Online question answering method based on artificial intelligence and related device

Abstract

The invention provides an online question answering method based on artificial intelligence and a related device. The method comprises the following steps: for a received user question, generating first answer track data by a question and answer large model; evaluating first reasoning process data in the first answer trajectory data, and outputting error attribution information aiming at the first answer trajectory data under the condition that reasoning errors exist in the first reasoning process data; according to the first answer trajectory data and the error attribution information, calling the question and answer large model to obtain second answer trajectory data for the user question; and performing reasoning process evaluation on second reasoning process data in the second answer trajectory data by the large evaluation model, and feeding back second answer data to the user client sending the user question under the condition that the second answer trajectory data is determined to be correct answer trajectory data. The reliability of the online question and answer content can be improved to a certain degree.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and related apparatus for generating product description text. Background Technology

[0002] In the context of internet healthcare, online question-and-answer systems typically rely on large language models to automatically generate answers, providing users with instant consultation services without the need for human doctors.

[0003] However, due to limitations in the model's reasoning ability and the complexity of medical issues, question-answering models are prone to illogical reasoning processes, biased understanding of facts, or non-compliant medication recommendations when generating answers. Summary of the Invention

[0004] In view of this, one or more embodiments of this application provide a method and related apparatus for generating product description text, which can improve the reliability of online Q&A content to a certain extent.

[0005] In a first aspect, one or more embodiments of this application propose an online question-answering method based on artificial intelligence, comprising: generating first answer trajectory data by a question-answering big model for a received user question; wherein the first answer trajectory data includes first reasoning process data and first answer data of the question-answering big model for the user question; the first reasoning process data is used to express the reasoning process that leads to the first answer data; evaluating the first reasoning process data by a judging big model, and outputting error attribution information for the first answer trajectory data if the first reasoning process data is determined to have a reasoning error; wherein the error attribution information is used to express the reason why the question-answering big model has an error in the reasoning process; invoking the question-answering big model according to the first answer trajectory data and the error attribution information to obtain second answer trajectory data for the user question; wherein the second answer trajectory data includes second reasoning process data and second answer data for the user question; the second reasoning process data is used to express the reasoning process that leads to the second answer data; evaluating the second reasoning process data by the judging big model, and feeding back the second answer data to the user client that issued the user question if the second answer trajectory data is determined to be correct answer trajectory data.

[0006] Secondly, one or more embodiments of this application propose a model training method for training a large question-answering model as described above. The method includes: acquiring erroneous answer trajectory data generated by the large question-answering model in response to a user's question; wherein the erroneous answer trajectory data indicates that the derived answer data contains errors; generating corrected answer trajectory data by calling a specified higher-order large model based on the erroneous answer trajectory data; wherein the corrected answer trajectory data includes corrected reasoning process data and corrected answer data; the corrected reasoning process data, in addition to expressing the correct reasoning content that leads to the correct corrected answer data, also includes reflective content regarding the errors in the reasoning process in the erroneous answer trajectory data; the model size of the specified higher-order large model is larger than the model size of the large question-answering model; and training the large question-answering model based on the corrected answer trajectory data as samples.

[0007] Thirdly, one or more embodiments of this application propose an online question-answering device based on artificial intelligence, comprising: a receiving module, configured to generate first answer trajectory data from a question-answering big model for a received user question; wherein the first answer trajectory data includes first reasoning process data and first answer data of the question-answering big model for the user question; the first reasoning process data is used to express the reasoning process leading to the first answer data; an output module, configured to evaluate the first reasoning process data by a judging big model, and output error attribution information for the first answer trajectory data if the first reasoning process data is determined to have a reasoning error; wherein the error attribution information is used to express the reason for the error in the reasoning process of the question-answering big model; a calling module, configured to call the question-answering big model according to the first answer trajectory data and the error attribution information to obtain second answer trajectory data for the user question; wherein the second answer trajectory data includes second reasoning process data and second answer data for the user question; the second reasoning process data is used to express the reasoning process leading to the second answer data; and a feedback module, configured to evaluate the second reasoning process data by the judging big model, and feed back the second answer data to the user client that issued the user question if the second answer trajectory data is determined to be correct answer trajectory data.

[0008] Fourthly, one or more embodiments of this application propose a model training apparatus, comprising: an acquisition module, configured to acquire erroneous answer trajectory data generated by a large question-and-answer model in response to a user's question; wherein the erroneous answer trajectory data indicates that the reasoning process leading to the answer data contains errors; a generation module, configured to generate corrected answer trajectory data by calling a specified high-order large model based on the erroneous answer trajectory data; wherein the corrected answer trajectory data includes corrected reasoning process data and corrected answer data; the corrected reasoning process data, in addition to expressing the correct reasoning content leading to the correct corrected answer data, also includes reflection content regarding the errors in the reasoning process in the erroneous answer trajectory data; the model size of the specified high-order large model is larger than the model size of the large question-and-answer model; and a training module, configured to train the large question-and-answer model based on the corrected answer trajectory data as samples.

[0009] Fifthly, one or more embodiments of this application provide a computer device including a memory and a processor, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the method as described above.

[0010] In a sixth aspect, one or more embodiments of this application provide a computer program product including computer instructions that, when executed by a processor, implement the method as described above.

[0011] In a seventh aspect, one or more embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0012] As can be seen from the above embodiments, multiple embodiments of this application generate first answer trajectory data containing reasoning process data by using a question-and-answer big model, and use a judgment big model to evaluate the reasoning process data of the first reasoning process in the first answer trajectory data to output error attribution information. Then, based on the first answer trajectory data and the error attribution information, the question-and-answer big model is called again to generate second answer trajectory data, and the judgment big model evaluates the second reasoning process data in the second answer trajectory data to determine its correctness. This achieves error identification and correction in the online question-and-answer reasoning process, thereby improving the reliability of online question-and-answer content. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating an application scenario of an online question-and-answer device provided in one embodiment of this application.

[0014] Figure 2 This is a flowchart illustrating an online question-and-answer method based on artificial intelligence, provided in one embodiment of this application.

[0015] Figure 3 This is a flowchart illustrating a model training method provided in one embodiment of this application.

