word and phrase in the final integrated answer, making it easier for users to query at any time which one or more models any word or phrase in the final integrated answer comes from, and making it easier for users to trace the source.
[0055] (4) The system and method of the present invention provide users with a visual display of answer differences, enabling users to easily view the consistency, difference, and difference points between the answer of any model and the final integrated answer, thereby reducing the time cost of reading and evaluating the answers of each large model and improving the efficiency of use. Attached Figure Description
[0056] Figure 1 is a schematic diagram showing the composition of the interactive question-answering system based on multi-model parallel reasoning of the present invention, and its relationship with users and model clusters.
[0057] Figure 2 is a schematic diagram of the initialization state of the interactive module.
[0058] Figure 3 is a schematic diagram of the interactive process of integrating answers.
[0059] Figure 4 is a schematic diagram of the process of tracing the source of the answer and visualizing it.
[0060] Figure 5 is a flowchart of the interactive question-and-answer method of the present invention, wherein the solid arrows represent the initial run process and the dashed arrows represent the process in which the user intervenes in the answer according to their preferences. Detailed Implementation
[0061] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.
[0062] Explanation of technical terms:
[0063] GloVe: Global Vectors for Word Representation
[0064] Word2vec: Word to vector, an open-source toolkit released by Google in 2013 for obtaining word vectors;
[0065] BERT model, Bidirectional Encoder Representations from Transformers, is a pre-trained language model;
[0066] Sentence-BERT: Sentence Bidirectional Encoder Representations from Transformers, a sentence embedding representation model;
[0067] GPT: Generative Pre-trained Transformer.
[0068] On the one hand, the present invention provides an interactive question-answering system based on multi-model parallel reasoning, as shown in Figure 1. The system includes three parts: an interaction module, a model database, and an answer integration module.
[0069] I. Interaction Module
[0070] The interaction module receives user input and sends it to the answer integration module, while simultaneously retrieving and displaying the final integrated answer from the answer integration module. User input differs at different stages of the question-and-answer process, and is categorized as follows:
[0071] Initialization phase: The interaction module receives questions from users, default model data selected by users, or custom model data, and sends them to the answer integration module, which then categorizes and displays the various models and their parameters in the model group.
[0072] Interactive phase of answer integration: When a user is not satisfied with the integrated answer, the interaction module receives the user's input of keywords related to the question, semantic similarity threshold, modified model information, preference coefficient, etc., and sends them to the answer integration module. Then, it retrieves the regenerated integrated answer from the answer integration module and displays it, while also showing the user's question parameters and information on the various models used in the integrated answer.
[0073] Once the final integrated answer is determined: the interaction module receives the user's word selection operation and visually displays the source information and statistical information of the selected words.
[0074] After the system starts, the interaction module initializes. Figure 2 shows the interactive interface during the initialization phase. Area A is the input interaction area, serving as the entry point for user questions. It uses text input boxes to provide raw input data for the subsequent question-and-answer process. Area B is used for model database management, employing a dual-mode switching mechanism. It supports two loading modes: default model data and custom model database. Default model data directly calls the system's preset model scores, tags, and other metadata. Custom model data allows users to modify basic model information (such as score weights and domain labels), and the modifications will be used as input parameters for calculating the model's answer weights. Area C displays model information, including static information such as model name and version number, as well as configurable information such as domain labels and score weights. Area D is used for model filtering, providing filtering dimensions such as domain, tags, and parameter count, enabling multi-condition combined filtering. By default, all options are selected.
[0075] As shown in Figure 3, during the interactive phase of answer integration, the interaction module receives the user's question, retrieves the integrated answer from the answer integration module, and displays it. Area A displays the final integrated answer, following the user-configured model weights and filtering conditions. Area B is used for parameter intervention. It first displays the parameters used in generating the final integrated answer, providing modification functionality to meet user needs. Simultaneously, it receives user-input parameters related to the question, such as keywords, semantic similarity thresholds, modified model information, and preference coefficients. After parameter modification, these parameters are sent to the answer integration module for calculation. The answer content displayed in Area A updates in real-time as the parameters change.
[0076] As shown in Figure 4, once the final integrated answer is determined, i.e. when the user is satisfied, the interaction module provides the word selection and source tracing function, receives the user's word selection operation, and starts the data source tracing and statistical process. Area B is used to display the source information and statistical information of the word. At the same time, some appropriate statistical information is visualized.
