Information generation method, training method and device based on large model

Abstract

The invention provides an information generation method and device based on a large model, and a training method and device, and relates to the technical field of artificial intelligence, in particular to the technical fields of multi-modal large models, AI medical treatment and the like. The specific implementation scheme of the information generation method comprises the following steps: analyzing historical doctor seeing information and state description information of a target object by using a large model to obtain a plurality of target problems and interaction paths; wherein the interaction path indicates an interaction sequence among the plurality of target questions used for interacting with the target object; according to the interaction sequence, performing multiple rounds of interaction with the target object based on the multiple target questions by using the large model to obtain multiple target answers associated with the multiple target questions; and performing information fusion on the plurality of target replies by using the large model to generate target information used for describing an inquiry result of the target object.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, in particular to the technical field of multi-modal large models, AI medical treatment, and the like, and specifically relates to a large model-based information generation method and a training method and device. BACKGROUND

[0002] A multi-modal large model is a large-scale artificial intelligence model that can understand, process, generate, and correlate multiple types of information, such as text, images, audio, and video, and achieve deeper semantic understanding and cross-modal content generation.

[0003] In the field of AI medical treatment, due to the obvious individual differences of users in the interaction process, and the fixed and mechanical medical inquiry process in related examples, it is difficult to collect complete and accurate information in the interaction process with the user. SUMMARY

[0004] The present disclosure provides a large model-based information generation method and a model training method and device.

[0005] According to an aspect of the present disclosure, a large model-based information generation method is provided, including: analyzing, by a large model, historical medical information and state description information of a target object to obtain a plurality of target questions and an interaction path; wherein the interaction path indicates an interaction order of the plurality of target questions for interacting with the target object; performing, by the large model, a plurality of rounds of interaction with the target object based on the plurality of target questions in the interaction order to obtain a plurality of target replies associated with the plurality of target questions; and performing, by the large model, information fusion on the plurality of target replies to generate target information for describing a medical inquiry result of the target object.

[0006] According to another aspect of the present disclosure, a training method of a large model is provided, including: analyzing, by an initial large model, sample historical medical information and sample state description information of a sample object to obtain a plurality of sample questions and a sample interaction path; wherein the sample interaction path indicates an interaction order of the plurality of sample questions for interacting with the sample object; and training, based on a target loss function, the initial large model by using the plurality of sample questions and a plurality of sample labels to obtain a trained large model; wherein the plurality of sample labels are a plurality of interaction questions for asking the sample object generated by calling at least two pre-trained language models through a plurality of rounds of simulated doctor-patient interactions.

[0007] According to another aspect of the present disclosure, a large model-based information generation device is provided, including a first analysis module, an interaction module, and a generation module.

[0008] The first analysis module is configured to analyze the historical medical information and the state description information of the target object by using a large model to obtain a plurality of target questions and an interaction path. The interaction path indicates an interaction sequence of the plurality of target questions.

[0009] The interaction module is configured to perform a plurality of rounds of interaction with the target object based on the plurality of target questions by using the large model in the interaction sequence to obtain a plurality of target answers associated with the plurality of target questions.

[0010] The generation module is configured to perform information fusion on the plurality of target answers by using the large model to generate target information for describing the diagnosis result of the target object.

[0011] According to another aspect of the present disclosure, a training device of a large model is provided, comprising a second analysis module and a training module.

[0012] The second analysis module is configured to analyze sample historical medical information and sample state description information of a sample object by using an initial large model to obtain a plurality of sample questions and a sample interaction path. The sample interaction path indicates an interaction sequence of the plurality of sample questions for interacting with the sample object.

[0013] The training module is configured to train the initial large model based on a target loss function by using the plurality of sample questions and a plurality of sample labels to obtain a trained large model. The plurality of sample labels are a plurality of interaction questions for asking the sample object generated by calling at least two pre-trained language models through a plurality of rounds of simulated doctor-patient interaction.

[0014] According to another aspect of the present disclosure, an intelligent agent is provided, comprising an input module, a processing module and an output module. The input module is configured to receive historical medical information and state description information of a target object. The processing module is configured to determine a target task based on the historical medical information and the state description information received by the input module, determine a target large model based on the target task, execute the information generation method described above by calling the target large model, and obtain target information for describing the diagnosis result of the target object. The output module is configured to output the target information obtained by the processing module.

[0015] According to another aspect of the present disclosure, an electronic device is provided, comprising at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0016] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method as described above.

[0017] According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the method as described above.

[0018] It should be understood that the matters described herein are intended to be illustrative rather than limiting. The scope of the present disclosure is to be determined by the following claims. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which are included to provide a further understanding of the application, illustrate embodiments of the application and together with the description serve to explain the principle of the application.

[0020] Figure 1 An exemplary system architecture to which the method and device for training a large model or the method for generating information based on a large model according to embodiments of the present disclosure can be applied is schematically shown;

[0021] Figure 2 A flowchart of the method for generating information based on a large model according to embodiments of the present disclosure is schematically shown;

[0022] Figure 3A A schematic diagram of analyzing historical medical information and state description information of a target object by using a large model to obtain a plurality of target questions and interaction paths according to embodiments of the present disclosure is schematically shown;

[0023] Figure 3B A schematic diagram of analyzing historical medical information and state description information of a target object by using a large model to obtain a plurality of target questions and interaction paths according to embodiments of the present disclosure is schematically shown;

[0024] Figure 3C A schematic diagram of analyzing historical medical information and state description information of a target object by using a large model to obtain a plurality of target questions and interaction paths according to embodiments of the present disclosure is schematically shown;

[0025] Figure 4A A schematic diagram of dynamically adjusting interaction questions in an interaction process according to embodiments of the present disclosure is schematically shown;

[0026] Figure 4B A schematic diagram of dynamically adjusting interaction questions in an interaction process according to embodiments of the present disclosure is schematically shown;

[0027] Figure 4CA schematic diagram of dynamically adjusting an interactive question in an interaction process according to an embodiment of the present disclosure is shown schematically.

[0028] Figure 4D A schematic diagram of dynamically adjusting an interactive question in an interaction process according to an embodiment of the present disclosure is shown schematically.

[0029] Figure 4E A schematic diagram of dynamically adjusting an interactive question in an interaction process according to an embodiment of the present disclosure is shown schematically.

[0030] Figure 5 A schematic diagram of a large model-based information generation method according to an embodiment of the present disclosure is shown schematically.

[0031] Figure 6 A schematic diagram of target information for characterizing a diagnosis result of a target object according to an embodiment of the present disclosure is shown schematically.

[0032] Figure 7 A flowchart of a large model training method according to an embodiment of the present disclosure is shown schematically.

[0033] Figure 8A A schematic diagram of a large model training method according to an embodiment of the present disclosure is shown schematically.

[0034] Figure 8B A schematic diagram of a large model training method according to an embodiment of the present disclosure is shown schematically.

[0035] Figure 9 A block diagram of a large model-based information generation apparatus according to an embodiment of the present disclosure is shown schematically.

[0036] Figure 10 A block diagram of a large model training apparatus according to an embodiment of the present disclosure is shown schematically.

[0037] Figure 11 A block diagram of an agent according to an embodiment of the present disclosure is shown schematically; and

[0038] Figure 12 A block diagram of an electronic device suitable for implementing a large model-based information generation method or a large model training method according to an embodiment of the present disclosure is shown schematically. DETAILED DESCRIPTION

[0039] Exemplary embodiments of the present disclosure are described herein with reference to the accompanying drawings, which are cited as illustrative examples. Various details of the embodiments of the present disclosure are described herein in order to provide a thorough understanding thereof. It will be understood by those of ordinary skill in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, in the following description, descriptions of well-known functions and constructions are omitted for clarity and conciseness.

[0040] In the related examples, the fixed and mechanical interrogation process is used to interact with the user, which is difficult to meet the personalized needs of the user in the AI medical interaction scene.

[0041] For example, a user has uploaded a recent physical examination report showing that the user's blood pressure is high. During the AI medical interaction process, the user is still asked whether there is a history of chronic diseases such as high blood pressure. This repeated interrogation process not only reduces the interrogation efficiency, but also causes negative emotions of the user in the case of suffering from pain, and reduces the user experience.