[0016] Figure 4 This is a schematic diagram of a module of an online question-and-answer device based on artificial intelligence provided in one embodiment of this application.

[0017] Figure 5 This is a schematic diagram of a model training device provided in one embodiment of this application.

[0018] Figure 6 This is a schematic diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments.

[0020] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0021] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0022] In the field of online medical Q&A, early online Q&A systems were mostly based on rule matching, knowledge base retrieval, or traditional machine learning models. They selected responses from a pre-set set of answers based on user-inputted symptom descriptions or keywords to provide basic consultation services without the need for human doctors. However, these systems have limited ability to understand questions and struggle to handle complex, multi-round interactions, or consultations involving medication safety judgments, resulting in a significant lack of professionalism in online Q&A.

[0023] With the widespread use of large-scale pre-trained language models, some online question-answering systems have begun to directly utilize these models to automatically generate natural language answers to user questions. This has improved their ability to understand free text and the fluency of their responses, while further reducing reliance on real-time human intervention. However, when faced with specialized and high-risk medical questions, large language models are prone to incomplete reasoning chains, factual misunderstandings, or inappropriate medication recommendations. If the one-time output of a large language model is directly returned to the user as the answer, the reliability and security of the online question-answering content remain difficult to guarantee.

[0024] In summary, the relevant technologies still have problems such as difficulty in timely detection and correction of online question-answering reasoning errors, resulting in low reliability of question-answering content, which require further improvement.

[0025] In several embodiments provided in this application, the artificial intelligence-based online question-answering method can be applied to electronic devices with certain computing power and network access capabilities. This electronic device can be a desktop computer, laptop computer, tablet computer, smartphone, or a server. Specifically, the electronic device includes a processor, memory, and a network access module for network communication. The server can be an electronic device with strong data processing capabilities; of course, a server can also refer to a server cluster formed by multiple electronic devices, or a quantum server built using a quantum computer.

[0026] Please see Figure 1 This application provides an example application scenario of an AI-based online question-answering method. This method can be applied to internet-based medical consultation platforms to automatically generate and verify intelligent question-and-answer results for user health consultation content. In this example scenario, the health questions submitted by users through the online consultation portal are used as input. A large question-answering model and a large evaluation model work collaboratively to perform multiple rounds of reasoning, verification, and correction on the questions, thereby improving the reliability of the final answer.

[0027] For example, a user submits the following question through a health consultation application on the user's client: "I have been experiencing a dull pain in my upper right abdomen for the past two days, which is more pronounced after meals, and I also occasionally experience acid reflux at night. What could be the cause?" The online Q&A device can use this user question as input data to be processed and as part of the trajectory data to initiate an online Q&A process for the user's question.

[0028] The online question-and-answer device first calls upon a large question-and-answer model to generate initial response trajectory data. Since trajectory data is used to completely record the inputs and outputs of the question-and-answer process, the initial response trajectory data includes not only the initial reasoning process data and initial answer data generated by the large question-and-answer model, but also the user question itself, which identifies the object being processed. When analyzing the user question, the large question-and-answer model generates initial reasoning process data, for example: "The patient describes right upper abdominal pain, which worsens after meals and is accompanied by acid reflux; digestive system-related diseases such as gastritis, cholecystitis, or bile reflux should be considered." Subsequently, the large question-and-answer model generates initial answer data based on the above analysis, for example: "This may be related to gastritis or gallbladder-related diseases; a light diet is recommended; if symptoms continue to worsen, medical examination should be sought." The online question-and-answer device combines the user question, initial reasoning process data, and initial answer data to obtain complete initial response trajectory data, which serves as the basic data unit for subsequent evaluation.

[0029] The online question-answering device can invoke a large-scale evaluation model to assess the reasoning process of the first answer trajectory data. Since the trajectory data includes the user's question itself, the large-scale evaluation model can perform logical consistency checks based on the original content of the user's question when reviewing the first reasoning process data. For example, the large-scale evaluation model might identify omissions in the first reasoning process data, such as failing to fully integrate the typical indicative nature of right upper quadrant abdominal pain and the relationship with food intake in gallbladder disease. The large-scale evaluation model can then output error attribution information, where the error type can be "reasoning omission," and the error cause analysis can be generated by the large-scale evaluation model based on its own thought process, for example: "The reasoning did not focus on the clinical significance of right upper quadrant abdominal pain in gallbladder disease, and also failed to distinguish the symptom differences between acid reflux and bile reflux."

[0030] The online question-and-answer device can re-invoke the question-and-answer big model based on the first answer trajectory data and error attribution information to generate second answer trajectory data. Since the trajectory data also includes the user question, the question-and-answer big model can directly re-infer based on the user question and error attribution information when generating the second inference process data. For example, the second inference process data might be: "Right upper abdominal pain is closely related to eating, more likely to be cholecystitis or gallstones; acid reflux symptoms need to differentiate gastroesophageal reflux, but gallbladder-related diseases should still be the primary suspicion." The question-and-answer big model forms the second answer data based on the revised inference, such as: "It is recommended to have a liver and gallbladder ultrasound examination, and avoid a high-fat diet in the near future." The online question-and-answer device combines the user question, the second inference process data, and the second answer data to form the second answer trajectory data.

[0031] The online question-and-answer device can invoke a large-scale evaluation model to reassess the reasoning process of the second response trajectory data, reviewing whether the revised reasoning is complete, the factual judgment is accurate, and the recommendations comply with medical safety guidelines. When the large-scale evaluation model determines that the second response trajectory data is the correct response trajectory, the online question-and-answer device can extract the second answer data and provide it to the user through the user's client as the final online question-and-answer result.

[0032] In this example scenario, by simultaneously recording user questions, reasoning process data, and answer data in the response trajectory data, both the question-and-answer model and the judging model can directly analyze the original questions at all reasoning and review stages. Furthermore, by providing only error attributions without offering correct solutions, the judging model encourages the question-and-answer model to autonomously complete the reasoning chain and improve logical judgment in the second round of reasoning, thereby enhancing the professionalism and security of the question-and-answer process. In this way, the entire online question-and-answer process can automatically identify and correct reasoning errors, thus improving the reliability of online question-and-answer content.

