Large model question and answer method and device, computer equipment and medium

Abstract

The invention discloses a large model question and answer method and device, computer equipment and a medium, and the large model question and answer method selects at least two target industry models from a preset industry model library according to question-related labels, and achieves multi-model cooperative work. The target industry models analyze questions from respective professional fields, and then the general large model is combined with the preset sharing mechanism to fuse analysis results, so that professional knowledge of multiple industries can be synthesized, targeted answers can be given, and the requirements of users for cross-professional questions are met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a large model question and answer method and device, a computer device and a medium. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models are increasingly widely used in the question and answer field. At present, the mainstream question and answer scheme in the prior art uses a single comprehensive large model or a single industry model to answer user questions. Although a single comprehensive large model can cover a wide range of knowledge, it is difficult to combine professional knowledge of various industries when facing complex problems across professions and fields, resulting in a lack of depth and pertinence in the answer. Although a single industry model has professional ability in a specific field, it cannot handle problems involving multiple industry intersections, resulting in question and answer results that cannot meet the actual needs of users. SUMMARY

[0003] The embodiments of the application provide a large model question and answer method, device, computer device and medium, aiming to solve the problem that the existing large model for question and answer cannot meet the actual needs of users for question and answer results of multiple industry intersections.

[0004] In a first aspect, the embodiments of the application provide a large model question and answer method, comprising: S101, vectorizing user question data to obtain vectorized question data; S102, using a trained keyword extraction large model to extract and classify the vectorized question data to generate question-related labels; S103, performing similarity matching from a pre-set industry model library according to the question-related labels, and selecting at least two target industry models according to the matching results; S104, analyzing the user question data through the selected target industry models to generate corresponding model analysis results; S105, using a general large model to combine a pre-set sharing mechanism to fuse all model analysis results to obtain a preliminary answer result; S106, obtaining evaluation feedback of the user on the preliminary answer result, if the evaluation feedback meets a pre-set standard, outputting the preliminary answer result as a final question and answer result; if the evaluation feedback does not meet the pre-set standard, supplementing question information based on the evaluation feedback and returning to S101 to re-execute the subsequent process.

[0005] In a second aspect, the embodiments of the application provide a large model question and answer device, comprising: a vectorization unit configured to vectorize user question data to obtain vectorized question data; The extraction and classification unit is configured to extract and classify the vectorized question data to generate a question-related label. The matching unit is configured to perform similarity matching from a preset industry model library according to the question-related label, and select at least two target industry models according to the matching result. The analysis unit is configured to analyze the user question data by using the selected target industry models to generate corresponding model analysis results. The fusion unit is configured to fuse all the model analysis results by using a general large model in combination with a preset sharing mechanism to obtain a preliminary answer result. The evaluation unit is configured to obtain an evaluation feedback of the user on the preliminary answer result, and if the evaluation feedback meets a preset standard, the preliminary answer result is output as a final question and answer result. If the evaluation feedback does not meet the preset standard, the evaluation feedback is supplemented with question information and returned to the vectorization unit to re-execute the subsequent process.

[0006] In a third aspect, an embodiment of the present application provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the program to implement the large model question and answer method provided by the embodiment of the present application.

[0007] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium storing a computer program, and the computer program is executed to implement the large model question and answer method provided by the embodiment of the present application.

[0008] In the embodiment of the present application, at least two target industry models are selected from a preset industry model library according to a question-related label to realize the collaborative work of multiple models. These target industry models analyze the question from their respective professional fields, and then the general large model is used in combination with a preset sharing mechanism to fuse the analysis results, so as to comprehensively utilize the professional knowledge of multiple industries and give a more in-depth and targeted answer to meet the user's demand for cross-professional questions. BRIEF DESCRIPTION OF DRAWINGS

[0009] The technical solutions of the present application and their beneficial effects will become apparent through the following detailed description of the specific embodiments of the present application combined with the accompanying drawings.

[0010] Figure 1 FIG. 1 is a flowchart of a large model question and answer method provided by an embodiment of the present application.

[0011] Figure 2 FIG. 2 is a structural schematic diagram of a large model question and answer device provided by an embodiment of the present application.