[0042] For another example, a user has a bias in understanding a certain problem, resulting in ambiguous or incorrect answers. In the related examples, since a fixed interrogation process is used, the same way of questioning is continued when the problem does not receive a clear reply, so that the interrogation process cannot continue.

[0043] Therefore, the embodiments of the present disclosure provide a large model-based information generation method, including: analyzing historical medical information and state description information of a target object by using a large model to obtain a plurality of target questions and an interaction path; wherein the interaction path indicates an interaction order between the plurality of target questions for interacting with the target object; performing a plurality of rounds of interaction between the target object and the plurality of target questions based on the interaction order by using the large model to obtain a plurality of target answers associated with the plurality of target questions; and performing information fusion on the plurality of target answers by using the large model to generate target information for describing an interrogation result of the target object.

[0044] Figure 1 An exemplary system architecture to which the large model-based information generation method or the training method and device of the large model according to the embodiments of the present disclosure can be applied is schematically shown.

[0045] It should be noted that, Figure 1The shown is only an example of a system architecture to which the embodiments of the present disclosure can be applied, to help those skilled in the art understand the technical content of the present disclosure, but does not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios. For example, in another embodiment, an exemplary system architecture to which the large model-based information generation method or the training method and device of the large model can be applied can include a terminal device, but the terminal device can not need to interact with a server to implement the large model-based information generation method or the training method and device of the large model provided by the embodiments of the present disclosure.

[0046] As shown in Figure 1 According to the system architecture 100 of the embodiment, the terminal device 101 can be loaded with a large model, and a user can interact with the large model by inputting historical consultation information and state description information to the terminal device 101. In the interaction process, the large model can first generate questions and an interaction sequence for interacting with the user based on the historical consultation information and the state description information. In the interaction process with the user according to the interaction sequence, the interaction sequence or the interaction questions in the next round of interaction process can also be dynamically adjusted based on the user's reply, so as to collect more accurate information for the current state of the user through a personalized interaction process, and finally obtain the target information 110. The target information 110 can include medical record information for representing the diagnosis result of the target object.

[0047] In the embodiments of the present disclosure, the ways of inputting information by the user include but are not limited to text input, image input, audio / video input, etc.

[0048] It should be understood that Figure 1 The number of terminal devices in may be any number according to the needs of implementation.

[0049] In the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision, disclosure and application of user personal information comply with the relevant legal regulations, necessary security measures are taken, and the public order and good customs are not violated.

[0050] In the technical solutions of the present disclosure, the authorization or consent of the user is obtained before the user's personal information is acquired or collected.

[0051] Figure 2 An exemplary flowchart of a large model-based information generation method according to the embodiments of the present disclosure is shown.

[0052] As shown in Figure 2 The method 200 includes operations S210-S230.

[0053] In operation S210, the history medical information and the state description information of the target object are analyzed by using a large model to obtain a plurality of target questions and an interaction path.

[0054] In operation S220, a plurality of target questions are interacted with the target object in a plurality of rounds based on the large model in an interaction order to obtain a plurality of target answers associated with the plurality of target questions.

[0055] In operation S230, the plurality of target answers are fused by using the large model to generate target information for describing the diagnosis result of the target object.

[0056] In some embodiments, the target object can be a patient interacting with an AI doctor in an AI medical interaction scenario. The history medical information can record the diagnosis and treatment results received by the target object in a predetermined history period. For example, the examination report generated by the target object using the medical examination equipment in the predetermined history period, the image of the lesion site collected, etc.

[0057] In some embodiments, the state description information can represent the subjective description of the physical state information of the target object before AI diagnosis. For example: “I have been having headaches recently” and the like.

[0058] The large model can be a large language model or a multi-modal large model. When the history medical information includes non-text information, a multi-modal large model can be used to analyze images, videos and other information to determine the physical state of the target object before AI diagnosis. Since the history medical information comes from historical medical records or examination reports and other medical diagnosis results, it is closer to objective facts than the subjective description of the target object.

[0059] Then, the large model can combine the history medical information and the state description information to generate a plurality of target questions that match the current state of the target object based on a deep understanding of the state and appeal of the target object.

[0060] Next, the large model simulates the diagnosis thinking of a doctor to sort the plurality of target questions and generate an interaction path. The interaction path indicates the interaction order of the plurality of target questions for interacting with the target object.

[0061] For example, the interaction path can be generated according to the diagnosis logic of pain nature, pain site, pain duration, severity, and accompanying symptoms, which facilitates the target object to describe the current physical state more clearly and explicitly, further improves the individualization degree of the AI diagnosis process, reduces the rejection of the target object to the mechanical diagnosis process, improves the diagnosis efficiency and the accuracy of the diagnosis result.

[0062] Then, the large model can be used to perform multiple rounds of interaction with the target object based on the multiple target questions in the interaction order, to obtain multiple target answers associated with the multiple target questions.

[0063] In some embodiments, the interaction path can be dynamically adjusted based on the answers of the target object. For example, the interaction path can be adjusted based on the risk level indicated by the diagnosis result. For example, when the answer of the target object to a certain question indicates that the target object is suffering from a high-risk disease, the order between the original questions to be interacted in the interaction path can be changed in the next round of diagnosis, and the question that can determine whether the target object is suffering from the high-risk disease is prioritized.

[0064] Finally, the large model can be used to fuse the multiple target answers. For example, the multiple target answers describing the accompanying symptoms can be sorted to obtain a text description with similar semantics but relatively concise. For example, the multiple target answers describing the severity can be screened to obtain a text description that clearly describes the severity level.

[0065] The target information describing the diagnosis result of the target object can be written into the corresponding field in the predetermined medical record template according to the fused text description. For example, the field describing drug allergies can be written into the field of medical history in the predetermined medical record template.

[0066] In the embodiments of the present disclosure, the large model is used to analyze the historical medical information and state description information of the target object to dynamically generate target questions for diagnosis that match the current state of the target object, and the interaction order between the multiple target questions for interaction with the target object is determined. According to the interaction order, the large model is used to perform multiple rounds of interaction with the target object based on the multiple target questions, which is more in line with the personalized needs of users than the fixed and mechanical diagnosis process, reduces repeated diagnosis, and further improves the diagnosis efficiency in the stage without intervention of doctors. In addition, the large model is used to fuse the target answers obtained through multiple rounds of interaction based on the natural language understanding capability, which can provide more targeted and accurate diagnosis result information for the formal diagnosis stage of doctors, and reduce the interference of redundant or incorrect information on the subsequent formal diagnosis process.

[0067] According to the embodiments of the present disclosure, the large model is used to analyze the historical medical information and state description information of the target object to obtain multiple target questions and an interaction path, which can include the following operations: the large model is used to analyze the state description information to generate multiple target questions associated with the state of the target object; and the large model is used to sort the multiple target questions based on the historical medical information to generate the interaction path.

[0068] Figure 3AThe illustration shows a schematic diagram of analyzing the historical medical information and status description information of a target object using a large model according to an embodiment of the present disclosure, to obtain multiple target questions and interaction paths.

[0069] like Figure 3A As shown, in this embodiment 300A, firstly, the state description information 301 is input into the large model 310 to generate a target problem 302 associated with the state of the target object. The target problem 302 may include: problem P1, problem P2, ..., problem P n .

[0070] Then, the target question 302 and historical medical information 303 are input into the large model 310 to generate an interaction path 304A. Based on a deep understanding of the historical medical information 303, the large model 310 can sort the questions in the target question 302 to obtain the interaction path 304A. For example... Figure 3A As shown, in interaction path 304A, question P n It is listed before question P1.

[0071] For example, historical medical information can include information that clearly indicates the outcome of the consultation, such as surgical history and drug allergy history. Based on a deep understanding of historical medical information, the large model 310 can also remove questions about the surgical history and drug allergy history of the target subject from the target question 302, and then sort the filtered questions to generate an interaction path 304A.

[0072] Based on historical medical records and status descriptions, a large-scale model is used in conjunction with past medical history and the target individual's subjective descriptions to achieve context-aware, dynamic generation of interaction paths that match the actual status of different target individuals. This reduces the probability of repeated consultations, improves consultation efficiency, and further enhances the user experience.