[0033] One embodiment of this application provides an example of an application scenario for a model training method. This method can be applied to the model training device of an online question-answering platform to improve the reasoning ability and answer accuracy of a large question-answering model in medical online question-answering scenarios. In this example scenario, the answer trajectory data generated by the online question-answering device in real user consultations is used as the training source. The model training device generates corrected answer trajectory data by calling a specified high-order large model, and combines the correct and incorrect answer trajectory data to train and update the large question-answering model and the evaluation model.

[0034] For example, during the operation of an online question-answering device, the question-answering model might generate an incorrect reasoning trajectory in response to a user's question: "Is it dangerous to experience dull pain in the upper right abdomen after a meal accompanied by fever?" This incorrect answer trajectory data might include flawed reasoning processes, such as misclassifying the symptoms as ordinary stomach discomfort or failing to consider the necessary risk assessment in conjunction with fever, leading to an overly hasty answer from the question-answering model. The model training device can obtain this incorrect answer trajectory data from the online question-answering system as one of the input data for subsequent training processing.

[0035] The model training device can invoke a designated high-order large model based on the aforementioned erroneous answer trajectory data to generate corrected answer trajectory data. The designated high-order large model is larger in scale and has stronger reasoning capabilities, enabling it to supplement the complete analysis path and provide more medically plausible corrected answers when generating corrected reasoning process data. For example, after receiving a user's question and erroneous answer trajectory, the designated high-order large model can generate corrected reasoning process data, such as "Right upper abdominal pain accompanied by fever is more likely to be a manifestation of inflammatory diseases such as cholecystitis, and further judgment should be made in conjunction with inflammatory indicators," and generate corresponding corrected answer data, such as "It is recommended to seek medical attention promptly for a liver and gallbladder ultrasound and infection indicators." Because the corrected reasoning process data also includes reflections on erroneous reasoning, the question-answering large model can more clearly learn the source of errors and correction methods in subsequent training.

[0036] The model training device can use corrected answer trajectory data as training samples to train and update the large-scale question-answering model. In this scenario example, the model training device can obtain correct answer trajectory data generated by the large-scale question-answering model. For example, for the user question "Does a slight cough require a CT scan?", the large-scale question-answering model may have already formed a reasonable reasoning process and given appropriate suggestions. The model training device can use this correct answer trajectory data and corrected answer trajectory data as mixed samples input into the training pipeline, allowing the large-scale question-answering model to simultaneously learn correct reasoning structures and ways to improve incorrect reasoning, thereby improving the overall model's answer quality.

[0037] During the training of the large-scale judging model, the model training device can train the model based on both correct and incorrect answer trajectory data. This enables the large-scale judging model to accurately identify errors in the reasoning process and, as required, output only error attribution information without providing improvement methods. For example, during training, incorrect answer trajectory data may teach the large-scale judging model error types such as "misuse of medical concepts," "ignoring key symptoms," and "logical leaps," while correct answer trajectory data can improve its judgment boundaries. In this way, the large-scale judging model can strictly fulfill its function of identifying errors and attributing causes in actual online question-answering scenarios, without directly providing the correct answer to the question-answering model, prompting the question-answering model to autonomously correct its logic based on error attribution information in the second reasoning.

[0038] When constructing model training tasks, the model training device can build training tasks with the same task paradigm based on both correct answer trajectory data and corrected answer trajectory data. For example, for correct answer trajectory data, the error feedback field in the model input is empty; while for corrected answer trajectory data, the error feedback field in the model input can contain corresponding error attribution information. This unified task paradigm enables the question-answering model to process structurally consistent input data during training, thereby enhancing the model's sensitivity and processing ability to error feedback fields, allowing it to more accurately utilize error attribution information for self-correction in subsequent online question-answering scenarios.

[0039] In this scenario example, the model training device, through a collaborative training process combining incorrect answer trajectory data, corrected answer trajectory data, and correct answer trajectory data, enables the large question-answering model to learn the complete path from incorrect to correct reasoning. It also enables the evaluation model to accurately identify the types of errors in reasoning and provide reasonable error attribution information. Thus, the model training method can effectively improve the reasoning ability, answer accuracy, and overall reliability of the large question-answering model in subsequent online question-answering scenarios.

[0040] Please see Figure 2 One embodiment of this application provides an artificial intelligence-based online question-answering method. The online question-answering method can be applied to an online question-answering device, which can be applied to the aforementioned electronic device possessing certain computing power and network access capabilities. Of course, in some embodiments, the online question-answering device can also be software running on an electronic device. The artificial intelligence-based online question-answering method may include the following steps.

[0041] Step S110: For the received user question, the question-and-answer big model generates first answer trajectory data; wherein, the first answer trajectory data includes the first reasoning process data and the first answer data of the question-and-answer big model for the user question; the first reasoning process data is used to express the reasoning process that leads to the first answer data.

[0042] Step S120: The evaluation model evaluates the first reasoning process data, and if it determines that there is a reasoning error in the first reasoning process data, it outputs error attribution information for the first answer trajectory data; wherein, the error attribution information is used to express the reason why the question-answering model has an error in the reasoning process.

[0043] Step S130: Based on the first answer trajectory data and the error attribution information, call the question-and-answer big model to obtain the second answer trajectory data for the user question; wherein, the second answer trajectory data includes the second reasoning process data and the second answer data for the user question; the second reasoning process data is used to express the reasoning process that leads to the second answer data.

[0044] Step S140: The evaluation model evaluates the reasoning process based on the second reasoning process data, and if the second answer trajectory data is determined to be the correct answer trajectory data, the second answer data is fed back to the user client that issued the user question.

[0045] In this embodiment, the online question-answering device can be deployed with a large question-answering model for performing online question-answering tasks and a large evaluation model for assessing the reasoning quality of the large question-answering model. Alternatively, the online question-answering device can also invoke the large question-answering model and the large evaluation model deployed remotely via a network. The online question-answering device can establish a communication connection with a user client, allowing users to submit health-related questions, such as symptom descriptions, past examination results, and medication information. User questions can be received by the online question-answering device as structured or unstructured text data and provided to the large question-answering model to initiate the online question-answering process. In this embodiment, the large question-answering model can primarily focus on reasoning based on the user question and generating answers, while the large evaluation model can primarily focus on reviewing the reasoning process of the large question-answering model and attributing errors, thus forming a clearly defined collaborative processing mechanism within the system.