[0012] Figure 3 FIG. 3 is a schematic block diagram of a computer device provided by an embodiment of the present application. Detailed Implementation

[0013] It should be noted that the terms "first," "second," and "third," etc., used in this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the listed steps or modules, but some embodiments also include steps or modules not listed, or some embodiments also include other steps or modules inherent to these processes, methods, products, or devices.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] This application provides a large model question answering method, apparatus, computer device, and medium. The execution entity of the large model question answering method can be the large model question answering apparatus provided in this application, which can be implemented in hardware or software.

[0016] Please see Figure 1 , Figure 1 This is a flowchart illustrating a large-model question-answering method provided in an embodiment of this application. The process may include: In S101, the user problem data is vectorized to obtain vectorized problem data.

[0017] In this embodiment, the user question data is the text question input by the user in the large model. The input question is vectorized to convert the text question data into a numerical vector form, resulting in vectorized question data. This vectorized question data is used as the input data for subsequent model processing.

[0018] In an optional embodiment, S101 includes: Break down user question data into multiple semantic units; Each semantic unit is converted into a corresponding semantic vector through the embedding model; All semantic vectors are weighted and fused to obtain vectorized problem data.

[0019] In this embodiment, tools such as jieba / NLTK can be used to segment the user question data into words, and part-of-speech tagging (such as nouns, verbs, and adjectives) can be combined to identify key components, thereby decomposing the user question data into multiple semantic units. Specifically, a BERT-base / large pre-trained model is used as the embedding model, and the parameters of the embedding model are configured, namely, the sequence length is set to 128 / 256 to adapt to the text length of the user question data and avoid truncating key information; the vector output dimension is fixed at 768 dimensions, and the configured embedding model is used to convert each semantic unit into a corresponding high-dimensional numerical semantic vector.

[0020] For each semantic unit, weights can be assigned based on the importance of the semantic unit to complete semantic vector fusion. Specifically, the weight is set as the first weight according to the domain keywords, the weight is set as the second weight according to the requirement keywords, and the weight is set as the third weight according to the limiting condition keywords. The first weight is greater than the second weight, and the second weight is greater than the third weight. Then, a weighted average algorithm is used to fuse all semantic vectors to obtain and output vectorized problem data.

[0021] In practical implementation, taking the medical field as an example, the domain keywords can be terms containing medical information, such as medical imaging diagnosis, clinical medication, and cancer. The demand keywords can be CT reports and medication regimens. The limiting keyword is early-stage cancer. Then, the first weight is set to 1, the second weight to 0.9, and the third weight to 0.6. Substituting these three weight values ​​into the weighted average formula yields the vectorized problem data. It should be noted that this example is only used in the medical field and does not mean that this embodiment can only be used in the medical field.

[0022] In S102, the trained keyword extraction model is used to extract and classify the vectorized question data, and generate question-related labels.

[0023] In this embodiment, in order to improve the accuracy of extraction and classification, the keyword extraction model is pre-trained so that the vectorized question data can be extracted and classified by the trained keyword extraction model, thereby generating structured question-related tags to ensure that the question-related tags can accurately map the domain needs and core demands of the user's questions.

[0024] In an optional embodiment, S102 includes: Collect a training dataset containing multi-domain questions and corresponding domain labels, and use the training dataset to train the initial keyword extraction model to obtain the trained keyword extraction model. The key feature words in the vectorized question data are identified by the large model using the keywords. The key feature words are matched according to the preset classification rules to generate question-related tags corresponding to the domain of the user's question.

[0025] In this embodiment, the collection of multi-domain training datasets covers common industry sectors such as architecture, finance, healthcare, education, new energy, fire protection, law, and psychology. Each sector collects 2000-5000 question samples, with a total sample size of no less than 50,000. Samples must include the original question text, corresponding vectorized data (a 768-dimensional vector consistent with the output of the embedding model in step S101), and the relevant domain label (if a single question contains multiple domain labels, connect the cross-domain labels with "+"). A lightweight DistilBERT model is selected as the initial keyword extraction model, with the training objective being vectorized data → key features + domain labels. The key feature extraction task uses a cross-entropy loss function to optimize the accuracy of key feature recognition, while the domain label matching task uses a multi-label classification loss function to optimize the accuracy of cross-domain label association. The trained initial keyword extraction model is then tested using a test set. Once the test meets the requirements, it can be used as the keyword extraction model for subsequent application stages.