[0073] In some embodiments, using a large model to sort multiple target questions based on historical medical information and generate an interaction path may include the following operations: using a large model to analyze multiple target questions based on historical medical information and generate a priority for the multiple target questions; and using a large model to sort the multiple target questions based on the priority and generate an interaction path.

[0074] For example, priority can characterize the level of questions asked to determine whether or to what extent the physical condition of a target is abnormal.

[0075] For example, when symptom A and symptom B occur simultaneously, the target individual has a higher probability of having a certain disease. Therefore, questions related to the accompanying symptoms can be assigned a higher priority to prioritize consultations for high-risk diseases, quickly identify or rule out high-risk factors, and further improve consultation efficiency.

[0076] Exemplarily, the priority can also represent a similarity between the historical medical information and the plurality of target problems, and the higher the similarity, the lower the priority.

[0077] For example, when the historical medical information clearly records that the target object has a history of drug allergy, the question for inquiring about the drug allergy history can be placed after other questions based on the priority, so as to reduce the probability that the target object generates a resistance emotion due to repeated questioning of known information in the early stage of interaction, and further improve the user experience.

[0078] Since the target problem is generated based on the subjective description of the target object, in actual application scenarios, there is a large individual difference in the subjective description, and the questioning direction of the target problem generated by the large model based on the subjective description can be biased.

[0079] For example, the historical medical information clearly indicates that the target object has a history of neural headache, but the target object only involves a cold and fever in the subjective description, so that the questioning direction of the target problem generated by the large model points to judging whether the target object has influenza, and no question related to meningitis is generated.

[0080] Therefore, in addition to screening and sorting, in some embodiments, the large model can also be used to adjust the questioning direction of at least one target problem based on the historical medical information, to generate at least one adjusted question; and the large model can be used to adjust the interaction path based on the at least one adjusted question, to generate an adjusted first interaction path.

[0081] Figure 3B A schematic diagram of analyzing the historical medical information and the state description information of the target object by using the large model to obtain a plurality of target problems and an interaction path according to an embodiment of the present disclosure is schematically shown.

[0082] As Figure 3B shown, the embodiment 300B is based on the interaction path generated by the embodiment 300A, and the question P1 and the historical medical information 303 are input into the large model 301, and the large model 301 is used to adjust the questioning direction of the question P1 with the questioning direction bias based on deeply understanding the past medical history of the target object recorded in the historical medical information, to generate the question P 1’ .

[0083] For example, the question P1 can be "Do you have symptoms such as shortness of breath or cough when you have a fever?" The historical medical information can record that the target object was hospitalized for meningitis for more than a month. However, the questioning direction of the current question P1 focuses on judging whether the fever is caused by upper respiratory tract infection or lung infection. It does not focus on judging whether the target object has a relapse of meningitis.

[0084] Therefore, based on the deep understanding of the target object having a history of meningitis, the large model adjusts the questioning direction of question P1 to focus on judging whether the target object has a meningitis recurrence, and generates question P 1’ Can you ask, "Do you have severe, explosive headaches while you are feverish? Are the headache symptoms similar to those you experienced when you were last diagnosed with meningitis?"

[0085] Then, question P 1’ and the interaction path 304A are input into the large model 310, and the large model 310 is used to adjust the interaction path 304A by determining the priority between each question to generate an interaction path 304B. Compared with the interaction path 304A, the interaction path 304B not only replaces question P1, but also changes the interaction order between each question. 1’

[0086] Adjusting the questioning direction of the question based on the historical consultation information can reduce the influence of the subjective feeling difference of the target object on the questioning direction, and further improve the accuracy of the questioning direction in the interaction process.

[0087] In some embodiments, the large model can also be used to generate a first supplementary question associated with the historical consultation information based on the historical consultation information and a plurality of target questions, and to generate an adjusted second interaction path by adjusting the interaction path based on the first supplementary question.

[0088] Figure 3C The schematic diagram shows that the large model analyzes the historical consultation information and the state description information of the target object to obtain a plurality of target questions and an interaction path according to an embodiment of the present disclosure.

[0089] As Figure 3C shown, in this embodiment 300C, the target question 302 and the historical consultation information 303 can be input into the large model 310 to generate a supplementary question SP1 associated with the historical consultation information.

[0090] For example, the subjective description information of the target object can include back pain and long-term desk work. Based on the understanding of the above description information, the large model generates a target question mainly for judging whether the target object has an abnormal cervical spine.

[0091] However, the historical consultation information records that the target object has a history of heart disease, and the back is a reflex area of heart disease, so the back pain may be caused by a recurrence of heart disease. The large model can be used to generate a supplementary question SP1 for judging whether the heart disease has recurred based on the historical consultation information and a plurality of target questions.

[0092] ​Then, the interactive path 304A is input into the large model 310 with the supplementary question SP1, and the large model 310 is used to adjust the interactive path 304A by determining the priority between each question, to generate an interactive path 304C. Compared with the interactive path 304A, the interactive path 304C adds the supplementary question SP1 and changes the original interaction order in the interactive path 304A.

[0093] Based on the historical consultation information, the probability of missing diagnosis caused by the subjective feeling difference of the target object can be reduced, and the accuracy of interactive diagnosis is further improved, and the diagnosis efficiency is improved.

[0094] In addition to adjusting the interactive path before interaction, during the interaction process, the questions can also be dynamically adjusted according to the reply state of the target object, so as to further improve the interaction efficiency.

[0095] The following will be described in detail Figure 4A through 4D The process of dynamically adjusting interactive questions in the interaction process will be described in detail.

[0096] When the completeness of the reply of the target object to a certain question is low, if the same question is repeatedly asked the target object multiple times to obtain a complete reply, the target object is easy to cause resentment.

[0097] Therefore, in some embodiments, the multiple rounds of interaction between the large model and the target object according to the interaction order based on the plurality of target questions can include the following operations: interacting with the target object based on a first target question in the plurality of target questions according to the interaction order, to obtain a first initial reply associated with the first target question; in response to determining that the completeness of the first initial reply is less than the completeness threshold, generating a second supplementary question based on the first target question and the first initial reply by using the large model; interacting with the target object based on the second supplementary question to obtain a first supplementary reply; and in response to determining that the completeness of the first supplementary reply is greater than or equal to the completeness threshold, performing information fusion on the first initial reply and the first supplementary reply by using the large model to generate a first target reply in the plurality of target replies.

[0098] Figure 4A A schematic diagram of dynamically adjusting interactive questions in the interaction process according to an embodiment of the present disclosure is schematically shown.

[0099] As Figure 4A shown, the first target question can be “In addition to coughing and chest tightness, do you have any other discomforts? Such as fever, sore throat, or feeling that breathing is more tiring than usual?” The first initial reply of the target object to the first target question can be “The throat is a little uncomfortable”.

[0100] As can be seen from the first target question, a plurality of accompanying symptoms is provided in a non-selective manner in the question about the accompanying symptoms, and obviously, the answer to this question is not a single selective question. It can be pre-configured that the completeness threshold is reached only when at least two of the plurality of accompanying symptoms are replied to. Therefore, when the first initial reply of the target object only replies to the "sore throat" option, it can be determined that the completeness of the first initial reply is less than the completeness threshold.

[0101] At this time, a second supplementary question can be generated based on the first target question and the first initial reply by using a large model. The second supplementary question can be specifically asked about the part of the first target question that the target object has not replied to, and can further refine the description of the symptoms corresponding to the part that the target object has not replied to, further improve the understanding degree of the target object to the supplementary question, and reduce the probability of the target object making a supplementary reply in the case that there is a deviation in the understanding of the supplementary question.

[0102] As shown in Figure 4A , the second supplementary question can be "So about the other conditions I just asked, let me confirm again: Have you measured your body temperature recently? Do you have a fever? In addition, do you feel short of breath or rapid breathing when walking or climbing stairs?"

[0103] Therefore, compared with the first target question, the second supplementary question further refines the description of "fever" as "Have you measured your body temperature recently? Do you have a fever?" and further refines the description of "shortness of breath" as "Do you feel short of breath or rapid breathing when walking or climbing stairs?" By converting some medical professional terms into more descriptions that are easier for non-professionals to understand, the understanding degree of the target object to the supplementary question is improved.