[0046] In this embodiment, the question-and-answer big model can be used to generate first answer trajectory data based on received user questions. First answer trajectory data is structured data used to record the thinking process and output results of the question-and-answer big model. The first answer trajectory data includes at least first reasoning process data and first answer data. The first reasoning process data is used to express, in text or other interpretable form, the process by which the question-and-answer big model analyzes and deduces from the user question. For example, it includes the gradual understanding of user symptoms, the internal judgment path regarding possible diseases or medication risks, and intermediate reasoning content that weighs different candidate conclusions. The first answer data represents the user-oriented answer content derived by the question-and-answer big model based on the first reasoning process data, such as suggested conclusions, medication tips, or medical advice. The online question-and-answer device can process the first answer trajectory data as the basic data unit for subsequent evaluation and error correction.

[0047] In this embodiment, the large-scale evaluation model can be used to evaluate the reasoning process based on the first answer trajectory data. Specifically, the online question-and-answer device can input the first answer trajectory data, which includes the first reasoning process data and the first answer data, into the large-scale evaluation model, which then examines the reasoning logic reflected in the first reasoning process data. For example, it checks whether there are logical jumps between reasoning steps, whether there are deviations in the understanding of user symptoms or factual information, and whether the medication recommendations involved in the first answer data conflict with general medical safety guidelines. When the large-scale evaluation model determines that there is a reasoning error in the first reasoning process data, it can output error attribution information for the first answer trajectory data. The error attribution information is used to express the reasons why the question-and-answer large-scale model has errors in the reasoning process of generating the first reasoning process data and the first answer data. For example, it points out the misused preconditions in the reasoning, the ignored key symptoms, the misinterpretation of test results, or the inappropriate medication judgment. In this embodiment, the large-scale evaluation model only outputs error attribution information to indicate the cause of the error, without outputting specific improvement steps or the correct answer approach, so that the large-scale evaluation model focuses on identifying and explaining the error itself, rather than replacing the question-and-answer large-scale model in re-reasoning and decision-making. Online question-answering devices can associate and store error attribution information with the corresponding first answer trajectory data, or provide them together as input to a large question-answering model.

[0048] In this embodiment, the online question-answering device can invoke the question-answering big model again based on the first answer trajectory data and error attribution information to obtain second answer trajectory data for the same user question. Specifically, after receiving the user question, the first answer trajectory data, and the error attribution information, the question-answering big model can explicitly refer to the error causes indicated in the error attribution information when generating the second reasoning process data, correct the erroneous reasoning paths in the first reasoning process data, or supplement missing analysis steps, thereby forming a more complete and reasonable second reasoning process data. Since the error attribution information only points out the cause of the error without directly providing the correct approach, the question-answering big model needs to rethink the appropriate analysis path and conclusion based on the error causes when generating the second reasoning process data. This allows the question-answering big model to maintain control over the overall reasoning chain during the error correction process, which is conducive to the question-answering big model forming a more rigorous reasoning strategy for the same or similar questions. Based on the second reasoning process data, the question-answering big model can generate second answer data, which is used to express the updated answer content for the user formed on the basis of the corrected reasoning. The online question-answering device can combine the second reasoning process data with the second answer data to form second answer trajectory data, which can then be used for subsequent evaluation processing. In this embodiment, by having the large judging model focus on outputting error attribution information, while the large question-answering model focuses on re-reasoning based on the error attribution information, a collaborative model with clearly defined responsibilities can be formed, which helps to improve the focus of each model on its respective task.

[0049] In this embodiment, the large-scale evaluation model can also be used to evaluate and review the reasoning process of the second answer trajectory data. Specifically, the online question-and-answer device can input the second answer trajectory data into the large-scale evaluation model, which will then review the second reasoning process data and the second answer data again to determine whether problems still exist in the second reasoning process data after correction based on error attribution information, such as logical errors, factual misunderstandings, or inappropriate medication. When the large-scale evaluation model determines that the second answer trajectory data is correct, the online question-and-answer device can extract the second answer data from the second answer trajectory data and use the second answer data as the online question-and-answer result, which is then fed back to the user who asked the question through the user client to complete the online question-and-answer process corresponding to the current user question. In some embodiments, the online question-and-answer device may choose not to display the complete reasoning process data to the user, but only use the first and second reasoning process data internally as the basis for quality control and model optimization, thereby achieving automated evaluation and error correction of the online question-and-answer reasoning process without changing the user interaction experience.

[0050] In some implementations, the online question-and-answer device can output an error type and an error cause analysis for the first answer trajectory data; wherein, the error attribution information includes the error type and the error cause analysis; the error type is used to express the type of error to which the first answer trajectory data belongs; the error cause analysis is used to analyze the reason why the first answer trajectory data has the type of error represented by the error type.

[0051] In this embodiment, after the online question-answering device inputs the first answer trajectory data into the evaluation model, the evaluation model can structurally identify and express errors based on the reasoning chain, factual understanding, and relevant medical standards reflected in the first reasoning process data and the first answer data. Specifically, the evaluation model can generate corresponding error types and error cause analyses based on the detected error content, and combine the two to form error attribution information for the question-answering model to refer to and correct in subsequent reasoning.

[0052] In this embodiment, the error type can be used to express the classification category to which the error appears in the first response trajectory data belongs. The error type can be one of several predefined categories used to identify the nature of the error, such as logical reasoning error, factual misunderstanding deviation, misuse of medical concepts, omission of key symptoms, insufficient risk assessment, or non-standard medication advice. The online question-and-answer device enables the evaluation model to categorize errors into corresponding error types based on the problems that occur in the reasoning chain when reviewing the data of the first reasoning process. This provides a clear categorized description of the error manifestation, making it easier for the question-and-answer model to pinpoint the error location in the next round of reasoning.