[0026] Vectorized data is input into a trained keyword extraction model. The model extracts key features through the following process: The Transformer layer of the DistilBERT model assigns weights to the semantic features in the vectorized problem data. The model's feature extraction head then filters out phrase-like key features of 2-8 characters in length from the semantic features, sorting them by domain association weight from high to low, and outputting a list of key features. Next, classification rules are constructed based on the keywords, i.e., a key feature word-domain rule base is established. The classification rules use a logic of feature word keyword matching and domain priority ranking. For example, the rule base could be: Key feature word: Cancer, Domain: Medical, Domain Priority: 1; Key feature word: Power Grid, Domain: Electrical Engineering, Domain Priority: 2. Key features from other domains in the pre-set rule base can be designed according to specific application scenarios.

[0027] The key feature word list output by the keyword extraction model is matched one by one with the preset rule base. The matched domains are integrated according to priority from high to low and the deduplication principle to generate question-related tags that only contain the domain to which the user's question belongs.

[0028] In S103, similarity matching is performed from a preset industry model library based on the problem-related tags, and at least two target industry models are selected based on the matching results.

[0029] In this embodiment, the preset industry model library includes multiple industry models from different fields (or different industries). In order to address users' cross-field / cross-industry questions, similarity matching can be performed from the preset industry model library based on question-related tags, and a target industry model can be selected for each field, so as to select at least two target industry models from different fields based on the matching results.

[0030] In an optional embodiment, S103 includes: Retrieve all industry models in the preset industry model library and the industry model tags associated with each industry model; The embedding model is used to vectorize the problem-related labels and the model labels of each industry to obtain problem label vectors and model label vector sets. Calculate the similarity between the label vector and each model label vector in the model label vector set, and select industry models with similarity greater than a preset similarity threshold as candidate industry models; The softmax algorithm is used to calculate the weights of each candidate industry model, and the candidate industry models whose weights are greater than a preset weight threshold are determined as the target industry models.

[0031] In this embodiment, the pre-set industry model library contains industry models from various fields, such as new energy vehicles, insurance, healthcare, education, construction, law, and psychology. The industry model library can be expanded according to actual application scenarios, and in principle, it should cover all core fields of the target question-and-answer scenario to ensure that cross-domain questions can be matched with corresponding models. Each industry model must be associated with a unique industry model tag, which can contain two-dimensional information: domain attributes and core capabilities, ensuring that the industry model tag accurately maps to the model's professional scope. For example, specific tag association rules can be exemplified in Table 1: Table 1

[0032] In other words, the model in the new energy vehicle field can be a battery technology model, the model in the insurance field can be a pricing scheme model, and so on. Then, the extracted correspondence between industry model name and industry model tag is stored in a structured form (such as JSON format) in the metadata management module of the industry model library to ensure that it can be quickly called during subsequent vectorization processing.

[0033] Using the same embedding model as step S101, the problem-related labels and the industry model labels are vectorized to obtain problem label vectors and model label vector sets (the model label vector set stores the model label vectors corresponding to the industry model labels associated with each industry model). These label vectors (including problem label vectors and all model label vectors) are consistent with the semantic space of the problem vectorized data.

[0034] Then, a semantic similarity algorithm is used to calculate the similarity between the question label vector and the label vectors of each model to filter candidate industry models. Specifically, industry models with a similarity greater than a preset similarity threshold are selected as candidate industry models. If more than two candidate industry models are selected, the softmax algorithm is used to calculate the weight of each candidate industry model. Candidate industry models with weights greater than a preset weight threshold are selected and sorted according to their weight values. The two candidate industry models with the highest weight values ​​are selected as target industry models. If exactly two candidate industry models are selected, these two candidate industry models are directly selected as target industry models. If fewer than two candidate industry models are selected, the preset similarity threshold is adjusted to obtain at least two candidate industry models. In specific implementations, the preset similarity threshold can be 0.8, and the preset weight threshold can be 0.6. These specific values ​​can be designed according to the application scenario.

