Question query method, electronic device, storage medium and computer program product

Abstract

The invention discloses a question query method, electronic equipment, a storage medium and a computer program product, and relates to the technical field of artificial intelligence, and the method comprises the following steps: receiving a multi-modal query question, and converting the multi-modal query question into query subtasks; for any query subtask in the query subtasks, allocating an optimal query model for the query subtask based on a preset model capability portrait library, and processing the query subtask based on the optimal query model to obtain a processing result, the effect of processing the query subtask by the optimal query model is better than the effect of processing the query subtask by other application models in the preset model capability portrait library; and correcting logic conflicts between the processing results corresponding to the query subtasks to obtain a query result which is used for solving the multi-modal query problem. The technical problem that a current problem query scheme based on a single large model is poor in accuracy is solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a problem query method, electronic device, storage medium, and computer program product. Background Technology

[0002] With the rapid development of artificial intelligence technologies such as large language models, multimodal interaction (such as voice, text, and gesture) has become the mainstream trend in human-computer interaction. However, in practical applications, a single user query often implies multi-dimensional and complex needs, such as a question involving both medical and multimedia recommendations. This places higher demands on the system's accurate response. Current mainstream solutions use a single large model to statically process all questions. However, due to the insufficient depth and expertise of a single large model in specific vertical fields (such as medicine and law), the accuracy of the query is affected. Therefore, current question query solutions based on a single large model suffer from poor accuracy.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide a question query method, electronic device, storage medium, and computer program product, which aims to solve the technical problem that current question query schemes based on a single large model have poor accuracy.

[0005] To achieve the above objectives, this application proposes a problem query method, which includes: Receive a multimodal query problem and transform it into various query subtasks; For any one of the query subtasks, an optimal query model is assigned to the query subtask based on a preset model capability profile library, and the query subtask is processed based on the optimal query model to obtain a processing result. The optimal query model performs better than other application models in the preset model capability profile library in processing the query subtask. The logical conflicts between the processing results corresponding to each query subtask are corrected to obtain the query result, wherein the query result is used to solve the multimodal query problem.

[0006] Optionally, the step of transforming the multimodal query problem into various query subtasks includes: The non-text information in the multimodal query question is converted into text form, and the non-text information in text form and the text information in the multimodal query question are used as the question description text; Intent recognition is performed on the problem description text to obtain the intent information represented by the problem description text; Extract entity information from the problem description text, and decompose the multimodal query problem based on the intent information and the entity information to obtain each query subtask.

[0007] Optionally, the preset model capability profile library includes each preset large model and the capability profile corresponding to each preset large model. The capability profile includes the domain label set, domain accuracy score, illusion probability, and response latency of the preset large model corresponding to the capability profile. The domain label set is used to characterize the applicable domain of the preset large model. The domain accuracy score is used to characterize the answer accuracy of the preset large model in the preset domain. The illusion probability is used to characterize the probability of the preset large model exhibiting model illusion. The response latency is used to characterize the average response time of the preset large model.

[0008] Optionally, the step of allocating the optimal query model to the query subtask based on the preset model capability profile library includes: The capability profiles of each preset large model are obtained based on the preset model capability profile library, wherein the preset domain is the domain corresponding to the query subtask; The matching degree between each preset large model and the query subtask is calculated based on the illusion probability, the response delay, the domain accuracy score, and the preset weight coefficient. The preset large model with the highest matching degree among the preset large models is taken as the optimal query model corresponding to the query subtask.

[0009] Optionally, the step of correcting logical conflicts between the processing results corresponding to each query subtask to obtain the query result includes: Detect whether there are logical conflicts between the processing results corresponding to each query subtask; In the event of a logical conflict between the processing results corresponding to each query subtask, each conflicting task, conflict description, and correction suggestion corresponding to the logical conflict are determined, wherein the other query subtasks besides the conflicting tasks are non-conflicting tasks. Based on the conflict tasks, the conflict descriptions, and the correction suggestions, update the processing results corresponding to each conflict task, and use the processing results corresponding to each conflict task and the processing results corresponding to the non-conflict tasks as the new processing results corresponding to each query subtask. Based on the processing results corresponding to each of the new query subtasks, return to the step of detecting whether there is a logical conflict between the processing results corresponding to each of the query subtasks, and record the number of returns; If there is no logical conflict between the processing results corresponding to each query subtask, the processing results corresponding to each query subtask are integrated into a query result; When the number of returns reaches the preset maximum number, the processing results of the high-priority tasks in each conflicting task, the processing results of the non-conflicting tasks, and the conflict descriptions are integrated into a query result. The high-priority tasks have a higher priority than other conflicting tasks in each conflicting task. The priority is determined by the preset domain priority and the domain corresponding to each conflicting task.

[0010] Optionally, the step of detecting whether there is a logical conflict between the processing results corresponding to each query subtask includes: The processing results corresponding to each query subtask are input into a preset logic verification model. Based on the preset logic verification model, a preset knowledge base, and preset verification instructions, the processing results corresponding to each query subtask are logically verified to obtain a verification result. The verification result includes a conclusion that there is / is no logical conflict between the processing results corresponding to each query subtask. If there is a logical conflict between the processing results corresponding to each query subtask, the verification result also includes the conflict description and the correction suggestion.

[0011] Optionally, the step of updating the processing results corresponding to each conflict task based on the conflict task, the conflict description, and the correction suggestion includes: The priority of each conflicting task is determined based on the preset domain priority and the domain corresponding to each conflicting task, and the first query subtask with the highest priority is obtained. Among the conflicting tasks, the other conflicting tasks besides the first query subtask are the second query subtasks. The processing results of the second query subtask, the first query subtask, the second query subtask, the conflict description, and the correction suggestion are input into the optimal query model corresponding to the second query subtask to obtain the updated processing result of the second query subtask. The updated processing result of the second query subtask and the processing result of the first query subtask are used as the corresponding processing results of each conflicting task.