[0053] In this implementation, error cause analysis can be used to express the specific reasons leading to the error type. The evaluation model can further analyze the causes of the error after identifying the error type, based on its understanding of the data from the first reasoning process. For example, error cause analysis can point out factors such as insufficient preconditions used in a reasoning step, misunderstanding of key symptoms in the user's question, neglect of necessary medical background knowledge, over-reliance on inappropriate assumptions, or insufficient consideration of drug contraindications in medication recommendation-related reasoning. Error cause analysis can be used as a component in generating error attribution information, combined with the error type, to provide structured error localization information for the question-answering model.

[0054] In this implementation, the large-scale evaluation model is primarily used to output error types and error cause analyses, thereby forming error attribution information. It's important to note that the large-scale evaluation model only outputs content expressing the error and does not provide any direct correction schemes or correct practices. For example, the large-scale evaluation model does not tell the question-answering model how to modify the reasoning steps, nor does it directly provide the correct conclusion or improved reasoning path. Thus, the large-scale evaluation model remains focused on its responsibility of error identification and cause analysis, without interfering with the question-answering model's autonomous construction of the correct reasoning process. This allows the question-answering model to rethink and deduce upon receiving error attribution information, optimizing the reasoning chain based on the error causes, thereby improving the accuracy of the subsequently generated second-response trajectory data.

[0055] In some implementations, the error type is enumerated data; the error cause analysis is generative divergent data of the thought process of the evaluation model.

[0056] In this embodiment, the error type can be used as enumerated data to represent the category to which the error belongs in the first answer trajectory data. Enumerated data can be used to select from multiple preset error categories, ensuring that the value range of the error type is fixed and has clear classification boundaries. The online question-answering device allows the evaluation model to select the error category that best matches the current error from a preset set of error types after identifying the reasoning problem in the first answer trajectory data. Examples include logical reasoning errors, factual misunderstanding biases, omission of key symptoms, misuse of medical concepts, or non-compliant medication recommendations. Using enumerated data to represent error types helps to standardize the error identification results, making subsequent error attribution information more structurally consistent and facilitating the question-answering model to adopt corresponding reasoning correction strategies based on different error categories.

[0057] In this embodiment, error cause analysis can serve as generative divergent data used to express the specific causes of errors. Generative divergent data can be generated by the large judging model based on its internal reasoning mechanism when analyzing the first reasoning process data and the first answer data. It is used to summarize the causes leading to a particular error type. For example, after identifying a certain error type, the large judging model can output the analysis content of its internal thinking process from the context of the reasoning chain, including explanations of erroneous logic, hints of misused concepts, descriptions of omitted symptoms, or elaboration on the reasons for insufficient reasoning steps. Because error cause analysis is generative divergent data, its content typically has a higher degree of expressive freedom and semantic coverage, and can fully express the large judging model's understanding of the root causes of errors without being limited to a fixed format.

[0058] In this implementation, by designing error types as enumerated data and error cause analysis as generative divergent data, the evaluation model can achieve a combination of structured and unstructured content when outputting error attribution information. Enumerated data clarifies the error category, making error localization more precise; generative divergent data expresses the causal explanations formed by the evaluation model during its internal thinking process, making error attribution more detailed and comprehensive. The combination of these two approaches allows error attribution information to possess both structured classification capabilities and the flexibility of natural language expression. This helps the question-answering model to comprehensively consider both error categories and causes in subsequent reasoning, thereby more effectively correcting the reasoning path and generating more accurate second-response trajectory data.

[0059] Please see Figure 3This application also provides a model training method. The model training method can be applied to a model training device. The model training device can be applied to the aforementioned electronic device with certain computing power and network access capabilities. Of course, in some embodiments, the model training device can also be software running on the electronic device. The model training method includes the following steps.

[0060] Step S210: Obtain the error answer trajectory data generated by the question-and-answer big model for the user's question; wherein, the error answer trajectory data is the result of incorrect answer data.

[0061] Step S220: Based on the incorrect answer trajectory data, call the specified high-order large model to generate corrected answer trajectory data; wherein, the corrected answer trajectory data includes corrected reasoning process data and corrected answer data; the corrected reasoning process data expresses the correct reasoning content that leads to the correct corrected answer data, and also includes reflection content on the errors in the reasoning process in the incorrect answer trajectory data; the model size of the specified high-order large model is larger than the model size of the question-and-answer large model.

[0062] Step S230: Train the question-answering model based on the corrected answer trajectory data as samples.

[0063] In this embodiment, the model training device can be used to acquire the answer trajectory data required for training, call a specified high-order large model to generate corrected answer trajectory data, and train and update the question-and-answer large model based on the corrected answer trajectory data, so as to improve the reasoning quality and answer accuracy of the question-and-answer large model in online question-and-answer scenarios.

[0064] In this embodiment, the model training device can be used to acquire erroneous answer trajectory data generated by the question-answering model in response to user questions. Since the answer data generated by the erroneous answer trajectory data contains errors, the reasoning process data leading to the answer data will also inevitably contain errors. Specifically, the erroneous answer trajectory data can include reasoning process data and answer data generated by the question-answering model, where the answer data contains errors, causing the answer trajectory data to be judged as erroneous answer trajectory data during the evaluation of the large model. For example, the erroneous answer trajectory data can reflect incorrect reasoning premises, omission of key information, misuse of medical concepts, or inappropriate risk judgments by the question-answering model when understanding user symptoms.

[0065] In this embodiment, the model training device can call a designated high-order large model to generate corrected answer trajectory data based on the erroneous answer trajectory data. The designated high-order large model can be a generative artificial intelligence model with a larger model size than the question-and-answer large model, possessing stronger reasoning ability, information association ability, and complex scene handling ability. The model training device can input user questions and erroneous answer trajectory data into the designated high-order large model, which will then generate corrected answer trajectory data containing corrected reasoning process data and corrected answer data. The corrected reasoning process data can express the correct reasoning content of the designated high-order large model when forming the corrected answer data, and further include reflections on the errors in the erroneous answer trajectory data, such as pointing out the reasons for errors in the original reasoning chain of the question-and-answer large model, or providing comparative explanations for erroneous conclusions. Because the designated high-order large model has a stronger reasoning breadth and expressive ability, the corrected reasoning process data it generates is usually more complete and interpretable, helping to provide high-quality reasoning examples for the question-and-answer large model during training.