[0104] The target object can make a supplementary reply to the incomplete part in the first initial reply based on the second supplementary question, for example: "No fever, feel short of breath when climbing stairs."

[0105] In the embodiments of the present disclosure, the large model can be used to fuse the information of "throat is a little uncomfortable" and "no fever, feel short of breath when climbing stairs" to generate a first target reply, for example, "no fever, mild sore throat and shortness of breath when exercising."

[0106] It can be understood that information fusion is a new reply text generated by fusing the semantics of multiple replies based on the semantic understanding of the large model to the multiple replies.

[0107] When the reply of the target object to a certain question is ambiguous, a question with clear direction can be generated to improve the interaction efficiency.

[0108] In some embodiments, following the interaction sequence, a large model is used to conduct multiple rounds of interaction with a target object based on multiple target questions to obtain multiple target responses associated with the multiple target questions. The method further includes the following operations: using the large model to perform intent recognition on a first initial response and generating an intent recognition result; in response to determining that the intent recognition result indicates an anomaly, using the large model to generate a third supplementary question based on the first initial response and the first target question; interacting with the target object based on the third supplementary question to obtain a second supplementary response; and in response to determining that the intent recognition result of the second supplementary response indicates normal, using the large model to perform information fusion on the second supplementary response and the first initial response to generate a second target response.

[0109] Figure 4B The illustration shows a schematic diagram of dynamically adjusting the interaction problem during the interaction process according to an embodiment of the present disclosure.

[0110] like Figure 4B As shown in embodiment 400B, the first target question could be "How long has this chest tightness lasted?" The first initial response could be "For a long time." In the target's first initial response, "for a long time" does not specify a concrete duration, making it a word with ambiguous intent. Therefore, it can be determined that the intent recognition result indicates an anomaly.

[0111] When generating a third supplementary question using a large model, to ensure the target audience provides a clear and explicit answer, explicit instructions can be given to the model, such as: "Please provide explicit options when generating the question." This allows for selective questioning of the target audience. For example: "Does 'how long' specifically mean more than three days? Or closer to a week or two?"

[0112] In this way, when the target audience interacts with the third supplementary question, they can choose from the options in the third supplementary question, thus giving a response with a clear intent. For example: "It's been almost a week."

[0113] Finally, a large model can be used to fuse information from "almost a week" and "for quite a while" to generate a second target response, such as: "Chest tightness has lasted for almost a week."

[0114] When the target audience's response intent is ambiguous, the natural language understanding capabilities of the large model can be used to generate specific supplementary questions, such as duration or pain level, so that the target audience can provide a clear response during the interaction, further improving interaction efficiency.

[0115] In real-world applications, the target's responses may be contradictory. In such cases, the contextual understanding capabilities of the large model can be leveraged to generate supplementary questions addressing the inconsistencies, thereby further improving the accuracy of the target information.

[0116] In some embodiments, the obtaining, by the large model, the plurality of target replies associated with the plurality of target questions based on the plurality of target questions and the target object in the plurality of rounds of interaction in the interaction order can further include: in response to determining that a matching degree between the at least two second initial replies of the at least two second target questions associated with each other is less than a predetermined matching degree threshold, generating, by the large model, a fourth supplementary question based on the at least two second target questions and the at least two second initial replies; interacting with the target object based on the fourth supplementary question to obtain a third supplementary reply; and modifying, by the large model, the at least two second initial replies based on the third supplementary reply to generate at least two third target replies.

[0117] Figure 4C An illustrative diagram of dynamically adjusting an interaction question in an interaction process according to an embodiment of the present disclosure is schematically shown.

[0118] As shown in the embodiment 400C, the at least two second target questions associated with each other can refer to that the question asking directions are associated or the same, for example, the question asking direction of “If you give a score of 0-10 to the pain of stomachache, what score do you give?” is “degree of pain”. The question asking direction of “Does the stomachache affect your life or work when it occurs?” is also “degree of pain”. Figure 4C

[0119] However, for the above two questions with the same question asking direction, the replies of the target object are “quite painful” and “actually not too bad, it does not affect my eating and work” respectively. Obviously, the degrees of pain indicated by the two replies are not matched, and it can be determined that the matching degree between the two replies is less than the predetermined matching degree threshold.

[0120] At this time, if it is explicitly pointed out that the replies of the target object before and after are contradictory, in the case of the target object suffering from illness, it can cause negative emotions of the target object. Therefore, when the large model is used based on the at least two second target questions and the at least two second initial replies, a reference example or a specific requirement for generating a supplementary question can be added in the Prompt, for example: when generating a supplementary question, do not directly point out that the replies of the target object before and after are contradictory, and in the generated question, the explanation reason why the target object may give contradictory replies is guessed, so as to guide the target object to give more accurate replies.

[0121] ​For example: "Thank you for your explanation. I want to confirm with you again, so that I can better understand your feelings. You mentioned that the pain is about 7 points, and you feel that it has little impact on daily life. Is this pain high in score, but it comes and goes, and you feel good when you are not in pain? Or is it a continuous but tolerable dull pain, so you have gotten used to it?" Let the target object feel understood and respected, try to reduce the negative emotions caused by being questioned, so that the target object can cooperate more with the AI diagnosis, further improve the efficiency of the diagnosis, and improve the user experience.

[0122] In this way, the target object can give a more accurate reply when facing the fourth supplementary question for interaction, such as "the pain comes and goes, and lasts for about 1 minute, then it slowly goes away".

[0123] Finally, the large model can correct "about 7 points, quite painful" based on the third supplementary reply, and the corrected target reply can be, for example, "the pain is 7 points when it comes and goes". The large model can correct "it doesn't affect me much in eating and working" based on the third supplementary reply, and the corrected target reply can be, for example, "the pain lasts for about 1 minute and then goes away, so it doesn't affect me much in eating and working".

[0124] In actual application scenarios, since AI diagnosis is aimed at a wide range of objects, there are large individual differences in the understanding of the objects to the questions. There are often many obscure medical terms in the questions of AI diagnosis interaction, which causes certain difficulties for the target object to understand. Therefore, when it is recognized that the target object has a low understanding of the question, the large model can be used to adjust the question expression to obtain an accurate reply from the target object.

[0125] In some embodiments, in the order of interaction, the large model is used to interact with the target object based on the plurality of target questions to obtain a plurality of target replies associated with the plurality of target questions, and the operation can further include: analyzing the first initial reply by the large model to generate an understanding degree of the target object to the first target question; modifying the first target question based on the understanding degree by the large model to generate a modified first question; and interacting with the target object based on the modified first question to obtain a fourth target reply.

[0126] Figure 4D A schematic diagram of dynamically adjusting an interactive question in an interaction process according to an embodiment of the present disclosure is schematically shown.

[0127] As Figure 4DAs shown, in this embodiment 400D, the first target question could be, "Besides pain, do you have any other uncomfortable feelings? For example, acid reflux, nausea, bloating, or any recent changes in your bowel movements?" The first initial response could be, "I don't quite understand the words you're using."

[0128] In some embodiments, a large model can be used to analyze the first initial response to generate a comprehension score, which can be represented by a score or a rating. If the first initial response explicitly contains fields indicating "unable to understand" or "does not comprehend," the comprehension score can be determined to be 1-2 points. The large model can also be used to analyze the logical correlation between the first initial response and the first target question. If the logical correlation is low, for example, if the answer is irrelevant, the comprehension score can also be determined to be low.

[0129] Then, a large model can be used to refine the first target question based on the level of understanding. For example, the way the first target question is expressed can be modified to generate a revised first question. The first target question and the revised first question have the same semantics.

[0130] In some embodiments, historical AI-generated consultation interactions that match a given level of comprehension can be pre-selected from a knowledge base. This knowledge base may store reference interactions matching different levels of comprehension. These reference interactions may be pre-generated using a large language model to simulate AI interactions between objects with varying levels of comprehension, complying with relevant legal regulations.