[0012] Furthermore, to achieve the above objectives, this application also proposes a problem query device, which includes: The task decomposition module is used to receive multimodal query questions and transform the multimodal query questions into various query subtasks. The task processing module is used to assign an optimal query model to any one of the query subtasks based on a preset model capability profile library, and process the query subtask based on the optimal query model to obtain a processing result. The optimal query model performs better than other application models in the preset model capability profile library in processing the query subtask. The conflict correction module is used to correct logical conflicts between the processing results corresponding to each query subtask to obtain query results, wherein the query results are used to resolve the multimodal query problem.

[0013] In addition, to achieve the above objectives, this application also proposes an electronic device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the problem query method described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the problem query method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the problem query method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: The system receives multimodal query questions and decomposes complex multimodal query problems into various query subtasks. This transforms fuzzy, comprehensive problems into precise, professionally manageable technical tasks, reducing the complexity and uncertainty of subsequent processing. Based on a pre-defined model capability profile library, it assigns optimal query models to query subtasks. This allows for real-time and precise scheduling of the most capable expert model (the optimal query model) within the specific vertical domain, tailored to the fine-grained characteristics of each subtask. The system processes subtasks based on this optimal model, significantly improving the accuracy of each subtask and ensuring high-quality overall response. Logical conflicts between the processing results of each subtask are corrected to obtain the final query results. By introducing a global logical consistency verification process, the system collaboratively verifies and resolves conflicts between the outputs of multiple models (i.e., the processing results of each subtask). This effectively identifies and corrects potential errors or contradictory outputs from individual models, ultimately outputting collaboratively verified, logically consistent query results. This greatly enhances the overall reliability and user trust of the system response. Compared to current problem query solutions based on a single large model, this application's solution decomposes the problem into multiple sub-tasks and assigns the optimal query model to each sub-task, integrating the domain expertise of each model and introducing an effective cross-validation mechanism to correct logical conflicts in the output results of each model. This significantly improves the accuracy, professionalism, and reliability of the response, overcoming the shortcomings of poor accuracy of a single large model. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the problem query method for this application. Figure 2 This is an overall flowchart of the problem query method provided in the first embodiment of this application; Figure 3 This is a flowchart illustrating the second embodiment of the problem query method of this application; Figure 4 This is a schematic diagram of the module structure of the problem query device for this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the problem query method in this application embodiment.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Current mainstream solutions use a single large model to statically process all problems. However, because a single large model lacks sufficient in-depth knowledge and expertise in specific vertical fields (such as medicine and law), it can affect the accuracy of queries.

[0024] This application provides a solution that receives multimodal query problems and decomposes complex multimodal query problems into various query sub-tasks. This transforms fuzzy, comprehensive problems into precise, professionally manageable technical tasks, reducing the complexity and uncertainty of subsequent processing. Based on a pre-defined model capability profile library, the optimal query model is assigned to each query sub-task. This allows for real-time and precise scheduling of the most capable expert model in the vertical domain (i.e., the optimal query model) according to the fine-grained characteristics of each query sub-task. The optimal query model is then used to process the query sub-tasks, significantly improving the accuracy of each sub-task and ensuring high-quality overall response. Logical conflicts between the processing results of each query sub-task are corrected to obtain the query results. By introducing a global logical consistency verification process, the outputs of multiple models (i.e., the processing results of each query sub-task) are collaboratively verified and conflict-resolved. This effectively identifies and corrects errors or contradictory outputs that may arise from individual models, ultimately outputting collaboratively verified, logically consistent query results, greatly improving the overall reliability and user trust of the system response. Compared to current problem query solutions based on a single large model, this application's solution decomposes the problem into multiple sub-tasks and assigns the optimal query model to each sub-task, integrating the domain expertise of each model and introducing an effective cross-validation mechanism to correct logical conflicts in the output results of each model. This significantly improves the accuracy, professionalism, and reliability of the response, overcoming the shortcomings of poor accuracy of a single large model.

[0025] It should be noted that the execution subject of each embodiment of the problem query method of this application can be an electronic device with data processing, network communication and program running functions, such as a tablet computer, personal computer, mobile phone, etc.

[0026] Reference Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the problem query method of this application.

[0027] In this embodiment, the problem query method includes steps S10 to S30: Step S10: Receive the multimodal query problem and transform it into various query subtasks; Multimodal query problems are query requests that include multiple modal information (such as text, images, voice, etc., combinations of two or more modalities). For example, a user uploads an image and attaches a voice description, asking for information about a certain object in the image.

[0028] Query subtasks are independent tasks derived from the multimodal query problem. A query subtask is used to handle a subproblem in the multimodal query problem. Multiple query subtasks are interconnected and together represent the goal that the multimodal query problem aims to achieve.

[0029] Understandably, by decomposing the multimodal query problem into multiple independent query subtasks, the complex problem that was originally handled by a single model is broken down into subtasks handled by multiple models respectively. This avoids factual errors and logical fallacies caused by the limitations of the breadth and depth of knowledge of a single model, lays the foundation for subsequent scheduling, and thus improves the accuracy of the system response.

[0030] Step S20: For any query subtask in each query subtask, the optimal query model is assigned to the query subtask based on the preset model capability profile library, and the query subtask is processed based on the optimal query model to obtain the processing result. The optimal query model performs better than other application models in the preset model capability profile library in processing the query subtask. The preset model capability profile library is a pre-built database containing information on various models and their capability characteristics. For each query subtask, there is a corresponding optimal query model in the preset model capability profile library. This optimal query model achieves better processing results compared to other application models in the library when handling the query subtask. After the optimal query model processes the query subtask, it will produce a corresponding processing result, which is the specific information or answer provided for that query subtask.