[0077] II. Model Database
[0078] The model database is used to store information sets M = {M1, M2, M3, ..., Mn} of n large models. i ,…,M n}, where any large model M i The information is represented using a tuple: M i =(name i ,source i ,parameters i ,score_original i score_client i ,domain_orginal i ,domain_client i tags_original i ,tags_client i weight i description i ,answer i ,…)
[0079] Among them, name i Represents the name of the i-th large model; source i Indicates the source of the i-th large model (which company, institution, or university released it); parameters i Indicates the number of parameters in the i-th large model; score_original i Represents the public score of the i-th large model on the network; score_client i This represents the user's rating of the i-th large model; domain_orginal i The domain_client represents the application domain targeted by the publisher of the i-th large model. i It refers to the application area that the user deems the i-th large model suitable based on their own needs; tags_original i Tags_client represents the tags assigned to the publisher of the i-th large model. i These are the labels that users assign to the i-th large model based on their own preferences; weight iThe weight score of the i-th large model is derived from a model weighting algorithm adjusted by user preferences; description i This represents the publisher's description of the i-th large model, which users can also modify; answer i Used to store the answer obtained by the i-th large model in the reasoning of the user's question.
[0080] III. Answer Integration Module
[0081] The answer integration module is divided into a weight calculation submodule, an answer integration submodule, and a statistical information calculation and visualization submodule.
[0082] 3.1. Weight Calculation Submodule
[0083] The input to the weight calculation submodule is the user-submitted question and information from each large model; the output is the weight w of each large model for the user-submitted question. i The weight calculation submodule has a built-in model weighting algorithm that is adjusted according to user preferences to calculate the weights w of each major model for the user's question. i The specific implementation process is as follows:
[0084] (1) Initialization configuration
[0085] First, score the users of all large models using score_client. i Initialize to 5.0 (representing no preference), and adjust the corresponding large model's rating to the user input value based on the user input.
[0086] (2) Vector representation of model labels
[0087] Using machine learning methods (such as Word2vec or GloVe), each large model label t in the set of tags consisting of all large model labels in the model database is analyzed. i Convert to vector representation v i Construct a model tag vector set (tags) v ; where tags = {t1, t2, ..., t i ,…,t x}, tags v ={v1,v2,…,v i ,…,v x Additionally, for each large model M in the model database... i Construct a model label matrix for all corresponding labels. Where 'a' represents the total number of tags.
[0088] (3) Problem tag extraction
[0089] Keywords from user questions are extracted using neural network-based methods in natural language processing (such as the BERT model). A classification label for the question is then provided based on a set of tags composed of all large model labels. The k tags with the highest probabilities (k is input by the user, defaults to 3, k≤x) are selected according to the algorithm results, and a question label matrix is constructed. And record the probability value of each label.
[0090] (4) Tag matching
[0091] The extracted user question classification labels are matched against all labels available in the large model database. Specifically, the question label vector V is... q With each model M in the model database i Model label matrix Calculate the weighted cosine vector similarity matrix S i If S i Each row in the table has a similarity threshold h that is not less than the user-defined threshold. s If the value is , then the model M is considered to be . i Match successful, and S recorded. i In each row, the value is not less than h s The number of values is used as the model matching score; finally, the successfully matched models are sorted from high to low matching scores, and the matching scores are normalized to obtain the matching weight values.
[0092] The formula for calculating the similarity matrix is as follows:
[0093] Where ⊙ represents the Hadamard product operation, S i Let be a matrix of size j×k, representing the similarity of each question label to each model label.
[0094] (5) Model weight calculation
[0095] Based on user selection, the scores of all models are aggregated using either publicly available online ratings or user-defined ratings, and then normalized. The model scores are denoted as "scores". norm ={s1,…,s i ,…,s n} Calculate the weight value w of the i-th model for the question raised by the current user. i And calculate the weights w of each model. i The result is passed to the answer integration submodule for subsequent calculations.
[0096] The weight value w of the i-th model in response to the question raised by the current user. iThe calculation formula is as follows:
[0097] Where z is the user-set preference coefficient, used to indicate the overall control system's tendency to make decisions.