[0035] In S104, the user problem data is analyzed using the selected target industry model to generate the corresponding model analysis results.

[0036] In this embodiment, the user-input question data (i.e., user question data) is parsed using at least two target industry models selected in step S103, and each target industry model generates a model parsing result.

[0037] In S105, the results of all model analysis are fused using a general large model combined with a preset sharing mechanism to obtain a preliminary answer to the question.

[0038] In this embodiment, the text content of the model's parsing results is merged (fused) using a general large model combined with a preset sharing mechanism to obtain a preliminary answer to the question. The general large model can be the GPT model, and the preset sharing mechanism can be a centralized Critic sharing mechanism. In specific implementation, the Critic sharing mechanism adopts a learning structure based on the Actor-Critic framework. The Actor-Critic framework learning structure includes an Actor network and a Critic network. The Actor network is deployed in each target industry model and is responsible for generating parsing results based on the user-input question data. The Critic network is deployed in the general large model as a centralized evaluator, receiving the parsing results and global state information output by all Actor networks and outputting the state value function Vφ(s).

[0039] In an optional embodiment, S105 includes: The centralized Critic sharing mechanism is initiated using a general large model, and the model analysis results and corresponding weights output by all target industry models are received. The weights of each target industry model are adjusted based on the centralized Critic sharing mechanism, and the same or similar solution points in the corresponding model analysis results are merged. By logically organizing and linguistically structuring the key points of the integrated answer, a preliminary response to the question is obtained.

[0040] In this embodiment, the general large model initiates a centralized Critic sharing mechanism. It first receives the model analysis results output by all target industry models. Simultaneously, through this sharing mechanism, it acquires the local input data (o, i.e., the local representation of vectorized question data within the model), output action data (u, i.e., the vector representation of the model analysis results), and global state data (s, i.e., the summary information of the local input data and output action data of all target industry models) generated by each target industry model during the analysis process. Based on the centralized Critic sharing mechanism, a state value function Vφ(s) is defined (where φ is the state value function parameter, trained using user historical question-and-answer records; higher-scoring historical results correspond to higher Vφ(s)). The larger the value, the higher the matching degree between the analysis results of each model and the global state data (s). Model analysis results with a Vφ(s) value greater than the preset threshold (0.7) and conforming to the user's historical rating preferences are obtained. The weights of each target industry model are adjusted, and the analysis results of the corresponding models after weight adjustment are fused. Identical or similar answer points are complemented according to their weight proportions (60% for target industry models with a weight of 0.6, and 40% for target industry models with a weight of 0.4). Logically related answer points are linked according to industry knowledge logic. Simultaneously, the fusion progress is synchronized to each target industry model through a centralized Critic sharing mechanism to ensure the synergy of the fusion process. The general model performs language organization and logical verification on the fused content, forming a preliminary answer to the question with a complete structure and fluent expression.

[0041] In the specific implementation process, the state value function Vφ(s) is obtained by training the Critic network. The training process is as follows: historical question-and-answer records are collected in advance. Each historical question-and-answer record includes global state data, user ratings, and question domain labels. The global state data is obtained by concatenating the local input data and output action data of each target industry model and aggregating them through an attention mechanism, with a dimension of 1024. The user rating is the user's satisfaction rating with the question-and-answer results, with a value range of [0,1]. The question domain labels are used to distinguish questions in different domains. The training set contains 100,000 samples, covering fields such as new energy vehicles, insurance, medical care, education, construction, law, and psychology. The mean squared error is used as the loss function L(φ).

[0042] Where N is the batch size, s i Let R be the global state data for the i-th sample. iThe Critic network is trained to rate users. Using the Adam optimizer with a learning rate of 0.001 and a batch size of 128, the Critic network is trained until the loss converges. The trained Critic network is then used in the online question-answering process to evaluate the quality score (Vφ(s)) of the model output for each target industry in real time.