[0031] In one feasible implementation, the preset model capability profile library includes each preset large model and the capability profile corresponding to each preset large model. The capability profile includes the domain label set, domain accuracy score, illusion probability, and response latency of the preset large model corresponding to the capability profile. The domain label set is used to characterize the applicable domain of the preset large model. The domain accuracy score is used to characterize the answer accuracy of the preset large model in the preset domain. The illusion probability is used to characterize the probability of the preset large model exhibiting model illusion. The response latency is used to characterize the average response time of the preset large model.

[0032] Pre-set large models are pre-trained large models with capabilities such as language processing and knowledge reasoning. Examples include M_medical, which excels in the field of medical health; M_audio, which excels in the field of music recommendation; and M_weather, which excels in the field of weather forecasting.

[0033] A capability profile is a comprehensive description of the query capabilities of a pre-defined large-scale model, encompassing its performance in various aspects. These capability profiles are continuously updated through offline benchmarking and online performance analysis. Offline benchmarking refers to a series of tests conducted on the pre-defined large-scale model based on a pre-defined test set before its deployment. Continuous online performance updates refer to the regular (e.g., daily or weekly) monitoring of the model's performance in actual use after deployment, and the recalculation of various metrics based on model performance to dynamically adjust the capability profile information. Model performance is represented by user satisfaction ratings. For example, if the pre-defined large-scale model receives a large amount of positive feedback on recommending classical music, its accuracy score in the "Music - Classical" subdomain will increase accordingly; conversely, if the pre-defined large-scale model frequently receives negative feedback when recommending heavy metal music, it indicates that the model is erroneous or misleading, and its corresponding score and probability of misleading performance will be adjusted.

[0034] Furthermore, satisfaction evaluation includes both explicit and implicit feedback. For explicit feedback, user evaluation information is directly received after using a pre-defined large model, for example, through… it works" / " Buttons such as "Useless" are used to determine user satisfaction ratings. "Useful" indicates user satisfaction rating as "satisfied". "Useless" indicates that the user's satisfaction rating is unsatisfactory. For implicit feedback, satisfaction can be inferred by analyzing the user's behavior after using the preset big model. If the user ends the conversation about the current question after using the preset big model, it indicates that the user's satisfaction rating is satisfactory. If the user continues to ask follow-up questions related to the current question after using the preset big model, it indicates that the user's satisfaction rating is unsatisfactory. For example, after receiving the answer from the preset big model, does the user immediately end the conversation or ignore the answer and ask other questions (which may indicate satisfaction), or does the user continue to ask related questions (which may indicate dissatisfaction or insufficient information)?

[0035] It's worth noting that the capability profile can also be adjusted by real-time monitoring of service quality indicators (SMIs) such as response latency (L_i), throughput, and concurrent requests of the preset large model. By determining whether the SMIs provide positive or negative feedback, the model's capability profile can be adjusted accordingly. Specific adjustment methods are detailed above and will not be elaborated further here. When determining whether the SMIs provide positive or negative feedback, a preset judge model can be used for analysis and evaluation. Furthermore, the preset judge model can be used independently to periodically or randomly audit the output of the preset large model to determine its accuracy and the presence of any illusions, and the capability profile can be adjusted based on the assessment results. The judge model is a large language model with analytical reasoning capabilities, which can be used for model reliability verification.

[0036] A domain label set is a collection of labels used to identify the applicable domains of a pre-defined large model. Each label represents a knowledge domain, such as {medical, music, history, programming, ...}. The applicable domain refers to the knowledge domain or application scenario in which the pre-defined large model can handle the corresponding problem or task. These applicable domains are determined based on the characteristics of the model's training data and the model's performance. For example, if music-related data is primarily used during model training, then the trained model will primarily be applied to the music domain.

[0037] Domain accuracy score is a quantitative assessment of the accuracy of a pre-defined large model's answers within a pre-defined domain. Different pre-defined domains have different accuracy scores; a higher score indicates a higher accuracy rate and stronger capability in that domain. This score can be calculated using historical test data. A pre-defined domain is a pre-defined area used to evaluate the model's capabilities during the capability assessment and capability profile construction of the pre-defined large model. It can be selected based on the user's actual application needs and the distribution of the model's training data. Answer accuracy refers to the degree to which the pre-defined large model's answers to questions match the standard answer or facts.

[0038] The illusion probability is used to measure the likelihood of model illusion in a pre-defined large model. The higher the illusion probability, the greater the possibility that the model generates inaccurate or unreasonable content. Model illusion is a phenomenon that occurs in the process of generating content by a pre-defined large model, that is, the generated content is inconsistent with the facts, lacks logical basis, or contains fictitious information.

[0039] Response latency refers to the average response time required for a pre-defined large model to generate and output a response from the moment it receives input information. It characterizes the processing speed and efficiency of the pre-defined large model; a shorter response latency indicates that the model can react to input information and provide an answer more quickly. Average response time is the average response latency of the pre-defined large model during multiple processing sessions of input information and response generation. It comprehensively considers the processing speed of the pre-defined large model under different conditions and can more fully reflect the response efficiency of the pre-defined large model.

[0040] Understandably, by constructing a pre-defined model capability profile library, integrating various pre-defined large models and their capability profiles—including domain label sets, domain accuracy scores, illusion probabilities, and response latency—the problem of unclear information in a single large model is solved. This allows for accurate matching of applicable models and evaluation of response accuracy during queries, achieving precise adaptation and effectively improving response accuracy compared to queries using a single large model. By dynamically assigning the best-performing query model from the pre-defined model capability profile library to each query subtask, the technical problem of static scheduling strategies failing to accurately match domain expert models is solved. This improves the professionalism and accuracy of the processing results of each subtask, overcoming the poor accuracy caused by insufficient domain knowledge depth in a single large model, and laying a solid foundation for high accuracy in the final response.