[0131] For example, when the target audience explicitly indicates a lack of understanding of medical terms like "acid reflux" or "bloating," the revised question can be translated into more colloquial terms, such as "acidic fluid rising in the throat." The revised first question could be, "Excuse me, let me ask you in a simpler way. Do you feel acidic fluid rising in your throat? Or feel nauseous? Also, have you had diarrhea or constipation recently?" Interacting with the target audience regarding the revised first question yields a more accurate fourth-target response, such as, "Oh, I do feel a little acidic fluid, but my bowel movements are normal." This further improves the target audience's understanding of the question, thereby increasing the probability of obtaining an accurate answer.

[0132] In addition to differences in comprehension levels, different individuals may also have differences in emotional management. Large models can be used to analyze the interaction styles of target individuals and adjust the way questions are asked so that AI consultations can proceed smoothly.

[0133] In some embodiments, using a large model to modify the first target question based on comprehension and generate a modified first question may further include the following operations: using the large model to analyze the first initial response and generate the interaction style of the target object; using the large model to modify the first target question based on the interaction style and generate a modified second question; and interacting with the target object based on the modified second question to obtain a fifth target response.

[0134] In this embodiment of the disclosure, interaction style can be used to characterize the preferences exhibited by the target object during the AI ​​consultation interaction process. For example, some prefer to respond by making choices, while others prefer to respond by providing verbal descriptions.

[0135] Figure 4E The illustration shows a schematic diagram of dynamically adjusting the interaction problem during the interaction process according to an embodiment of the present disclosure.

[0136] like Figure 4E As shown, in this embodiment 400E, the first target question can be "Please describe the specific feeling of your headache?" The first initial response can be "Here we go again, full of nonsense, completely useless."

[0137] By analyzing the first initial response using a large model and gaining a deep understanding of its semantics, it can be determined that the first initial response does not directly address the first target question, revealing that the target audience exhibits signs of frustration. This allows for the generation of the target audience's interaction style. For example, it could be a "concise and to the point" style. Target audiences of this type tend to use selective responses.

[0138] Therefore, when refining the first target question using a large model based on interaction style, the "feeling of a headache" in the first target question can be visualized, and options can be provided to vividly describe different types of headaches for the target audience to choose from. For example: "a throbbing pain like a tight band around the head." This allows the target audience to make a selection, thereby generating a fifth target response, such as: "somewhat like A."

[0139] Meanwhile, by analyzing the first initial response, it can be determined that the subject is in a state of agitation. Therefore, when revising the first target question, a reference example can be added to the Prompt. This reference example can include sentences that can resonate with the target subject and provide example text to soothe their emotions. This allows the revised second question generated by the large model to include sentences such as "I understand, headaches are very distressing" to alleviate the target subject's negative emotions and facilitate the smooth progress of the AI ​​consultation process.

[0140] Figure 5 A schematic diagram of an information generation method based on a large model according to an embodiment of the present disclosure is shown.

[0141] like Figure 5 As shown, in this embodiment 500, firstly, the status description information 301, historical medical information 303 and reference problem information 501 can be input into the large model 310, and the interaction path 304B can be output.

[0142] Then, interact with the target object according to the interaction order between the target questions indicated by interaction path 304B.

[0143] Finally, the target responses collected during the interaction are input into the large model 310 for information fusion to generate target information 502 that describes the consultation results of the target object.

[0144] In this embodiment of the disclosure, the target information 502 may include information to be consulted and target response information. Information to be consulted may include, for example, medical history, pain level, etc. Reference question information 501 may be candidate questions to be asked regarding the information to be consulted. These candidate questions may be collected from publicly available medical records in medical books or materials, or may be determined based on medical record templates from different departments in a hospital.

[0145] For example, a reference question regarding "pain level" could be, "If you were to rate your pain on a scale of 1 to 10, what score would you consider appropriate?"

[0146] In some embodiments, using a large model to fuse information from multiple target responses to generate target information describing the consultation results of a target object may include the following operations: extracting target fields associated with predetermined fields and the sources of the target fields from multiple target responses; and using a large model to fuse information from the target fields and the sources of the target fields to generate target information.

[0147] In this embodiment of the disclosure, the predetermined field can refer to the field corresponding to the information to be collected in the AI ​​consultation result template. For example, the onset of the illness, symptom characteristics, severity, etc. The target field associated with the predetermined field can refer to the field in multiple target responses used to describe the specific content corresponding to the predetermined field. For example, the target response "no fever, mild sore throat and shortness of breath during exercise" is a target field associated with the predetermined field "symptom characteristics". The source of this target field can be "AI interaction".

[0148] In addition, the source of the target field can also include historical medical information, such as examination reports and laboratory test results.

[0149] In some embodiments, using a large model to fuse target information and the source of target information to generate target information may include the following operations: using a large model to fuse target fields and the source of target fields based on a predetermined format to generate target information.

[0150] The predefined format refers to the layout or structure of the fields in the AI ​​consultation result template. In real-world applications, different hospitals and departments may require different format templates. Therefore, the template can be configured according to the specific format requirements of a particular hospital or department during deployment.

[0151] Figure 6 A schematic diagram illustrating target information for characterizing the consultation results of a target object according to an embodiment of the present disclosure is shown.

[0152] like Figure 6 As shown, this embodiment illustrates the specific content of the target information 502 obtained in embodiment 500. The fields marked with gray shading in the target information 502 are all predetermined fields.

[0153] When generating target information, the target field and the source of the target field associated with the predetermined field can be extracted from multiple target responses to determine the specific descriptive information corresponding to each predetermined field, and the source of the information is marked in each specific descriptive information so that the information can be traced during the doctor's formal consultation.

[0154] Then, the AI ​​consultation result template and multiple target responses can be input into the large model. Using the large model, information fusion is performed on the target fields and their sources based on a predetermined format to generate target information 502.

[0155] Figure 7 A flowchart illustrating a training method for a large model according to an embodiment of the present disclosure is shown schematically.

[0156] like Figure 7 As shown, the training method 700 may include operations S710 to S720.

[0157] When operating the S710, the large model is used to analyze the sample object's sample history and sample status description information to obtain multiple sample questions and sample interaction paths.

[0158] When operating the S720, based on the objective loss function, a large model is trained using multiple sample questions and multiple sample labels to obtain a trained large model.

[0159] According to the embodiments of this disclosure, the sample historical medical information, sample status description information, and multiple sample questions and sample interaction paths are the same as the historical medical information, status description information, target questions, and interaction paths defined in the large model-based information generation method described above, and will not be repeated here.

[0160] According to embodiments of this disclosure, the sample interaction path indicates the interaction order among multiple sample questions used to interact with a sample object.

[0161] In this embodiment of the disclosure, the multiple sample labels are multiple interactive questions to be posed to the sample objects, generated by calling at least two pre-trained language models through multiple rounds of simulated doctor-patient interaction.

[0162] For example, one pre-trained language model can simulate a doctor, and another pre-trained language model can simulate a patient to interact, in order to build rich sample data.

[0163] In the embodiments disclosed herein, the target loss function may be the cross-entropy loss function or other loss functions, and the undisclosed embodiments do not specifically limit this.

[0164] During training, the initial large model generates the next round of interaction questions for a certain answer of the sample object. It can calculate the loss value with the sample interaction questions corresponding to a certain answer in the sample labels, so that the initial large model can learn the consultation logic in the AI ​​consultation process.

[0165] For example, sample labels can include multi-turn interactive dialogues, Q 10 -A1、Q 20 -A2…, during training, when the sample object is targeted by Q 10 When the response is A1, then the Q-value in the sample interaction path generated by the initial large model can be calculated based on the loss function. 10 The following Q2 and Q 20 The loss value between the two is used to perform supervised fine-tuning of the initial large model based on the loss value until the loss value converges or the maximum number of iterations is reached, thus obtaining the trained large model.

[0166] The initial large model can be a multimodal large model to enable the parsing of images, text, audio and video information in historical medical records.

[0167] Because the consultation logic and the information required to be collected during the consultation process vary across different departments and hospitals in real-world applications, a reference information set can be constructed based on a large number of publicly available medical books and publicly available medical records (which have already been anonymized) to further improve the computational capabilities of the finely trained large-scale model. When constructing the reference information set, the natural language understanding capabilities of the pre-trained language model can also be used to perform text sorting and optimization on the publicly available medical books and medical records.