[0098] 3.2 Answer Integration Submodule
[0099] The input to the answer integration submodule is the answers provided by each major model to the user's question, and the weights w of each model passed from the weight calculation submodule. i Including user preference settings, the output is the final integrated answer, designed to address the issues of user inability to intervene in the answer integration process and the inability to trace the source of the final answer. The answer integration submodule has a built-in answer integration algorithm that is adjustable by user preferences, and the specific execution process is as follows:
[0100] 1) Text preprocessing
[0101] First, the answers output by each model for this problem are read from the model database and standardized one by one to eliminate the influence of format and syntax and ensure the accuracy of semantic calculation, including unifying capitalization, removing punctuation, and removing stop words.
[0102] 2) Semantic weight calculation
[0103] The Sentence-BERT model is used to map the answer output by each model into a high-dimensional semantic space, generating a corresponding semantic embedding vector. Then, the semantic similarity between all answer pairs is calculated based on cosine similarity, and a symmetric similarity matrix is constructed, denoted as Sim(i,j), where i and j represent different answers, and Sim(i,j) represents the semantic similarity between answer i and answer j.
[0104] 3) Integrated weight calculation
[0105] The final integrated weight for each answer is calculated by combining the model weights and semantic similarity. Assuming there are n answers, the integrated weight α for the i-th answer is... i The calculation formula is:
[0106] Among them, w i This represents the model weight of the i-th model, obtained by a model weighting algorithm adjusted by user preferences. The initial value is set to the weight obtained by the weight calculation module, reflecting its reliability. Users can subsequently adjust the model weight according to their own preferences to intervene in the content of the final integrated answer.
[0107] λ is a hyperparameter that adjusts the influence of semantic similarity, with a value range of [0,1]. Users can adjust it according to task requirements. For tasks that require a high degree of consistency in answers (such as computational problems), increase the value of λ; for tasks that require a comprehensive and diverse perspective (such as open-ended questions), decrease the value of λ.
[0108] ∈ is a semantic similarity threshold used to filter out answer pairs with low semantic similarity to reduce interference from irrelevant information. An initial value of 0.5 is recommended, and it can be adjusted based on experimental data to suit specific scenario needs. Users can control answer integration by adjusting the ∈ value. Increasing the ∈ value will cause the system to more strictly filter highly relevant answers, suitable for scenarios with high consistency requirements; decreasing the ∈ value will allow the system to introduce more potentially relevant answers, suitable for questions requiring thought-provoking discussion.
[0109] 4) Answer generation
[0110] A large-scale model with contextual understanding, reasoning, and text generation capabilities is selected as the final answer integration model (specific criteria: large-scale multi-task language understanding of at least 80%; context window size of at least 32kb; GPT is selected in this embodiment). Logically coherent and grammatically fluent answers are generated by inputting the original answers from each model and the final integration weights. Users can also influence the semantic orientation of the final integrated answer by setting keywords. In the generated answers, the answer integration submodule maps each word in the final integrated answer back to the original answer (preserving the mapping relationship between the original and final integrated answers) to achieve answer traceability, meaning that each word in the final integrated answer can be traced back to its main contributing model. Users can understand the impact of each model on the final integrated answer through traceability information and can influence the content of the final integrated answer according to their own preferences by adjusting various parameters.
[0111] 3.3 Statistical Information Calculation and Visualization Submodule
[0112] The input to the statistical information calculation and visualization submodule is the final integrated answer and the answers of each model to the user's question. The output is the statistical information of each word in the final integrated answer in all model answers, such as the percentage of answers that use the word, the number of times the word appears in each answer, the number of synonyms / antonyms, etc. Appropriate data formats are displayed using corresponding visualization methods such as pie charts, bar charts, word clouds, etc.
[0113] On the other hand, the present invention also provides an interactive question-answering method based on multi-model parallel reasoning, as shown in Figure 5. The method includes the following steps:
[0114] Step 1: The interaction module sets model information, loads default values, selects a model group, receives user questions, and sends user questions to all large models in the model cluster for answers;
[0115] Step 2: The weight calculation submodule in the answer integration module classifies and locates user questions, and calculates the weight of each major model in the model database for the user's question;
[0116] Step 3: The answer integration submodule in the answer integration module outputs the integrated answer based on the answers given by each model to the user's question, the weights of each model passed by the weight calculation submodule, and the user's preference settings.