[0041] In one feasible implementation, step S20, which involves assigning the optimal query model to the query subtask based on a preset model capability profile library, includes: Step S201: Obtain the capability profile of each preset large model based on the preset model capability profile library, wherein the preset domain is the domain corresponding to the query sub-task; The preset domain is the domain corresponding to the query subtask, indicating that the domain accuracy score in the capability profile is a preset large model for the domain corresponding to the query subtask.

[0042] Step S202: Calculate the matching degree between each preset large model and the query subtask based on the hallucination probability, response latency, domain accuracy score and preset weight coefficient; Step S203: Select the preset large model with the highest matching degree among the preset large models as the optimal query model corresponding to the query subtask.

[0043] Alternatively, the matching degree calculation formula can be:

[0044] in, For the pre-set large model For query subtasks The degree of matching, For the pre-set large model In query subtasks Corresponding fields Domain accuracy score, For the pre-set large model The probability of hallucination, For the pre-set large model Response delay, , , These are preset weighting coefficients for domain accuracy score, illusion probability, and response latency, which can be adjusted according to the user's actual needs to balance accuracy, reliability, and response speed. Based on the matching degree calculation formula, the matching degree between each preset large model and the query subtask can be obtained, and then the preset large model with the highest matching degree can be selected as the optimal query model for the query subtask.

[0045] Understandably, by acquiring capability profiles that include key indicators such as domain accuracy scores and illusion probabilities, a data foundation is provided for accurate evaluation. By comprehensively calculating the matching degree of each model through preset weight coefficients, the model with the highest matching degree is determined as the optimal query model. This ensures that each subtask is handled by the model with the best accuracy processing effect in the corresponding domain, thereby improving the professionalism and accuracy of the processing and effectively overcoming the defect of poor accuracy caused by insufficient domain knowledge depth of a single large model.

[0046] Step S30: Correct the logical conflicts between the processing results corresponding to each query subtask to obtain the query results, which are used to solve the multimodal query problem.

[0047] Logical conflict refers to a situation where there are contradictions, illogicalities, or non-compliance with preset logical rules among the processing results corresponding to various query subtasks. Preset logical rules can include inconsistent factual information, contradictory reasoning processes, and disordered causal relationships.

[0048] It should be noted that, since the multimodal query problem is broken down into multiple query subtasks for processing, logical inconsistencies or contradictions (i.e., logical conflicts) may occur between the processing results of different query subtasks. Therefore, it is necessary to correct these logical conflicts. After correcting the logical conflicts between the processing results of each query subtask, the final comprehensive information used to solve the original multimodal query problem is the query result.

[0049] Understandably, by correcting the logical conflicts between the processing results of each query subtask, the final query result is ensured to have overall logical consistency, effectively intercepting the propagation of errors generated by the model, thereby improving the accuracy and reliability of the output and overcoming the defect of poor accuracy caused by the limitations of a single large model due to its own system.

[0050] In one feasible implementation, step S10, which involves transforming the multimodal query problem into various query subtasks, includes: Step S101: Convert the non-text information in the multimodal query question into text form, and use the non-text information in text form and the text information in the multimodal query question as the question description text; Non-textual information refers to information in multimodal query problems that takes forms other than textual modality. This information carries semantic content in a way different from text encoding, such as images, audio, and video. Textual form refers to the text encoding form that computers can directly process and understand. The question description text is a complete text content that integrates the textual information from the multimodal query problem with the converted non-textual information. It comprehensively covers all the information involved in the user query, providing a foundation for subsequent intent recognition and task decomposition.

[0051] Optionally, when the non-text information in a multimodal query problem is speech, it can be converted into text using an Automatic Speech Recognition (ASR) model. ASR can convert the lexical content of speech into computer-readable text encoding, thereby enabling voice interaction between humans and machines.

[0052] Optionally, when the non-textual information in a multimodal query problem is gestures or facial expressions, it can be converted into text using a computer vision (CV) model. CV can recognize gestures, facial expressions, and other information in image or video data, understand the relationships and meanings between these information, and thus convert them into corresponding text encoding.

[0053] Step S102: Perform intent recognition on the problem description text to obtain the intent information represented by the problem description text; Intent recognition refers to the process of analyzing and understanding the question description text to determine the goal or purpose that a user expects to achieve when making a multimodal query. The result of intent recognition is intent information, which can be extracted from the question description text through the contextual understanding and reasoning capabilities of a large language model (LLM).

[0054] Step S103: Extract entity information from the problem description text, decompose the multimodal query problem based on intent information and entity information, and obtain each query subtask.

[0055] Entity information refers to the specific objects, things, or concepts involved in the problem description text. This entity information is usually the core focus of the user's query and is an important element for understanding the query question and breaking down the task. It can be names of people, places, organizations, things, etc.

[0056] It's important to note that entity information possesses domain relevance, task-driven nature, and information indispensability. Domain relevance refers to the association of entity information with the knowledge domain involved in the multimodal query problem. For example, in the question "What audio relaxation is recommended when experiencing abnormal heart rate?", the entity information "abnormal heart rate" is associated with the medical and health domain, while the entity information "audio relaxation" is associated with the multimedia recommendation domain. Task-driven nature means that entity information is used to trigger corresponding query subtasks. Through entity information, the processing scheme for query subtasks can be determined, i.e., which pre-defined main model should be called to handle the query subtask. For example, identifying the entity "abnormal heart rate" tells us that a medical expert model needs to be called to handle the corresponding query subtask, and identifying the entity "audio relaxation" tells us that an audio recommendation model needs to be called to handle the corresponding query subtask. Information indispensability means that without entity information, it is impossible to decompose the query subtasks. For example, if only "recommendation" is identified without "audio," it is impossible to know whether to recommend an article or a video, and further impossible to determine how to design the query subtasks, thus making decomposition impossible.