[0168] In this way, when generating sample data, we can first obtain sample interaction problems that match the state of the simulated patient from the reference information set based on the patient profile.

[0169] By calling at least two pre-trained language models to generate multiple interactive questions for posing to sample objects through multiple rounds of simulated doctor-patient interactions, the impact of sample scarcity or uneven distribution of different types of samples on model training efficiency can be reduced.

[0170] The following is combined with Figure 8A and Figure 8B The training methods for large models are described in detail.

[0171] In some embodiments, the above method may further include the following operations: using a first pre-trained language model to generate sample interaction questions by simulating a doctor based on predetermined information to be collected; using a second pre-trained language model to generate sample responses for replying to the sample interaction questions by simulating a patient based on the sample interaction questions, predetermined interaction styles, sample historical medical information, and sample status description information; and using the first pre-trained language model to generate sample labels by simulating a doctor based on the sample responses and predetermined information to be collected.

[0172] In the embodiments of this disclosure, the predetermined information to be collected can be determined based on predetermined fields in the medical record templates of different departments in the actual application scenario. In the embodiments of this disclosure, the predetermined interaction style can be configured according to actual needs, such as "concise", "nitpicking", "fuzzy reply", etc.

[0173] Figure 8A A schematic diagram illustrating a training method for a large model according to an embodiment of the present disclosure is shown.

[0174] like Figure 8A As shown in the embodiment 800A, firstly, a pre-trained language model LLM can be used to sort the text of medical books 810 and public medical records 820 to generate a reference information set 830.

[0175] Then, information to be collected 804 can be randomly sampled from the reference information set 830, and the information to be collected 804 can be input into the first pre-trained language model LLM1 used to simulate doctors to generate sample interaction questions.

[0176] Simultaneously, patient information 1201 can be randomly sampled from the reference information set 830, and the patient profile 802, interaction style 803, patient information 801, and database 840 storing examination reports or test results are input into the second pre-trained language model LLM2 used to simulate the patient to generate sample responses to the sample interaction questions.

[0177] Next, using the first pre-trained language model LLM1, sample labels are generated by simulating a doctor's response based on the sample and the pre-defined information to be collected.

[0178] After multiple rounds of simulated interaction, the interaction data collected from these rounds can be used as sample data 805 to train the initial large model 850.

[0179] When generating sample responses, pre-trained language models can be used to randomly generate or optimize patient profiles and interaction styles. Patient profiles can refer to the basic state information of simulated patients, so that the pre-trained language model used to simulate patients can interact with another pre-trained language model based on the interaction style, which further improves the similarity between simulated interaction scenarios and real interaction scenarios, and further enhances the robustness and generalization ability of the large model.

[0180] To reduce the interference of the randomness of the pre-trained model output on the training process, a third pre-trained language model can be introduced to supervise the interaction process. For example, the above training method can also include the following operations: using the third pre-trained language model to analyze the sample interaction questions and sample responses, and generating analysis results; and in response to the determination that the analysis results indicate that the sample responses match the sample interaction questions, using the first pre-trained language model, by simulating a doctor based on the sample responses and predetermined information to be collected, generating sample labels.

[0181] Figure 8B A schematic diagram illustrating a training method for a large model according to an embodiment of the present disclosure is shown.

[0182] As shown in Figure 800B, this embodiment 800B adds operation S810 to 800A to determine whether the sample interaction question 8051 and sample response 8052 match. If they match, the interaction data collected from multiple rounds of interaction is used as sample data 805 to train the initial large model 850. If they do not match, the mismatched sample interaction questions and responses are discarded. This achieves quality control of the interaction data generated during the simulated interaction process, reducing the interference of erroneous data on training accuracy.

[0183] Figure 9 A block diagram of a large-model-based information generation apparatus according to an embodiment of the present disclosure is shown schematically.

[0184] like Figure 9 As shown, the information generation device 900 may include: a first analysis module 910, an interaction module 920, and a generation module 930.

[0185] The first analysis module 910 is used to analyze the historical medical information and status description information of the target object using a large model to obtain multiple target questions and interaction paths; wherein, the interaction path indicates the interaction order between the multiple target questions used to interact with the target object.

[0186] Interaction module 920 is used to conduct multiple rounds of interaction with target objects based on multiple target questions according to the interaction sequence, and obtain multiple target responses associated with multiple target questions.

[0187] The generation module 930 is used to fuse information from multiple target responses using a large model to generate target information that describes the consultation results of the target object.

[0188] According to an embodiment of this disclosure, the first analysis module 910 includes a first analysis submodule and a sorting submodule.

[0189] The first analysis submodule is used to analyze the state description information using a large model and generate multiple target questions associated with the state of the target object.

[0190] The sorting submodule is used to sort multiple target questions based on historical medical information using a large model and generate interaction paths.

[0191] According to embodiments of this disclosure, the sorting submodule includes: a first analysis unit and a sorting unit.

[0192] The first analysis unit is used to analyze multiple target questions based on historical medical information using a large model, and to generate priorities for the multiple target questions.

[0193] The sorting unit is used to prioritize multiple target problems using a large model and generate interaction paths.

[0194] According to embodiments of this disclosure, the first analysis module 910 further includes a problem adjustment submodule and a path adjustment submodule.

[0195] The question adjustment submodule is used to adjust the questioning direction of at least one target question based on historical medical information using a large model, and generate at least one adjusted question.

[0196] The path adjustment submodule is used to adjust the interaction path based on at least one adjusted question using the large model, and generate the adjusted first interaction path.

[0197] According to an embodiment of this disclosure, the first analysis module 910 further includes: a first generation submodule and a second generation submodule.

[0198] The first generation submodule is used to generate a first supplementary question associated with the historical medical information by utilizing the large model based on historical medical information and multiple target questions.

[0199] The second generation submodule is used to adjust the interaction path based on the first supplementary question using the large model, and generate the adjusted second interaction path.

[0200] According to embodiments of this disclosure, the interaction module 920 may include: a first interaction submodule, a first supplementary submodule, a second interaction submodule, and a first fusion submodule.

[0201] The first interaction submodule is used to interact with the target object based on the first target question among multiple target questions, according to the interaction order, and to obtain the first initial response associated with the first target question.

[0202] The first supplementary submodule is used to generate a second supplementary question based on the first target question and the first initial response in response to determining that the completeness of the first initial response is less than a completeness threshold.

[0203] The second interaction submodule is used to interact with the target object based on the second supplementary question to obtain the first supplementary response.

[0204] The first fusion sub-Euclidean is used to fuse information between the first initial response and the first supplementary response in response to determining that the completeness of the first supplementary response is greater than or equal to the completeness threshold, and to generate the first target response among multiple target responses by using a large model.

[0205] According to embodiments of this disclosure, the interaction module 920 may further include: an identification submodule, a second supplementary submodule, a third interaction submodule, and a second fusion submodule.

[0206] The identification submodule is used to perform intent identification on the first initial response using a large model and generate intent identification results.

[0207] The second supplementary submodule is used to generate a third supplementary question based on the first initial response and the first target question in response to the determination that the intent recognition result indicates an anomaly.

[0208] The third interaction submodule is used to interact with the target object based on the third supplementary question to obtain the second supplementary response.

[0209] The second fusion submodule is used to respond to the determination that the intent recognition result of the second supplementary response is normal, and to use the large model to fuse the information of the second supplementary response and the first initial response to generate the second target response.

[0210] According to embodiments of this disclosure, the interaction module 920 may further include: a third supplementary submodule, a fourth interaction submodule, and a first correction submodule.

[0211] The third supplementary submodule is used to generate a fourth supplementary question based on the at least two second initial responses to at least two related second target questions in response to the determination that the matching degree between them is less than a predetermined matching degree threshold.

[0212] The fourth interaction submodule is used to interact with the target object based on the fourth supplementary question to obtain the third supplementary response.

[0213] The first correction submodule is used to use the large model to correct at least two second initial responses based on the third supplementary response, and generate at least two third target responses.

[0214] According to embodiments of this disclosure, the interaction module 920 may further include: a first analysis submodule, a second correction submodule, and a fifth interaction submodule.