[0117] Step 4: The integrated answer is displayed to the user in real time through the interaction module, and the user's instructions are received to confirm whether it is the final integrated answer; if yes, proceed to step 5; if no, the user's modified preference settings are received, and the answer integration submodule re-outputs the integrated answer.
[0118] Step 5: The statistical information calculation and visualization submodule in the answer integration module calculates the statistical information of each word in the final integrated answer in all model answers based on the final integrated answer and the answers of each major model. After visualization processing, the data is displayed by the interactive module.
[0119] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.
Claims
1. An interactive question-answering system based on multi-model parallel reasoning, characterized in that, Includes an interaction module, a model database, and an answer integration module; The interaction module is used to receive user input, send it to the answer integration module, and display the information. The model database is used to store information sets of n large models; The answer integration module includes a weight calculation submodule, an answer integration submodule, and a statistical information calculation and visualization submodule; wherein... The weight calculation submodule has a built-in model weighting algorithm that is adjusted by user preferences. Based on the question raised by the user and the information of each large model, it calculates the weight of each large model for the question raised by the user. The answer integration submodule has a built-in answer integration algorithm that is adjusted by user preferences. Based on the answers given by each large model to the user's question, the weights of each model passed by the weight calculation submodule, and the user's preference settings, the final integrated answer is calculated. The statistical information calculation and visualization submodule calculates the statistical information of each word in the final integrated answer in all model answers based on the final integrated answer and the answers of each model to the user's question, and then visualizes and displays it in the interactive module.
2. The interactive question-answering system based on multi-model parallel reasoning according to claim 1, characterized in that, The specific execution process of the weight calculation submodule is as follows: S1.1: Initialize the user ratings for all large models to unbiased ratings, and adjust the ratings for the corresponding large models to the user input values based on the user input. S1.2: Using machine learning methods, each large model label t in the set of tags consisting of all large model labels in the model database is selected. i Convert to vector representation v i Construct a model tag vector set (tags) v ; Where tags = {t1, t2, ..., t i ,…,t x }, tags v ={v1,v2,…,v i ,…,v x Additionally, for each large model M in the model database... i Construct a model label matrix for all corresponding labels. Where 'a' represents the total number of tags; S1.3: Extract keywords from user questions using neural network-based methods in natural language processing. Provide a classification label for the question based on a set of tags composed of all large model labels. Select the k tags with the highest probabilities from the algorithm results and construct a question label matrix. And record the probability value of each label. S1.4: Match the extracted user question classification labels against all labels owned by the large models in the model database, and calculate the similarity matrix S. i The successfully matched models are then sorted from highest to lowest matching score, and the matching scores are normalized to obtain matching weight values. S1.5: Based on user selection, use either publicly available ratings or user-defined ratings for the large models on the network to construct an aggregate of all large model ratings and normalize them. The model rating is denoted as "score". norm ={s1,…,s i ,…,s n } Calculate the weight w of the i-th large model for the question raised by the current user. i And calculate the weight w i The result is passed to the answer integration submodule for subsequent calculations.
3. The interactive question-answering system based on multi-model parallel reasoning according to claim 2, characterized in that, In step S1.4, the specific matching process for matching the extracted user question classification labels against all the labels owned by the large models in the model database is as follows: The problem label vector V q With each large model M in the model database i Model label matrix Calculate the weighted cosine vector similarity matrix S i If S i Each row in the table has a similarity threshold h that is not less than the user-defined threshold. s If the value is , then the model M is considered to be . i Match successful, and S recorded. i In each row, the value is not less than h s The number of values is used as the model matching degree.
4. The interactive question-answering system based on multi-model parallel reasoning according to claim 2, characterized in that, In S1.5, the weights w of each large model i The calculation formula is as follows: Where z is the user-set preference coefficient, used to indicate the overall control system's tendency to make decisions.