[0057] Optionally, entity information in the problem description text can be extracted using regular expressions or dictionaries containing preset rules. For example, using a pre-built medical dictionary containing disease names, when a word in the problem description text matches a word in the medical dictionary, that word is marked as a "medical entity". Alternatively, machine learning methods can be used to extract entity information from the problem description text. For instance, sequence labeling models (such as Hidden Markov Models, Conditional Random Fields, etc.) can be used to predict the entity type of each word by learning from a large amount of labeled text data, thereby obtaining the corresponding entity information. Entity information can also be extracted from the question description text using deep learning methods. For example, a fine-tuned language model (such as BERT, RoBERTa, etc.) can be used for entity extraction to obtain the corresponding entity information. This can be achieved through data preparation, model fine-tuning, and predictive inference. Data preparation involves preparing a large amount of text corpus with entity annotations as an annotation dataset. Model fine-tuning involves training a pre-trained natural language processing model on this annotation dataset, enabling the model to learn to recognize entities of the corresponding types in the annotation dataset. Predictive inference involves inputting the user's query text into the fine-tuned natural language processing model, which will output the entity label for each word, thus obtaining the corresponding entity information.

[0058] It should be noted that the decomposition of multimodal query problems is based on related knowledge domains and a pre-defined model capability profile library. Multimodal query problems often span multiple different related knowledge domains, therefore the requirements of these related knowledge domains need to be broken down so that they can be handled by specialized models for their respective domains. The pre-defined model capability profile library can determine the models that can be applied to the problem and the domains in which the models are applicable, thus allowing each subtask to determine the most suitable model.

[0059] Optionally, when decomposing a multimodal query problem, the entity information can first be analyzed using a pre-defined entity-domain association table or a pre-defined domain analysis model to determine the knowledge domain associated with the multimodal query problem. The pre-defined entity-domain association table records the knowledge domains associated with each entity; the pre-defined domain analysis model is a pre-trained large language model used to infer the corresponding knowledge domain from entity information. For example, the entity information "abnormal heart rate" belongs to the medical and health domain, while the entity information "recommended audio" belongs to the multimedia / entertainment domain. Since these two knowledge domains differ significantly, processing by models from different domains will yield more efficient and accurate results.

[0060] Furthermore, after obtaining the knowledge domain associated with the multimodal query question, the multimodal query question is decomposed by combining intent information and the domain label sets of each preset large model in the preset model capability profile library. The preset model capability profile library can clearly identify the available preset large models, and the domain label sets in it are used to characterize the applicable domains of each preset large model. Therefore, by matching the knowledge domain with the applicable domain, the preset large model corresponding to each knowledge domain can be determined. At the same time, the intent information can determine the tasks that the preset large model needs to perform, thereby obtaining each query sub-task and realizing the decomposition of the multimodal query question. For example, the user's multimodal query question is: "What audio relaxation is recommended when the heart rate is abnormal?" From this, the user's intent information (requesting audio suggestions suitable for a specific situation) and entity information ("abnormal heart rate", "audio relaxation") are identified. The preset model capability profile library includes medical models (medical and health domain) and music models (music recommendation domain). Therefore, the multimodal query question can be decomposed into: Query sub-task ST1: Analyze the health meaning of "abnormal heart rate" and provide precautions; Query sub-task ST2: Based on health suggestions, recommend suitable "relaxation audio".

[0061] For example, reasoning analysis can be performed using a collaborative reasoning engine (a pre-trained large language model used for understanding, reasoning, and problem decomposition, such as GPT-4, Claude, etc.) and corresponding prompt word templates. First, prompt word templates are constructed based on the multimodal query question, intent information, entity information, and a pre-defined model capability profile library. The extraction process for this information is described above and will not be elaborated upon here. The prompt word template includes the multimodal query question (e.g., "What audio is recommended for relaxation when heart rate is abnormal?"), decomposition objectives (represented by intent and entity information, such as "decompose the question into sub-tasks associated with each intent and entity, each sub-task clearly corresponding to a model's knowledge domain"), and a pre-defined model capability profile library (including the knowledge domains of each model, such as "M_medical: proficient in the medical and health domain, able to answer questions related to diseases and symptoms; M_audio: proficient in the music recommendation domain, able to recommend songs based on emotions or scenes; M_weather: proficient in weather forecasts"). The constructed prompt word templates are then input into the collaborative reasoning engine, which utilizes the engine's understanding, reasoning, and problem decomposition capabilities to decompose the multimodal query question.

[0062] Understandably, by unifying multimodal information into text, the standardized processing of heterogeneous data is achieved, providing a foundation for accurate analysis. By identifying the user's core intent, extracting key entities, and decomposing tasks based on this, it is ensured that each generated subtask accurately corresponds to a professional domain, thereby avoiding misunderstandings caused by confusion or mismatch in model knowledge domains and improving the accuracy of processing.

[0063] For example, refer to Figure 2 , Figure 2 This is an overall flowchart of the query method provided in the first embodiment of this application. First, a multimodal query question is received, and intent understanding is performed to parse the user's core needs. Next, task decomposition and model selection are performed, breaking down the multimodal query question into various query sub-tasks. Based on a preset model capability profile library, an optimal query model is assigned to each query sub-task. Each optimal query model processes its corresponding query sub-task in parallel to achieve task processing. Then, conflict correction is performed, verifying and correcting the logical consistency of the processing results of each model. Finally, the corrected results are integrated and output to obtain an accurate and reliable final answer, thus resolving the initial multimodal query question.

[0064] In this embodiment, firstly, multimodal queries are transformed into sub-tasks based on domain boundaries to ensure that each problem is handled by the corresponding domain; secondly, by quantitatively evaluating the model's capability profile, the model with the best processing effect is dynamically assigned to each sub-task to ensure the professionalism of each sub-result; finally, cross-validation between models is achieved through logical conflict correction to intercept the erroneous output of individual models, improve the accuracy and logical consistency of query results, and solve the accuracy defects of a single large model in cross-domain complex queries.