[0043] In an optional embodiment, the adjustment of the weights of each target industry model based on the centralized Critic sharing mechanism, and the fusion of identical or similar solution points in the corresponding model analysis results, includes: The completeness of the analytical results of each model is verified according to the preset quality assessment rules. If one of the model's analysis results contains errors or is missing key information, the corresponding model's analysis result will be removed, and the weights of the remaining target industry models will be readjusted.

[0044] In this embodiment, after the centralized Critic sharing mechanism is activated in the general large model, the completeness and accuracy of the model analysis results of all target industry models are first verified according to the preset quality assessment rules to screen high-quality and effective analysis results. Invalid results with errors or missing key information are marked and removed through the centralized Critic mechanism. Then, the weights of invalid results are adjusted (can be adjusted to 0). The weights of the remaining valid target industry models are recalculated and reassigned to ensure that the weights are positively correlated with model quality and problem matching degree. Finally, the same or similar answer points in the valid results are merged according to the readjusted weight ratio to form logically coherent and well-founded integrated points.

[0045] In S106, the user's evaluation feedback on the preliminary answer to the question is obtained. If the evaluation feedback meets the preset standard, the preliminary answer to the question is output as the final question and answer result. If the evaluation feedback does not meet the preset standard, the question information is supplemented based on the evaluation feedback and the process returns to S101 to re-execute the subsequent process.

[0046] In this embodiment, the evaluation feedback interface provides various rating options. After the user completes the rating, the backend converts the rating into quantified evaluation data. If the evaluation does not meet the preset standard (e.g., the evaluation data is less than 0.8), supplementary question information (e.g., the question data re-entered by the user) is automatically extracted from the text supplementary explanation area using a keyword extraction algorithm. This supplementary question information is then merged with the original user question data (i.e., the question data entered the first time) to obtain new user question data, and the process returns to step S101 for re-vectorization. Simultaneously, the current process time step t is recorded, and the parameters of each target industry model are optimized using a loss function. The loss function formula is:

[0047] in, This represents the importance sampling ratio based on the policy parameters at time step t and sample i. , The probability distribution of the new strategy, This represents the probability distribution of the old strategy. This represents the action corresponding to time step t and sample i. This represents the observation corresponding to time step t and sample i. Let be the parameters of the i-th target industry model. This represents the expected value at time step t, where min() is the minimum value and clip() is the clipping operation. This is the estimated value of the dominance function. The editing parameter is set to 0.15. The target industry model parameters are updated using a gradient descent algorithm (learning rate 0.001) to improve the accuracy of subsequent question-and-answer sessions. If the evaluation meets the preset criteria (e.g., the evaluation data is greater than or equal to 0.8), the preliminary response result is directly output as the final question-and-answer result.

[0048] In an optional embodiment, the large model question answering method further includes: The user question data, vectorized question data, selected target industry model, model analysis results, preliminary question answer results and evaluation feedback are stored for each question and answer process to form a question and answer knowledge base.

[0049] In this embodiment, by forming a question-and-answer knowledge base, when new user question data is received subsequently, the knowledge base is first searched. If a historical question-and-answer record exists with a similarity greater than a preset similarity threshold, and the user evaluation of the final question-and-answer result in the historical record meets the preset standard, then the final question-and-answer result in that historical record is directly output as the answer to the current user question, without needing to repeat steps S101-S105. If no historical question-and-answer record meets the conditions, then the process of steps S101-S106 is executed. The preset similarity threshold is dynamically adjusted according to the number of historical question-and-answer records, and the similarity calculation can use a cosine similarity algorithm.

[0050] Please see Figure 2 , Figure 2 This is a schematic diagram of a large-scale question-answering device provided in an embodiment of this application. The large-scale question-answering device 200 includes: a vectorization unit 201, an extraction and classification unit 202, a matching unit 203, a parsing unit 204, a fusion unit 205, and an evaluation unit 206.