[0215] The first analysis submodule is used to analyze the first initial response using a large model to generate the target object's understanding of the first target question.

[0216] The first correction submodule is used to modify the first target question based on the understanding of the large model, generating a modified first question; the first target question and the modified first question have the same semantics.

[0217] The fifth interaction submodule is used to interact with the target object based on the revised first question to obtain the fourth target response.

[0218] According to embodiments of this disclosure, the interaction module 920 may further include: a second analysis submodule, a second correction submodule, and a sixth interaction submodule.

[0219] The second analysis submodule is used to analyze the first initial response using a large model to generate the interaction style of the target object.

[0220] The second correction submodule is used to modify the first target problem based on the interaction style using the large model, and generate the modified second problem.

[0221] The sixth interaction submodule is used to interact with the target object based on the revised second question to obtain the fifth target response.

[0222] According to embodiments of this disclosure, the generation module 930 may include an extraction submodule and a third fusion submodule.

[0223] The extraction submodule is used to extract the target fields and the source of the target fields that are associated with the predetermined fields from multiple target responses.

[0224] The third fusion submodule is used to fuse information from the target field and its source using the large model to generate target information.

[0225] According to embodiments of this disclosure, the third fusion submodule is used to utilize a large model to fuse information about the target field and its source based on a predetermined format, thereby generating target information.

[0226] Figure 10 A block diagram of a large model training apparatus according to an embodiment of the present disclosure is shown schematically.

[0227] like Figure 10 As shown, the training device 1000 may include a second analysis module 1010 and a training module 1020.

[0228] The second analysis module 1010 is used to analyze the sample historical medical information and sample status description information of the sample object using the initial large model to obtain multiple sample questions and sample interaction paths; wherein, the sample interaction path indicates the interaction order between the multiple sample questions used to interact with the sample object.

[0229] Training module 1020 is used to train an initial large model based on a target loss function, using multiple sample questions and multiple sample labels to obtain a trained large model; wherein, multiple sample labels are multiple interactive questions to be posed to sample objects by calling at least two pre-trained language models through multiple rounds of simulated doctor-patient interaction.

[0230] According to embodiments of this disclosure, the training device 1000 may further include: a first simulation module, a second simulation module, and a generation module.

[0231] The first simulation module is used to generate sample interaction questions by simulating a doctor based on predetermined information to be collected, using a first pre-trained language model.

[0232] The second simulation module is used to generate sample responses to sample interaction questions by simulating patients based on sample interaction questions, predetermined interaction styles, sample historical medical information, and sample status description information, using the second pre-trained language model.

[0233] The generation module is used to generate sample labels by simulating a doctor's response based on the sample and the pre-defined information to be collected, using a first pre-trained language model.

[0234] According to embodiments of this disclosure, the training device 1000 may further include a third analysis submodule and a generation submodule.

[0235] The third analysis submodule is used to analyze sample interaction questions and sample responses using the third pre-trained language model and generate analysis results.

[0236] The generation submodule, in response to the determination of the analysis results indicating that the sample response matches the sample interaction question, uses a first pre-trained language model to generate sample labels by simulating a doctor based on the sample response and predetermined information to be collected.

[0237] Figure 11 A block diagram of an intelligent agent according to an embodiment of the present disclosure is shown schematically.

[0238] In embodiments of this disclosure, the von Neumann architecture in modern computer theory is inspired, such as... Figure 11 As shown, the AI ​​agent 1100 may include three core modules: an input module 1110, an output module 1120, and a processing module 1130. The processing module 1130 may include a control unit 1131, a storage unit 1132, and a computing unit 1133.

[0239] The input module 1110 is responsible for receiving or sensing information such as queries, requests, instructions, signals, or data from the outside world (e.g., users or the external environment), and converting it into a format that the AI ​​agent 1100 can understand and process. The input module 1110 is the primary link for the AI ​​agent 1100 to interact with the outside world. It enables the AI ​​agent 1100 to efficiently and accurately obtain the necessary "sensory" information from the outside world and respond to this information.

[0240] In the example, the input information received by the input module 1110 can be the historical medical records and status description information of the target object described above.

[0241] In the example, processing module 1130 is the core support for the AI ​​agent 1100's ability to handle complex tasks. Processing module 1130 can determine the target task based on the input information received by input module 1110, determine a large model based on the target task, and execute the information generation method based on the large model described above by calling the large model to output target information describing the consultation results of the target object.

[0242] In the example, the control unit 1131 in the processing module 1130 will continuously interact with the storage unit 1132, the arithmetic unit 1133, and / or the output module 1120 during operation. However, it should be noted that in the embodiments of this disclosure, the control unit 1131 initiates communication with the storage unit 1132, the arithmetic unit 1133, and / or the output module 1120 as a single initiator, and there is no communication coupling between the storage unit 1132, the arithmetic unit 1133, and the output module 1120.

[0243] In the example, the performance of control unit 1131 is closely related to the large model on which AI agent 1100 is based. To fully leverage the capabilities of the large language model, the internal structure of control unit 1131 can be designed to be highly configurable and scalable to handle various types of tasks and requirements in real-world scenarios.

[0244] Storage unit 1132 can be responsible for remembering information such as historical dialogues and event streams. As mentioned above, the configuration information, target text, and data resources generated in each round can be included in storage unit 1132.

[0245] In the example, after receiving a configuration generation request, AI agent 1100 can determine the configuration intent from the initial text using an intent recognition model. The configuration intent can be stored in storage unit 1132. AI agent 1100 can retrieve relevant data resources from storage unit 1132 and feed them back to control unit 1131. Then, control unit 1131 can use the returned data resources to obtain configuration data corresponding to the initial text. It can also retrieve relevant text data from storage unit 1132 and feed it back to control unit 1131. Then, control unit 1131 can use the returned text data to obtain the target text and pass the target text and configuration data to output module 1120.

[0246] The computation unit 1133 can be viewed as a predefined tool library. As mentioned above, renderers and display controls can be included in the computation unit 1133.

[0247] In the example, when the AI ​​agent 1100 needs to render multiple output data, it can call the relevant renderer and display controls from the computing unit 1133 and feed them back to the control unit 1132. Then, the control unit 1132 can use the fed-back renderer and display controls to render the first search result and pass the first search result to the output module 1120. It is understandable that although large language models have excellent language understanding and generation capabilities, like humans, the tasks they can solve without any tools are very limited. When the AI ​​agent 1100 is given the ability to call tools, it can perform tasks such as using a calculator to complete mathematical calculations, using Python to complete data analysis, and using a search engine to complete prediction tasks.

[0248] In the example, output module 1120 can output the target information described above for describing the consultation results of the target object.

[0249] The AI ​​agent 1100 according to the embodiments of this disclosure can simply and effectively improve the level of intelligence, and enhance flexibility and versatility.

[0250] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0251] According to an embodiment of the present disclosure, an electronic device 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, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods described above.

[0252] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described above.

[0253] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0254] Figure 12A schematic block diagram of an example electronic device 1200 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0255] like Figure 12 As shown, device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1202 or a computer program loaded from storage unit 1208 into random access memory (RAM) 1203. The RAM 1203 may also store various programs and data required for the operation of device 1200. The computing unit 1201, ROM 1202, and RAM 1203 are interconnected via bus 1204. Input / output (I / O) interface 1205 is also connected to bus 1204.

[0256] Multiple components in device 1200 are connected to I / O interface 1205, including: input unit 1206, such as keyboard, mouse, etc.; output unit 1207, such as various types of monitors, speakers, etc.; storage unit 1208, such as disk, optical disk, etc.; and communication unit 1209, such as network card, modem, wireless transceiver, etc. Communication unit 1209 allows device 1200 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0257] The computing unit 1201 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1201 performs the various methods and processes described above, such as large model-based information generation methods or large model training methods. For example, in some embodiments, the large model-based information generation methods or large model training methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1208. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1200 via ROM 1202 and / or communication unit 1209. When the computer program is loaded into RAM 1203 and executed by the computing unit 1201, one or more steps of the large model-based information generation methods or large model training methods described above can be performed. Alternatively, in other embodiments, the computing unit 1201 may be configured by any other suitable means (e.g., by means of firmware) to perform a large model-based information generation method or a large model training method.