[0065] Based on the first embodiment described above, a second embodiment of the problem query method for this application is proposed. In this embodiment, content that is the same as or similar to that in the first embodiment can be referred to the above description, and will not be repeated hereafter. (Refer to...) Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the query method for this application. In this embodiment, step S30, which corrects logical conflicts between the processing results corresponding to each query subtask to obtain the query result, includes: Step S301: Check whether there is a logical conflict between the processing results corresponding to each query subtask; In one feasible embodiment, step S301 includes: Step A01: Input the processing results corresponding to each query subtask into the preset logic verification model. Based on the preset logic verification model, the preset knowledge base, and the preset verification instructions, perform logical verification on the processing results corresponding to each query subtask to obtain the verification results. The verification results include the conclusion that there are / are not logical conflicts between the processing results corresponding to each query subtask. If there are logical conflicts between the processing results corresponding to each query subtask, the verification results also include a conflict description and correction suggestions.

[0066] The preset logic verification model is an artificial intelligence model that performs logical analysis on input information (such as the processing results corresponding to each query subtask). It can analyze the processing results corresponding to each input query subtask based on built-in logical rules and algorithms, thereby determining whether there are logical contradictions, illogicalities, or non-compliance with preset logical relationships among these results.

[0067] The pre-defined knowledge base is a database that stores rich, accurate, and systematically organized knowledge information. This knowledge information covers basic knowledge, factual data, rules and principles in various fields, providing knowledge support for the pre-defined logic verification model to perform logical verification.

[0068] Preset verification instructions are a series of operational instructions and rules used to guide the preset logical verification model in performing logical verification on the processing results corresponding to each query subtask. These instructions specify the verification steps, methods, key points to focus on, and criteria for judging logical conflicts. For example, "Please determine whether there is a logical conflict between health advice R1 and recommended audio R2? For example, R1 suggests staying calm, while R2 contains exciting symphonic music?" Optionally, preset verification instructions can also be generated based on preset instruction templates. The preset instruction templates include basic verification content and key verification content. Basic verification content consists of the most fundamental verification logic and rules inherent in the instruction template (e.g., not violating factual basis), while key verification content (e.g., the specific logical relationships to be checked) can be dynamically generated and filled based on the nature of the query subtask and its output results. The generation and filling of key verification content can be achieved through a preset central large-scale model. This central large-scale model, by integrating and scheduling multiple specialized large-scale models, can identify key issues and potential conflicts in the query subtask.

[0069] Logical verification is a process that uses a preset logical verification model, combined with a preset knowledge base and preset verification instructions, to check and analyze the processing results of each query subtask in order to determine whether there are logical conflicts among these processing results.

[0070] The verification result is the final output of the logical verification process. It indicates the logical state between the processing results of each query subtask, including two cases: First, there is no logical conflict between the processing results of each query subtask, which means that all processing results are logically reasonable and conform to the preset logical rules and common sense. Second, there is a logical conflict between the processing results of each query subtask. In this case, the verification result will further provide a description of the conflict and suggestions for correction.

[0071] A conflict description is a detailed explanation and record of detected logical conflicts, indicating the query subtask where the conflict occurred, as well as the specific form and content of the conflict. Correction suggestions are solutions and improvement measures proposed for logical conflicts. Based on the analysis and judgment of the logical conflicts, they provide specific suggestions on how to adjust the processing results of the conflicting tasks to eliminate the logical conflicts.

[0072] For example, the processing results of each query subtask are input into a preset logic verification model. Logical verification is then performed on the processing results of each query subtask based on the preset logic verification model, a preset knowledge base, and preset verification instructions. For instance, when processing a query subtask based on the optimal query model, M_medical executes query subtask ST1 to generate health advice text R1, and M_audio executes query subtask ST2, receiving R1 as context to generate an audio recommendation list R2. R1 and R2 are the processing results, which are also the inputs to the preset logic verification model. The preset logic verification model performs logical consistency verification based on its internal knowledge. Verification instructions can be such as, "Please determine if there is a logical conflict between the health advice R1 and the recommended audio R2? For example, R1 suggests staying calm, while R2 contains exciting symphonic music?" The pre-defined logic verification model outputs the verification results after analysis. The verification results include multiple cases. If there is no logical conflict between the processing results of each query subtask, it directly indicates that there is no conflict. If there is a logical conflict, the verification results will not only indicate the existence of the conflict, but also provide a description of the conflict, detailing the conflict points. For example, "Health advice R1 requires maintaining calm, but the audio recommendation list R2 contains exciting symphonies, which contradict each other in creating an emotional atmosphere." At the same time, it will also propose correction suggestions for the conflict, such as "Re-determine the audio recommendation criteria. According to the requirements of health advice R1, remove the exciting symphonies from the recommendation list and replace them with soothing light music."

[0073] Understandably, by performing automated logical verification on the outputs of each model, verifying the consistency of facts based on the preset logical verification model and the preset knowledge base, and also outputting conflict conclusions and correction suggestions, the technical problem that a single large model cannot identify and correct internal logical errors due to the lack of an external verification mechanism is solved, significantly improving the logical consistency and factual accuracy of the final query results.

[0074] Step S302: In the case of logical conflicts between the processing results of each query subtask, determine each conflicting task, conflict description and correction suggestion corresponding to the logical conflict. Among them, the other query subtasks in each query subtasks other than the conflicting tasks are non-conflicting tasks. A conflict task is a query subtask that is identified as conflict-related when logical conflicts are detected in the processing results of the corresponding query subtasks. The processing results of these tasks are contradictory or logically inconsistent.