5. The interactive question-answering system based on multi-model parallel reasoning according to claim 2, characterized in that, The process by which the answer integration submodule calculates the final integrated answer is as follows: S2.1: First, read the answers output by each major model for this problem from the model database, and then standardize them one by one; S2.2: Map the answer output by each large model to a high-dimensional semantic space to generate the corresponding semantic embedding vector; then, calculate the semantic similarity between all answer pairs based on cosine similarity and construct a symmetric similarity matrix, denoted as Sim(i,j), where i and j represent different answers, and Sim(i,j) represents the semantic similarity between answer i and answer j. S2.3: Combining the weights w of various major models i Based on semantic similarity, calculate the final integrated weight for each answer: Where, α i denoted by , where represents the integration weight of the i-th answer, n represents the number of answers, λ is a hyperparameter that adjusts the influence of semantic similarity, with a value range of [0,1]; ∈ is the semantic similarity threshold; f(Sim(i,j),∈) is a piecewise function; S2.4: Select a large model with contextual understanding, reasoning, and text generation capabilities as the final answer integration model. By inputting the original answers of each large model and the final integration weight, generate a logically coherent and grammatically fluent answer. Receive user input of keywords related to the question and semantic similarity thresholds to adjust the semantic tendency of the final integrated answer. The answer integration submodule also maps each word in the final integrated answer back to the original answer.
6. The interactive question-answering system based on multi-model parallel reasoning according to claim 1, characterized in that, The interaction module specifically includes: Initialization phase: Receive questions from users, default model data selected by users, or custom model data, and display model information in categories; The interactive stage of answer integration: receiving keywords related to the user's question, semantic similarity threshold, modified model information, and preference coefficients; displaying the integrated answer obtained from the answer integration module; and displaying the user's question parameters and information on the various major models used in the integrated answer. Once the final integrated answer is determined: the user's word highlighting operation is received, and the source tracing and statistical information related to the final integrated answer is displayed.
7. The interactive question-answering system based on multi-model parallel reasoning according to claim 1, characterized in that, The information set of each large model in the model database is represented by a tuple: M i =(name i ,source i ,parameters i ,score_original i score_client i ,domain_orginal i ,domain_client i tags_original i ,tags_client i weight i description i ,answer i ,…) Among them, name i Represents the name of the i-th large model; source i Indicates the source of the i-th large model (which company, institution, or university released it); parameters i Indicates the number of parameters in the i-th large model; score_original i Represents the public score of the i-th large model on the network; score_client i This represents the user's rating of the i-th large model; domain_orginal i The domain_client represents the application domain targeted by the publisher of the i-th large model. i It refers to the application area that the user deems the i-th large model suitable based on their own needs; tags_original i Tags_client represents the tags assigned to the publisher of the i-th large model. i These are the labels that users assign to the i-th large model based on their own preferences; weight i The weight score of the i-th large model is derived from a model weighting algorithm adjusted by user preferences; description i This represents the publisher's description of the i-th large model, which users can also modify; answer i Used to store the answer obtained by the i-th large model in the reasoning of the user's question.
8. The interactive question-answering system based on multi-model parallel reasoning according to claim 1, characterized in that, The initial value of the semantic similarity threshold ∈ is 0.
5. Increasing the value of ∈ will filter out answers with higher relevance during the answer integration process, while decreasing the value of ∈ will introduce more potential related answers during the answer integration process.
9. The interactive question-answering system based on multi-model parallel reasoning according to claim 5, characterized in that, The standardization process specifically includes unifying capitalization, removing punctuation, and removing stop words.
10. An interactive question-answering method based on multi-model parallel reasoning, characterized in that, This method is implemented by the interactive question-answering system based on multi-model parallel reasoning as described in claim 1, and includes the following steps: Step 1: The interaction module sets model information, loads default values, selects a model group, receives user questions, and sends user questions to all large models in the model cluster for answers; Step 2: The weight calculation submodule in the answer integration module classifies and locates user questions, and calculates the weight of each major model in the model database for the user's question; Step 3: The answer integration submodule in the answer integration module outputs the integrated answer based on the answers given by each model to the user's question, the weights of each model passed by the weight calculation submodule, and the user's preference settings. Step 4: The integrated answer is displayed to the user in real time through the interactive module, and the user's instructions are received to confirm whether it is the final integrated answer; If yes, proceed to step five; if no, receive the user's modified preference settings parameters and have the answer integration submodule re-output the integrated answer. Step 5; The statistical information calculation and visualization submodule in the answer integration module calculates the statistical information of each word in the final integrated answer in all model answers based on the final integrated answer and the answers of each major model. After visualization processing, the data is displayed by the interactive module.