[0066] In this embodiment, the model training device can train and update the question-and-answer large model based on corrected answer trajectory data. Specifically, the model training device can input the corrected answer trajectory data as training samples into the question-and-answer large model, enabling the large model to learn the correct reasoning paths in the corrected reasoning process data and the reasoning improvement logic reflected in the reflected content, thereby improving the reasoning ability and answer accuracy of the large model when encountering similar user questions. In some embodiments, the model training device can employ supervised fine-tuning, reward modeling training, or other training strategies suitable for generative models, allowing the large model to gradually optimize its internal reasoning mechanism and reduce the occurrence of reasoning errors under the guidance of the corrected answer trajectory data. Thus, by using the corrected answer trajectory data generated by a specified high-order large model as training samples, the self-correction ability of the question-and-answer large model for erroneous reasoning can be effectively improved, making the answer results of the large model in subsequent online question-and-answer scenarios more reliable.

[0067] In some implementations, the model training device can acquire the correct answer trajectory data generated by the question-and-answer big model for user questions; wherein the reasoning process for deriving the answer data from the correct answer trajectory data is correct; the model training device can also use the correct answer trajectory data and the corrected answer trajectory data as mixed samples to train the question-and-answer big model.

[0068] In this embodiment, the model training device can be further used to acquire correct answer trajectory data generated by the question-answering big model in response to user questions. Correct answer trajectory data can be used to represent the answer trajectory generated by the question-answering big model when processing user questions, and judged as having correct reasoning process after review by the evaluation big model. The reasoning process data contained in the correct answer trajectory data can represent the complete reasoning chain of the question-answering big model in the process of progressively analyzing user questions, identifying key symptoms, explaining medical concepts, and forming the final answer, and can reflect characteristics such as logical coherence, accurate understanding of facts, and standardized medical advice. Unlike incorrect answer trajectory data, correct answer trajectory data does not contain reasoning errors pointed out by the evaluation big model and can be used as positive samples in the training of the question-answering big model.

[0069] In this embodiment, the model training device can, based on the generated or acquired corrected answer trajectory data, use both corrected and corrected answer trajectory data as mixed samples to train the large-scale question-answering model. The model training device can use the corrected answer trajectory data to provide the large-scale question-answering model with standardized inference path examples, enabling the model to learn the logical structure of high-quality inference chains. Simultaneously, the model training device can use the corrected answer trajectory data to provide the large-scale question-answering model with inference content corrected by a specified higher-order model, enabling the model to learn correction methods for typical errors and the inference optimization logic reflected in the reflection content. Thus, the mixed samples formed by combining corrected and corrected answer trajectory data can improve the reasoning ability of the large-scale question-answering model from both positive reasoning and error-correcting reasoning perspectives. For example, the model training device can employ supervised fine-tuning, reinforcement training, reward modeling training, or other training strategies suitable for generative models, enabling the large-scale question-answering model to more comprehensively understand user questions and form more reliable inference content under the guidance of mixed samples. In some embodiments, the proportion of corrected and corrected answer trajectory data used during training can be dynamically adjusted according to the training stage to improve the overall inference quality. By using mixed samples to train a large question-answering model, the reliability of the large question-answering model in real-world online question-answering scenarios can be effectively improved.

[0070] In some implementations, the model training device can train the large evaluation model based on the correct answer trajectory data and the incorrect answer trajectory data, and constrain the large evaluation model to only output error attribution information and not to output improvement practices.

[0071] In this embodiment, the model training device can also be used to train and update the large-scale evaluation model based on correct and incorrect answer trajectory data, enabling the large-scale evaluation model to identify whether errors exist in the reasoning process of the question-and-answer large-scale model and output corresponding error attribution information. The model training device can use correct answer trajectory data as positive samples representing complete reasoning chains, accurate factual understanding, and standardized medical recommendations, and use incorrect answer trajectory data as negative samples representing biased reasoning processes, incorrect reasoning premises, or inappropriate conclusions, to construct a training dataset for training the large-scale evaluation model. By comparing the differences in reasoning paths between correct and incorrect answer trajectory data, the model training device can enable the large-scale evaluation model to gradually learn characteristics that judge the quality of the reasoning process, such as logical coherence, the correctness of symptom analysis, and the rationality of medical recommendations.

[0072] In this embodiment, the model training device can be further used to constrain the output of the large-scale evaluation model, ensuring that it only outputs error attribution information and not improvements to the question-answering model. Specifically, to ensure that the question-answering model can deduce the correct reasoning chain based on the error attribution information during the generation of the second answer trajectory data, rather than relying on the direct answer or optimization path of the large-scale evaluation model, the model training device can constrain its output using specific training methods during training. For example, during training, when the large-scale evaluation model processes error answer trajectory data, the model training device can require it to only locate the position of the reasoning error, the type of error, and the cause of the error, without allowing it to generate content expressing correct reasoning steps, alternative answers, or specific correction schemes. Thus, the function of the large-scale evaluation model is limited to "identifying errors and explaining their causes," without participating in providing solutions, maintaining its functional independence from the large-scale question-answering model throughout the online question-answering system.

[0073] In this embodiment, the large-scale evaluation model obtained through the above training method can focus on performing the task of evaluating the quality of the reasoning process. It can establish a benchmark for identifying correct reasoning logic based on correct answer trajectory data and learn patterns and misuse conditions of common reasoning errors based on incorrect answer trajectory data. In practical applications, when identifying incorrect reasoning, the large-scale evaluation model can output error attribution information instead of content indicating the correct reasoning path. This forces the question-answering model to independently think and actively correct its reasoning chain based on the error attribution information when regenerating answer trajectory data, thereby promoting stronger self-reasoning capabilities during the error correction process. Thus, by using both correct and incorrect answer trajectory data to train the large-scale evaluation model and imposing content constraints on its output during training, a clear division of functions between the question-answering and evaluation models can be achieved, helping to improve the independence and synergy between reasoning evaluation and reasoning generation in online question-answering systems.