[0075] Step S303: Update the processing results of each conflict task based on each conflict task, conflict description and correction suggestion, and take the processing results of each conflict task and the processing results of non-conflict tasks as the new processing results of each query subtask. In one feasible embodiment, step S303, which involves updating the processing results corresponding to each conflict task based on each conflict task, conflict description, and correction suggestion, includes: Step B01: Determine the priority of each conflict task based on the preset domain priority and the domain corresponding to each conflict task, and obtain the first query subtask with the highest priority. Among them, the other conflict tasks besides the first query subtask are the second query subtasks. Preset domain priorities are the priority order that is set for each domain in advance based on the importance and application needs of different domains. For example, medical advice has a higher priority than entertainment recommendations.

[0076] The first query subtask is the highest priority query subtask among all conflicting tasks. The second query subtask refers to all other conflicting tasks besides the first query subtask.

[0077] Step B02: Input the processing results of the second query subtask, the first query subtask, the second query subtask, the conflict description, and the correction suggestions into the optimal query model corresponding to the second query subtask to obtain the updated processing result of the second query subtask. Step B03: The updated processing result of the second query subtask and the processing result of the first query subtask are used as the corresponding processing results for each conflicting task.

[0078] For example, priorities are determined based on preset domain priorities and the domains corresponding to each conflicting task. For instance, in the preset domain priorities, medical advice has a higher priority than entertainment recommendations. If there are two conflicting tasks, medical and entertainment, the medical task is the first query subtask with the highest priority, and the entertainment task is the second query subtask. The processing results of the second query subtask (e.g., the entertainment task), the first query subtask (e.g., the medical task result), the second query subtask's original entertainment task result, the conflict description (e.g., "there is a logical conflict between the medical task's suggestion to maintain calm health advice and the entertainment task's recommendation of exciting symphonies"), and the correction suggestion are input into the optimal query model corresponding to the second query subtask. The model analyzes the conflict information, adjusts the processing logic, and outputs the updated processing result of the second query subtask (e.g., correcting the entertainment task to recommend soothing music). Finally, the updated processing result of the second query subtask and the processing result of the first query subtask are used together as the processing results corresponding to each conflicting task, ensuring that the processing results of the conflicting tasks are logically consistent and meet the preset domain priority requirements.

[0079] Understandably, by prioritizing domains to determine the baseline task (first query subtask), the stability of the core domain conclusions is ensured. Conflicting contexts and correction suggestions are fed back to the optimal query model of the secondary priority task (second query subtask), driving it to make targeted corrections. Using the high-priority expert output as a reliable benchmark, the low-priority model is guided to self-correct, eliminating logical contradictions between results and ensuring the overall consistency and factual accuracy of cross-domain query results.

[0080] Step S304: Based on the processing results corresponding to each new query subtask, return the step of checking whether there is a logical conflict between the processing results corresponding to each query subtask, and record the number of returns; The return count refers to the number of times the "check for logical conflicts between the processing results of each query subtask" step is executed during the logical conflict correction process. It controls the number of loops in the correction process to avoid infinite loops. Each time the processing result of a conflicting task is updated, a new logical conflict needs to be checked again. Each time such a check and return operation is performed, the return count increases by 1.

[0081] Step S305: If there is no logical conflict between the processing results corresponding to each query subtask, integrate the processing results corresponding to each query subtask into a query result. Step S306: When the number of returns reaches the preset maximum number of times, the processing results corresponding to the high-priority tasks, the processing results corresponding to the non-conflicting tasks, and the conflict descriptions in each conflict task are integrated into the query results. Among them, the priority of the high-priority tasks is higher than that of other conflict tasks in each conflict task. The priority is determined by the preset domain priority and the domain corresponding to each conflict task.

[0082] High-priority tasks are the highest-priority query subtasks identified from the conflicting tasks when the maximum number of returns (e.g., 5) has been reached, but logical conflicts still exist. High-priority tasks have higher priority than other conflicting tasks within the same group, and their processing results are given priority in the final query result integration. Priority is determined by a preset domain priority and the corresponding domain of each conflicting task. By identifying the domain of each conflicting task and determining the priority order of the conflicting tasks within the preset domain priority, the high-priority task can be identified.

[0083] It should be noted that when the preset maximum number of returns is reached, the processing results of higher-priority tasks are prioritized. This ensures that the most important needs are addressed. For example, medical needs are related to personal health and have a higher priority than entertainment needs; therefore, medical needs are the more important needs. Compared to integrating the processing results of each query subtask into a single query result, while this method cannot guarantee complete satisfaction of multimodal query problems, it ensures that the most important needs are met. Although it is slightly less effective than query results without logical conflicts, it can still be used to solve multimodal query problems.

[0084] In this embodiment, by detecting logical conflicts between results, the master-slave relationship of conflicting tasks is determined based on preset domain priorities, and the results of low-priority tasks are corrected. By introducing a maximum number of iterations, the process is kept controllable. When conflicts cannot be completely eliminated, the results are integrated according to domain priorities. This solves the problem of inconsistent outputs that may occur when multiple models work together, effectively eliminates logical fallacies caused by knowledge differences between models or illusions of individual models, and improves the logical consistency and accuracy of the system response.

[0085] This application also provides a problem query device; please refer to... Figure 4 The problem query device includes: The task decomposition module 10 is used to receive multimodal query problems and transform them into various query subtasks. The task processing module 20 is used to assign the optimal query model to any query subtask in each query subtask based on the preset model capability profile library, and process the query subtask based on the optimal query model to obtain the processing result. The optimal query model is better than other application models in the preset model capability profile library in processing the query subtask. The conflict correction module 30 is used to correct logical conflicts between the processing results of each query subtask to obtain query results, which are used to solve multimodal query problems.

[0086] The problem query device provided in this application adopts the problem query method in the above embodiments. Compared with the prior art, the beneficial effects of the problem query device provided in this application are the same as those of the problem query method provided in the above embodiments. Moreover, the other technical features in the problem query device are the same as those disclosed in the method of the above embodiments, and will not be repeated here.