[0051] The vectorization unit 201 is used to perform vectorization processing on the user question data to obtain vectorized question data; The extraction classification unit 202 is used to extract and classify the vectorized question data and generate question-related labels; The matching unit 203 is used to perform similarity matching from a preset industry model library based on the problem-related tags, and select at least two target industry models based on the matching results; The parsing unit 204 is used to parse the user problem data through the selected target industry model and generate the corresponding model parsing results; The fusion unit 205 is used to fuse the analysis results of all models using a general large model combined with a preset sharing mechanism to obtain a preliminary answer to the question. Evaluation unit 206 is used to obtain user feedback on the preliminary answer to the question. If the feedback meets the preset standard, the preliminary answer to the question is output as the final question and answer result. If the feedback does not meet the preset standard, the question information is supplemented based on the feedback and the process is returned to the vectorization unit to re-execute the subsequent process.

[0052] In an optional embodiment, the vectorization unit 201 can be used to: decompose user question data into multiple semantic units; convert each semantic unit into a corresponding semantic vector through an embedding model; and perform weighted fusion of all semantic vectors to obtain vectorized question data.

[0053] In an optional embodiment, the extraction classification unit 202 can be used to: collect a training dataset containing multi-domain questions and corresponding domain labels; train an initial keyword extraction model using the training dataset to obtain a trained keyword extraction model; use the keyword extraction model to identify key feature words in the vectorized question data; match the key feature words according to preset classification rules to generate question-related labels corresponding to the domain to which the user's question belongs.

[0054] In an optional embodiment, the matching unit 203 may be used to: obtain all industry models in a preset industry model library and the industry model tags associated with each industry model; use an embedding model to vectorize the question-related tags and each industry model tag to obtain a question tag vector and a model tag vector set; calculate the similarity between the tag vector and each model tag vector in the model tag vector set, and select industry models with similarity greater than a preset similarity threshold as candidate industry models; use a softmax algorithm to calculate the weight of each candidate industry model, and determine the candidate industry model with a weight greater than a preset weight threshold as the target industry model.

[0055] In an optional embodiment, the preset sharing mechanism is a centralized Critic sharing mechanism. The fusion unit 205 can be used to: initiate the centralized Critic sharing mechanism using a general large model, and receive the model analysis results and corresponding weights output by all target industry models; adjust the weights of each target industry model based on the centralized Critic sharing mechanism, and fuse the same or similar answer points in the corresponding model analysis results; and perform logical sorting and language organization on the fused answer points to obtain a preliminary answer to the question.

[0056] In an optional embodiment, the fusion unit 205 can also be used to: perform integrity verification on the analysis results of each model according to preset quality assessment rules; if one of the model analysis results has erroneous information or missing key information, the corresponding model analysis result is removed, and the weights of the remaining target industry models are readjusted.

[0057] In an optional embodiment, the large model question answering method further includes a storage unit for storing user question data, vectorized question data, selected target industry model, model parsing results, preliminary question answer results and evaluation feedback for each question answering process, forming a question answering knowledge base.

[0058] Please see Figure 3 This application provides a computer device 300, including a processor 320 and a memory, wherein the memory stores a computer program 332, and when the processor executes the program, it calls the computer program 332 stored in the memory to execute a large model question answering method provided in this embodiment.

[0059] In this embodiment, the computer device 300 includes a processor 320, a memory, and a network interface 350 connected via a system bus 310. The memory may include a non-volatile storage medium 330 and internal memory 340. The non-volatile storage medium 330 stores an operating system 331 and a computer program 332. The computer program 332 includes program instructions that, when executed, cause the processor 320 to execute a large model question-answering method. The processor 320 provides computational and control capabilities to support the operation of the entire computer device 300. The internal memory 340 provides an environment for the execution of the computer program 332 in the non-volatile storage medium 330. When the processor 320 executes the computer program 332, it causes the processor 320 to execute a large model question-answering method.

[0060] This network interface 350 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 300 to which the present application is applied. The specific computer device 300 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0061] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program 332 instructing related hardware. The computer program includes program instructions, and the computer program 332 can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0062] Therefore, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed on a computer, causes the computer to perform a large-scale question-answering method as provided in this embodiment.