[0074] In some implementations, the model training device can construct model training tasks based on the correct answer trajectory data and the corrected answer trajectory data respectively; wherein, the model training task constructed based on the correct answer trajectory data and the model training task constructed based on the corrected answer trajectory data have the same thinking method and thought path.

[0075] In this embodiment, the model training device can also be used to construct model training tasks for the large question-answering model based on correct answer trajectory data and corrected answer trajectory data, respectively, so that the large question-answering model can learn high-quality reasoning examples from two different sources simultaneously in the same training process. Both correct answer trajectory data and corrected answer trajectory data contain reasoning process data and answer data in their information structure, and are used to express how to form a reasonable conclusion from the user's question. Therefore, both can be used as effective training inputs for the large question-answering model.

[0076] In this embodiment, the model training device can construct a first type of model training task based on correct answer trajectory data. Specifically, this first type of model training task can be used to enable the question-answering large model to learn the internal logical structure of correct reasoning process data, such as a reasonable understanding path of symptoms, complete referencing relationships of factual information, and standardized generation methods for medical advice. The goal of the first type of model training task is to enhance the ability of the question-answering large model to reproduce correct reasoning chains, enabling the question-answering large model to generate correct answers based on its existing knowledge when faced with similar user questions.

[0077] In this embodiment, the model training device can also construct a second type of model training task based on the corrected answer trajectory data. This second type of model training task, in addition to including the structural content of the correct reasoning process data, also includes the reflection content generated by the specified high-order large model in response to incorrect reasoning. By introducing corrected reasoning process data into the training, the model training device enables the question-answering large model to understand the causes of typical errors while learning correct reasoning, thereby proactively avoiding similar errors in subsequent reasoning generation. In some embodiments, the model training task constructed based on the correct answer trajectory data and the model training task constructed based on the corrected answer trajectory data have the same thinking methods and thought paths, and both reflect attention and consideration to error feedback during the thinking process, thereby enhancing the model's sensitivity and attention to error feedback.

[0078] In this embodiment, the model training device can make the model training task based on correct answer trajectory data and the model training task based on correct answer trajectory data have the same task paradigm. This task paradigm is a unified training format for large question-answering models, used to ensure that different types of training samples can be input into the same training framework without difference. Specifically, the model training device can construct both types of training tasks into a unified form of "generating answer data consistent with reasoning from user questions and corresponding reasoning process data," so that the model does not need to distinguish the source of samples during training, but only needs to follow a consistent training logic. Since the two types of training tasks use the same data format, the same input field structure, and the same training objective function, the large question-answering model can learn the correct reasoning path and reflection content simultaneously under the unified paradigm, thereby improving its ability to self-identify reasoning errors and the stability of reasoning in complex medical scenarios.

[0079] In this embodiment, by using the same task paradigm to construct the model training task, the model training device can achieve seamless mixed input of training data during the training process. This avoids training instability caused by differences in sample formats and is beneficial to enhancing the generalization ability and reasoning consistency of the large question-answering model. Ultimately, this enables the large question-answering model to have stronger reasoning reliability when facing diverse medical consultation questions.

[0080] In some implementations, the model input data for the model training task includes an error feedback field; wherein, in the model training task constructed based on the correct answer trajectory data, the content of the error feedback field is empty; and in the model training task constructed based on the correct answer trajectory data, the content of the error feedback field includes error attribution information.

[0081] In this embodiment, the model training device can use the error feedback field as part of the model input data for the model training task. This allows the question-answering model to distinguish between different types of training sample sources during training, thereby guiding the model to utilize sample information in different ways while maintaining a unified task paradigm. The error feedback field can be used to express whether the training sample contains error attribution information for the original error answer trajectory data, and when the sample comes from corrected answer trajectory data, to provide the error-related attribution content to the question-answering model for reference, thereby improving its ability to identify error reasoning patterns. When constructing the model training task, the model training device can fill or leave the error feedback field blank according to different answer trajectory data, enabling the question-answering model to automatically perceive sample attributes during training without introducing an additional sample labeling system.

[0082] In this embodiment, the model training device can construct a first type of model training task based on correct answer trajectory data. Since the correct answer trajectory data is autonomously generated by the question-answering model and its reasoning process is confirmed as correct by the evaluation model, there are no incorrect reasoning chains or attribution content related to errors in the correct answer trajectory data. Therefore, when constructing the corresponding model training task based on the correct answer trajectory data, the model training device can set the content of the error feedback field to empty, so that the question-answering model only focuses on the correct reasoning process data and the answer data itself when training this type of sample, thereby strengthening the learning of the correct reasoning path.

[0083] In this embodiment, the model training device can construct a second type of model training task based on corrected answer trajectory data. The corrected answer trajectory data is generated by a specified high-order large model and includes reflections on reasoning errors in the original incorrect answer trajectory data, building upon the correct reasoning process data. These reflections originate from error attribution information generated from the incorrect answer trajectory data and evaluate the error type and cause analysis output by the large model for the incorrect reasoning chain. Therefore, when constructing a model training task based on the corrected answer trajectory data, the model training device can fill the error attribution information into the error feedback field, allowing the error feedback field to express the error location and error analysis content contained in the corrected answer trajectory data. By inputting a training task containing the error feedback field into the question-answering large model, the model can learn the correct reasoning path while further understanding the causes and typical characteristics of incorrect reasoning, thereby proactively avoiding similar errors in future reasoning processes.

[0084] Please see Figure 4 The embodiments of this application also provide an online question-and-answer device based on artificial intelligence. The online question-and-answer device includes: a receiving module, an output module, a calling module, and a feedback module.

[0085] The receiving module is used to generate first answer trajectory data from a question-and-answer big data model for a received user question; wherein, the first answer trajectory data includes first reasoning process data and first answer data of the question-and-answer big data model for the user question; the first reasoning process data is used to express the reasoning process that leads to the first answer data.

[0086] The output module is used to evaluate the first reasoning process data by the evaluation model, and if it is determined that there is a reasoning error in the first reasoning process data, it outputs error attribution information for the first answer trajectory data; wherein, the error attribution information is used to express the reason why the question-answering model has an error in the reasoning process.