[0087] This application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the problem query method in Embodiment 1 above.

[0088] The following is for reference. Figure 5 The diagram illustrates a structural schematic of an electronic device suitable for implementing embodiments of this application. The electronic devices in these embodiments may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0089] like Figure 5As shown, the electronic device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagrams show electronic devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0090] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0091] Compared with the prior art, the beneficial effects of the electronic device provided in this application embodiment are the same as the beneficial effects of the problem query method provided in the above embodiment, and other technical features in the electronic device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0093] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the problem query method described in the above embodiments.

[0094] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0095] The aforementioned computer-readable storage medium may be included in an electronic device or may exist independently without being assembled into an electronic device.

[0096] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the functions defined in the methods of the embodiments disclosed in this application.

[0097] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0098] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0099] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0100] The readable storage medium provided in this application embodiment is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described problem query method. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application embodiment are the same as the beneficial effects of the problem query method provided in the above-described embodiments, and will not be repeated here.

[0101] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the problem query method described above.

[0102] Compared with the prior art, the beneficial effects of the computer program product provided in this application embodiment are the same as the beneficial effects of the problem query method provided in the above embodiment, and will not be repeated here.

[0103] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for querying questions, characterized in that, The problem query method includes: Receive a multimodal query problem and transform it into various query subtasks; For any one of the query subtasks, an optimal query model is assigned to the query subtask based on a preset model capability profile library, and the query subtask is processed based on the optimal query model to obtain a processing result. The optimal query model performs better than other application models in the preset model capability profile library in processing the query subtask. The logical conflicts between the processing results corresponding to each query subtask are corrected to obtain the query result, wherein the query result is used to solve the multimodal query problem.

2. The problem query method as described in claim 1, characterized in that, The steps of transforming the multimodal query problem into various query subtasks include: The non-text information in the multimodal query question is converted into text form, and the non-text information in text form and the text information in the multimodal query question are used as the question description text; Intent recognition is performed on the problem description text to obtain the intent information represented by the problem description text; Extract entity information from the problem description text, and decompose the multimodal query problem based on the intent information and the entity information to obtain each query subtask.

3. The problem query method as described in claim 1, characterized in that, The preset model capability profile library includes each preset large model and the capability profile corresponding to each preset large model. The capability profile includes the domain label set, domain accuracy score, illusion probability, and response latency of the preset large model corresponding to the capability profile. The domain label set is used to characterize the applicable domain of the preset large model. The domain accuracy score is used to characterize the answer accuracy of the preset large model in the preset domain. The illusion probability is used to characterize the probability of the preset large model exhibiting model illusion. The response latency is used to characterize the average response time of the preset large model.

4. The problem query method as described in claim 3, characterized in that, The step of assigning the optimal query model to the query subtask based on the preset model capability profile library includes: The capability profiles of each preset large model are obtained based on the preset model capability profile library, wherein the preset domain is the domain corresponding to the query subtask; The matching degree between each preset large model and the query subtask is calculated based on the illusion probability, the response delay, the domain accuracy score, and the preset weight coefficient. The preset large model with the highest matching degree among the preset large models is taken as the optimal query model corresponding to the query subtask.

5. The problem query method as described in claim 1, characterized in that, The step of correcting logical conflicts between the processing results corresponding to each query subtask to obtain the query result includes: Detect whether there are logical conflicts between the processing results corresponding to each query subtask; In the event of a logical conflict between the processing results corresponding to each query subtask, each conflicting task, conflict description, and correction suggestion corresponding to the logical conflict are determined, wherein the other query subtasks besides the conflicting tasks are non-conflicting tasks. Based on the conflict tasks, the conflict descriptions, and the correction suggestions, update the processing results corresponding to each conflict task, and use the processing results corresponding to each conflict task and the processing results corresponding to the non-conflict tasks as the new processing results corresponding to each query subtask. Based on the processing results corresponding to each of the new query subtasks, return to the step of detecting whether there is a logical conflict between the processing results corresponding to each of the query subtasks, and record the number of returns; If there is no logical conflict between the processing results corresponding to each query subtask, the processing results corresponding to each query subtask are integrated into a query result; When the number of returns reaches the preset maximum number, the processing results of the high-priority tasks in each conflicting task, the processing results of the non-conflicting tasks, and the conflict descriptions are integrated into a query result. The high-priority tasks have a higher priority than other conflicting tasks in each conflicting task. The priority is determined by the preset domain priority and the domain corresponding to each conflicting task.

6. The problem query method as described in claim 5, characterized in that, The step of detecting whether there is a logical conflict between the processing results corresponding to each query subtask includes: The processing results corresponding to each query subtask are input into a preset logic verification model. Based on the preset logic verification model, a preset knowledge base, and preset verification instructions, the processing results corresponding to each query subtask are logically verified to obtain a verification result. The verification result includes a conclusion that there is / is no logical conflict between the processing results corresponding to each query subtask. If there is a logical conflict between the processing results corresponding to each query subtask, the verification result also includes the conflict description and the correction suggestion.

7. The problem query method as described in claim 5, characterized in that, The step of updating the processing results corresponding to each conflict task based on the conflict task, the conflict description, and the correction suggestion includes: The priority of each conflicting task is determined based on the preset domain priority and the domain corresponding to each conflicting task, and the first query subtask with the highest priority is obtained. Among the conflicting tasks, the other conflicting tasks besides the first query subtask are the second query subtasks. The processing results of the second query subtask, the first query subtask, the second query subtask, the conflict description, and the correction suggestion are input into the optimal query model corresponding to the second query subtask to obtain the updated processing result of the second query subtask. The updated processing result of the second query subtask and the processing result of the first query subtask are used as the corresponding processing results of each conflicting task.

8. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the problem query method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the problem query method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the problem query method as described in any one of claims 1 to 7.

Images