[0063] In this embodiment, the processor 320 in the computer device loads the executable code corresponding to the processes of one or more applications into the memory according to the following instructions, and the processor 320 runs the applications stored in the memory, thereby achieving: S101. Vectorize the user problem data to obtain vectorized problem data; S102. The trained keyword extraction model is used to extract and classify the vectorized question data, and generate question-related tags. S103. Based on the problem-related tags, perform similarity matching from a preset industry model library, and select at least two target industry models based on the matching results; S104. Analyze the user problem data using the selected target industry model to generate the corresponding model analysis results; S105. By using a general large model combined with a preset sharing mechanism, the analysis results of all models are fused to obtain a preliminary answer to the question; S106. Obtain user feedback on the preliminary answer to the question. If the feedback meets the preset criteria, output the preliminary answer as the final answer. If the feedback does not meet the preset criteria, supplement the question information based on the feedback and return to S101 to re-execute the subsequent process.

[0064] In an optional embodiment, when the processor 320 executes S101, it may perform the following: decompose the user question data into multiple semantic units; convert each semantic unit into a corresponding semantic vector through an embedding model; and perform weighted fusion of all semantic vectors to obtain vectorized question data.

[0065] In an optional embodiment, when processor 320 executes S102, it may perform the following: collect a training dataset containing multi-domain questions and corresponding domain labels; train an initial keyword extraction model using the training dataset to obtain a trained keyword extraction model; use the keyword extraction model to identify key feature words in the vectorized question data; match the key feature words according to a preset classification rule to generate question-related labels corresponding to the domain to which the user's question belongs.

[0066] In an optional embodiment, when processor 320 executes S103, it may perform the following: obtain all industry models in a preset industry model library and the industry model labels associated with each industry model; use an embedding model to vectorize the problem-related labels and each industry model label to obtain a problem label vector and a model label vector set; calculate the similarity between the label vector and each model label vector in the model label vector set, and select industry models with similarity greater than a preset similarity threshold as candidate industry models; use a softmax algorithm to calculate the weight of each candidate industry model, and determine the candidate industry model with a weight greater than a preset weight threshold as the target industry model.

[0067] In an optional embodiment, when processor 320 executes S105, it may perform the following: initiate the centralized Critic sharing mechanism using a general large model, and receive the model analysis results and corresponding weights output by all target industry models; adjust the weights of each target industry model based on the centralized Critic sharing mechanism, and merge the same or similar answer points in the corresponding model analysis results; perform logical sorting and language organization on the merged answer points to obtain a preliminary answer to the question.

[0068] In an optional embodiment, when the processor 320 executes the adjustment of the weights of each target industry model based on the centralized Critic sharing mechanism and merges the same or similar solution points in the corresponding model analysis results, it may perform the following: perform integrity verification on each model analysis result according to the preset quality assessment rules; if one of the model analysis results has error information or missing key information, then the corresponding model analysis result is removed and the weights of the remaining target industry models are readjusted.

[0069] In an optional embodiment, the processor 320 may also perform the following: storing user question data, vectorized question data, selected target industry model, model parsing results, preliminary question answer results and evaluation feedback for each question-and-answer process, to form a question-and-answer knowledge base.

[0070] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the detailed description of a large model question answering method above, which will not be repeated here.

[0071] The large model question answering device provided in this application embodiment belongs to the same concept as the large model question answering method in the above embodiment. Any method provided in the large model question answering method embodiment can be run on the large model question answering device. For details of its implementation process, please refer to the large model question answering method embodiment, which will not be repeated here.

[0072] It should be noted that, for the purposes of this application, those skilled in the art will understand that all or part of the processes of the embodiments of this application can be implemented by a computer program controlling related hardware. The computer program can be stored in a computer-readable storage medium, such as a memory, and executed by at least one processor. During execution, it can include the process of an embodiment of a large-scale question-answering method. The computer-readable storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), etc.