[0258] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0259] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0260] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on 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 fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0261] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0262] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0263] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0264] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0265] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An information generation method based on a large model, comprising: By analyzing the historical medical records and status descriptions of the target object using a large model, multiple target questions and interaction paths are obtained; wherein, the interaction path indicates the interaction order among the multiple target questions used to interact with the target object. Following the interaction sequence, the large model engages with the target object in multiple rounds based on the multiple target questions, yielding multiple target responses associated with the multiple target questions; and The large model is used to fuse information from the multiple target responses to generate target information that describes the consultation results of the target object.

2. The method according to claim 1, wherein, The method utilizes a large model to analyze the target object's historical medical information and status description information, resulting in multiple target questions and interaction paths, including: The state description information is analyzed using the large model to generate multiple target questions associated with the state of the target object; and The large model is used to sort the multiple target questions based on the historical medical information to generate the interaction path.

3. The method according to claim 2, wherein, The step of using the large model to sort the multiple target questions based on the historical medical information and generating the interaction path includes: The large model is used to analyze the multiple target questions based on the historical medical information, and a priority is generated for the multiple target questions; and Based on the priority, the multiple target questions are sorted using the large model to generate the interaction path.

4. The method according to claim 2 or 3, wherein, The method of analyzing the historical medical information and status description information of the target object using a large model to obtain multiple target questions and interaction paths also includes: Using the large model based on the historical medical information, the questioning direction of at least one target question is adjusted to generate at least one adjusted question; and The interaction path is adjusted based on the at least one adjusted question using the large model to generate an adjusted first interaction path.

5. The method according to any one of claims 2-4, wherein, The method of analyzing the historical medical information and status description information of the target object using a large model to obtain multiple target questions and interaction paths also includes: Using the large model based on the historical medical records and the multiple target questions, a first supplementary question associated with the historical medical records is generated; and The interaction path is adjusted based on the first supplementary question using the large model to generate the adjusted second interaction path.

6. The method according to claim 1, wherein, The step of using the large model to conduct multiple rounds of interaction with the target object based on the multiple target questions, according to the interaction sequence, to obtain multiple target responses associated with the multiple target questions, includes: According to the interaction sequence, the target object is interacted with based on the first target question among the plurality of target questions to obtain a first initial response associated with the first target question; In response to determining that the completeness of the first initial response is less than a completeness threshold, a second supplementary question is generated using the large model based on the first target question and the first initial response; Based on the second supplementary question, the interaction with the target object yields a first supplementary response; and In response to determining that the completeness of the first supplementary response is greater than or equal to the completeness threshold, the large model is used to fuse the information of the first initial response and the first supplementary response to generate the first target response among the multiple target responses.

7. The method according to claim 6, wherein, The step of using the large model to conduct multiple rounds of interaction with the target object based on the multiple target questions, according to the interaction sequence, to obtain multiple target responses associated with the multiple target questions, further includes: The large model is used to perform intent recognition on the first initial response, and intent recognition results are generated. In response to determining that the intent recognition result indicates an anomaly, a third supplementary question is generated using the large model based on the first initial response and the first target question; Based on the interaction with the target object regarding the third supplementary question, a second supplementary response is obtained; and In response to the determination that the intent recognition result of the second supplementary response is normal, the second supplementary response and the first initial response are fused using the large model to generate a second target response.

8. The method according to claim 6 or 7, wherein, The step of using the large model to conduct multiple rounds of interaction with the target object based on the multiple target questions, according to the interaction sequence, to obtain multiple target responses associated with the multiple target questions, further includes: In response to the determination that the matching degree between at least two second initial responses to at least two related second target questions is less than a predetermined matching degree threshold, a fourth supplementary question is generated using the large model based on the at least two second target questions and the at least two second initial responses; Based on the interaction with the target object regarding the fourth supplementary question, a third supplementary response is obtained; and Using the large model, the at least two second initial responses are modified based on the third supplementary response to generate at least two third target responses.

9. The method according to claim 6, wherein, The step of using the large model to conduct multiple rounds of interaction with the target object based on the multiple target questions, according to the interaction sequence, to obtain multiple target responses associated with the multiple target questions, further includes: The first initial response is analyzed using the large model to generate the target object's understanding of the first target question; Using the large model, the first target problem is modified based on the level of understanding to generate a modified first problem; the first target problem and the modified first problem have the same semantics; and Based on the revised first question, the interaction with the target object yields the fourth target response.

10. The method according to claim 9, wherein, The step of using the large model to modify the first target problem based on the level of understanding, and generating a modified first problem, further includes: The first initial response is analyzed using the large model to generate the interaction style of the target object; Using the large model, the first target problem is modified based on the interaction style to generate a modified second problem; and Based on the revised second question, an interaction is conducted with the target object to obtain a response to the fifth target.

11. The method according to any one of claims 1-10, wherein, The process of fusing information from the multiple target responses using the large model to generate target information describing the consultation results of the target object includes: Extract the target field associated with the predetermined field and the source of the target field from multiple target responses; and The target information is generated by fusing information from the target field and its source using the large model.

12. The method according to claim 11, wherein, The process of using the large model to fuse the target information and its sources to generate the target information includes: Using the large model, information is fused between the target field and its source based on a predetermined format to generate the target information.

13. A training method for a large model, comprising: The initial large model is used to analyze the sample historical medical information and sample status description information of the sample objects to obtain multiple sample questions and sample interaction paths; wherein, the sample interaction path indicates the interaction order among the multiple sample questions used to interact with the sample object; and Based on the objective loss function, the initial large model is trained using the multiple sample questions and multiple sample labels to obtain the trained large model; The multiple sample labels are generated by calling at least two pre-trained language models through multiple rounds of simulated doctor-patient interactions, and are used to pose multiple interactive questions to the sample objects.

14. The method of claim 13, further comprising: Using the first pre-trained language model, sample interaction questions are generated by simulating a doctor based on predetermined information to be collected; Using a second pre-trained language model, a sample response to the sample interaction question is generated by simulating a patient's interaction question based on the sample interaction question, a predetermined interaction style, sample historical medical information, and sample status description information. as well as Using the first pre-trained language model, the sample labels are generated by simulating a doctor's response based on the sample and the predetermined information to be collected.

15. The method according to claim 14, wherein, Also includes: The sample interaction questions and responses are analyzed using a third pre-trained language model to generate analysis results; and In response to determining that the analysis results indicate that the sample response matches the sample interaction question, the sample label is generated by simulating a doctor based on the sample response and the predetermined information to be collected, using the first pre-trained language model.

16. An information generation device based on a large model, comprising: The first analysis module is used to analyze the historical medical information and status description information of the target object using a large model to obtain multiple target questions and interaction paths; wherein, the interaction path indicates the interaction order between the multiple target questions used to interact with the target object; An interaction module is configured to, according to the interaction sequence, utilize the large model to conduct multiple rounds of interaction with the target object based on the multiple target questions, to obtain multiple target responses associated with the multiple target questions; and The generation module is used to perform information fusion on the multiple target responses using the large model to generate target information describing the consultation results of the target object.

17. A training device for a large model, comprising: The second analysis module is used to analyze the sample historical medical information and sample status description information of the sample objects using the initial large model, to obtain multiple sample questions and sample interaction paths; wherein, the sample interaction path indicates the interaction order among the multiple sample questions used to interact with the sample object; and The training module is used to train the initial large model based on the target loss function, using the multiple sample questions and multiple sample labels, to obtain the trained large model. The multiple sample labels are generated by calling at least two pre-trained language models through multiple rounds of simulated doctor-patient interactions, and are used to pose multiple interactive questions to the sample objects.

18. An intelligent agent, comprising: The input module is used to receive the target object's historical medical records and status description information; The processing module is configured to analyze the requirements and determine the target task based on the historical medical information and the status description information received by the input module, determine the target large model based on the target task, and execute the method described in any one of claims 1-12 by calling the target large model to obtain target information for describing the consultation results of the target object; as well as An output module is used to output the target information obtained by the processing module.

19. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-15.

20. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-15.

21. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-15.

Images