[0087] The calling module is used to call the question-and-answer big model based on the first answer trajectory data and the error attribution information to obtain the second answer trajectory data for the user question; wherein, the second answer trajectory data includes the second reasoning process data and the second answer data for the user question; the second reasoning process data is used to express the reasoning process that leads to the second answer data.

[0088] The feedback module is used to evaluate the second reasoning process data by the evaluation model, and if the second answer trajectory data is determined to be the correct answer trajectory data, to provide feedback of the second answer data to the user client that issued the user question.

[0089] In this embodiment, the functions and effects of the AI-based online question-and-answer device can be explained in comparison with the aforementioned embodiments, and will not be repeated here.

[0090] Please see Figure 5 The embodiments of this application also provide a model training apparatus for training the aforementioned large question-answering model. The model training apparatus includes: an acquisition module, a generation module, and a training module.

[0091] The acquisition module is used to acquire the error answer trajectory data generated by the question-and-answer big data model for user questions; wherein, the error answer trajectory data is due to errors in the reasoning process that leads to the answer data.

[0092] The generation module is used to call a specified high-order large model to generate corrected answer trajectory data based on the erroneous answer trajectory data; wherein, the corrected answer trajectory data includes corrected reasoning process data and corrected answer data; the corrected reasoning process data expresses the correct reasoning content that leads to the correct corrected answer data, and also includes reflection content on the errors in the reasoning process in the erroneous answer trajectory data; the model size of the specified high-order large model is larger than the model size of the question-and-answer large model.

[0093] The training module is used to train the question-answering model based on the corrected answer trajectory data as samples.

[0094] In this embodiment, the functions and effects achieved by the model training device can be explained in comparison with the aforementioned embodiments, and will not be repeated here.

[0095] Please see Figure 6 This application also provides a computer device comprising: a memory and a processor, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the method described above.

[0096] The memory, processor, and communication interface in the computer device can communicate with each other via the system bus and network communication.

[0097] In this embodiment, the functions and effects implemented by the computer device can be explained by referring to the foregoing embodiments, and will not be repeated here.

[0098] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0099] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0100] This application also provides a computer program product containing instructions, including a computer program / instructions that, when executed by a processor, implement the method as described above.

[0101] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0102] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments of this application, and are not intended to limit the scope of the invention.

[0103] It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0104] It is understood that the various implementation methods described in this application can be implemented individually or in combination, and the implementation methods in this application are not limited in this respect.

[0105] Unless otherwise stated, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0106] It is understood that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0107] It is understood that the memory in the embodiments of this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Specifically, non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0108] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0111] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0112] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0113] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. An online question-answering method based on artificial intelligence, characterized in that, include: For the received user question, the question-and-answer big data model generates first answer trajectory data; wherein, the first answer trajectory data includes first reasoning process data and first answer data of the question-and-answer big data model for the user question; the first reasoning process data is used to express the reasoning process that leads to the first answer data; The evaluation model assesses the first reasoning process data and, if it determines that there is a reasoning error in the first reasoning process data, outputs error attribution information for the first answer trajectory data; wherein, the error attribution information is used to express the reason why the question-answering model has an error in the reasoning process; Based on the first answer trajectory data and the error attribution information, the question-and-answer big model is invoked to obtain second answer trajectory data for the user question; wherein, the second answer trajectory data includes second reasoning process data and second answer data for the user question; the second reasoning process data is used to express the reasoning process that leads to the second answer data; The evaluation model assesses the second reasoning process data, and if it determines that the second answer trajectory data is the correct answer trajectory data, it feeds back the second answer data to the user client that issued the user question.

2. The method according to claim 1, characterized in that, Output error attribution information for the first response trajectory data, including: Output the error type and error cause analysis for the first answer trajectory data; wherein, the error attribution information includes the error type and the error cause analysis; the error type is used to express the error category to which the first answer trajectory data belongs; the error cause analysis is used to analyze the reason why the first answer trajectory data has the error category represented by the error type.

3. The method according to claim 2, characterized in that, The error types are enumerated data; the error cause analysis is generative divergent data of the thinking process of the evaluation model.

4. A model training method, characterized in that, For training a large question-answering model as described in any one of claims 1 to 3; the method includes: Obtain the error answer trajectory data generated by the question-and-answer model for user questions; wherein, the error answer trajectory data is the result of erroneous answer data. Based on the erroneous answer trajectory data, a specified high-order large model is invoked to generate corrected answer trajectory data; wherein, the corrected answer trajectory data includes corrected reasoning process data and corrected answer data; the corrected reasoning process data expresses the correct reasoning content that leads to the correct corrected answer data, and also includes reflection content on the errors in the reasoning process in the erroneous answer trajectory data; the model size of the specified high-order large model is larger than the model size of the question-and-answer large model; The question-answering model is trained using the corrected answer trajectory data as samples.

5. The method according to claim 4, characterized in that, The method further includes: obtaining correct answer trajectory data generated by the question-and-answer model for user questions; wherein the reasoning process for deriving the answer data from the correct answer trajectory data is correct; Training the question-and-answer model based on the corrected answer trajectory data as samples includes: using the correct answer trajectory data and the corrected answer trajectory data as mixed samples to train the question-and-answer model.

6. The method according to claim 5, characterized in that, The method further includes: Based on the correct answer trajectory data and the incorrect answer trajectory data, the evaluation model is trained, and the evaluation model is constrained to only output error attribution information and not to output improvement methods.

7. The method according to claim 5, characterized in that, Using the correct answer trajectory data and the correct answer trajectory data as mixed samples, the question-answering large model is trained, including: Model training tasks are constructed based on the correct answer trajectory data and the corrected answer trajectory data, respectively; wherein, the model training task constructed based on the correct answer trajectory data and the model training task constructed based on the correct answer trajectory data have the same thinking method and thought path.

8. The method according to claim 7, characterized in that, The model input data for the model training task includes an error feedback field; In the model training task built based on the correct answer trajectory data, the content of the error feedback field is empty; in the model training task built based on the correct answer trajectory data, the content of the error feedback field includes error attribution information.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 8.

10. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the method as described in any one of claims 1 to 8.

11. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images