[0073] It is understood that in the specific implementation of this application, user information, such as application usage behavior data, logs and other related data, is involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0074] For a large-scale question-and-answer device according to an embodiment of this application, its functional modules can be integrated into a single processing chip, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer device, such as a processor or memory, or in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0075] The foregoing has provided a detailed description of a large-scale question-answering method, apparatus, computer device, and medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A large-scale question-answering method, characterized in that, include: S101. Vectorize the user problem data to obtain vectorized problem data; S102. The trained keyword extraction model is used to extract and classify the vectorized question data, and generate question-related tags. S103. Based on the problem-related tags, perform similarity matching from a preset industry model library, and select at least two target industry models based on the matching results; S104. Analyze the user problem data using the selected target industry model to generate the corresponding model analysis results; S105. By using a general large model combined with a preset sharing mechanism, the analysis results of all models are fused to obtain a preliminary answer to the question; S106. Obtain user feedback on the preliminary answer to the question. If the feedback meets the preset criteria, output the preliminary answer to the question as the final answer result. If the evaluation feedback does not meet the preset standard, supplement the problem information based on the evaluation feedback and return to S101 to re-execute the subsequent process.

2. The large-scale question-answering method according to claim 1, characterized in that, S101 includes: Break down user question data into multiple semantic units; Each semantic unit is converted into a corresponding semantic vector through the embedding model; All semantic vectors are weighted and fused to obtain vectorized problem data.

3. The large-scale question-answering method according to claim 1, characterized in that, S102 includes: Collect a training dataset containing multi-domain questions and corresponding domain labels, and use the training dataset to train the initial keyword extraction model to obtain the trained keyword extraction model. The key feature words in the vectorized question data are identified by the large model using the keywords. The key feature words are matched according to the preset classification rules to generate question-related tags corresponding to the domain of the user's question.

4. The large-scale question-answering method according to claim 1, characterized in that, S103 includes: Retrieve all industry models in the preset industry model library and the industry model tags associated with each industry model; The embedding model is used to vectorize the problem-related labels and the model labels of each industry to obtain problem label vectors and model label vector sets. Calculate the similarity between the label vector and each model label vector in the model label vector set, and select industry models with similarity greater than a preset similarity threshold as candidate industry models; The softmax algorithm is used to calculate the weights of each candidate industry model, and the candidate industry models whose weights are greater than a preset weight threshold are determined as the target industry models.

5. The large-scale question-answering method according to claim 4, characterized in that, The preset sharing mechanism is a centralized Critic sharing mechanism, and S105 includes: The centralized Critic sharing mechanism is initiated using a general large model, and the model analysis results and corresponding weights output by all target industry models are received. The weights of each target industry model are adjusted based on the centralized Critic sharing mechanism, and the same or similar solution points in the corresponding model analysis results are merged. By logically organizing and linguistically structuring the key points of the integrated answer, a preliminary response to the question is obtained.

6. The large-scale question-answering method according to claim 5, characterized in that, The centralized Critic sharing mechanism adjusts the weights of each target industry model and merges similar or identical solution points in the corresponding model analysis results, including: The integrity of the analytical results of each model is verified according to the preset quality assessment rules. If one of the model's analysis results contains errors or is missing key information, the corresponding model's analysis result will be removed, and the weights of the remaining target industry models will be readjusted.

7. The large-scale question-answering method according to claim 1, characterized in that, Also includes: The user question data, vectorized question data, selected target industry model, model analysis results, preliminary question answer results and evaluation feedback are stored for each question and answer process to form a question and answer knowledge base.

8. A large-scale question-and-answer device, characterized in that, include: The vectorization unit is used to vectorize user question data to obtain vectorized question data; Extract classification units, used to extract and classify questions from the vectorized question data, and generate question-related labels; The matching unit is used to perform similarity matching from a preset industry model library based on the problem-related tags, and select at least two target industry models based on the matching results; The parsing unit is used to parse the user question data using the selected target industry model and generate the corresponding model parsing results; The fusion unit is used to fuse the analysis results of all models using a general large model combined with a preset sharing mechanism to obtain a preliminary answer to the question; An evaluation unit is used to obtain user feedback on the preliminary answer to the question. If the feedback meets the preset criteria, the preliminary answer to the question is output as the final answer result. If the evaluation feedback does not meet the preset standard, the problem information is supplemented based on the evaluation feedback and returned to the vectorization unit to re-execute the subsequent process.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the large model question answering method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the large model question-answering method as described in any one of claims 1 to 7.

Images