Medical question and answer model training method and device, electronic equipment, storage medium and program product

Abstract

The invention discloses a medical question and answer model training method and device, electronic equipment, a storage medium and a program product. The medical question and answer model training method comprises the steps of obtaining a to-be-trained initial medical question and answer model; based on the medical theoretical question-answer pair, performing first-stage model parameter adjustment on the initial medical question-answer model to obtain a first medical question-answer model; based on the medical record sample data, performing second-stage model parameter adjustment on the first medical question and answer model to obtain a second medical question and answer model; based on a medical question and answer sample group, model parameter adjustment of a third stage is carried out on the second medical question and answer model to obtain a third medical question and answer model, the medical question and answer sample group comprises the medical question and positive reply information and negative reply information of the medical question, and multi-stage training of the medical question and answer model is achieved. And the accuracy of the medical question-answer model is improved.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a training method, apparatus, electronic device, storage medium, and program product for a medical question-and-answer model. Background Technology

[0002] In the medical field, Large Language Models (LLMs) based on the Transformers architecture play a crucial role in medical question-answering scenarios. Therefore, the training of LLMs has a significant impact on their accuracy.

[0003] In existing technologies, multi-source heterogeneous data, such as training data of basic theoretical types and training data of clinical cases, are mixed and then used for single-stage training of LLM. During single-stage training, LLM cannot be trained specifically for different types of training data, resulting in low accuracy. Summary of the Invention

[0004] This invention provides a training method, apparatus, electronic device, storage medium, and program product for a medical question-answering model, to achieve multi-stage training of the medical question-answering model.

[0005] According to one aspect of the present invention, a method for training a medical question-answering model is provided, comprising:

[0006] Obtain the initial medical question-answering model to be trained;

[0007] Based on medical theoretical question-and-answer pairs, the initial medical question-and-answer model is adjusted in the first stage to obtain the first medical question-and-answer model;

[0008] Based on medical record sample data, the model parameters of the first medical question-answering model were adjusted in the second stage to obtain the second medical question-answering model;

[0009] Based on the medical question-and-answer sample set, the model parameters of the second medical question-and-answer model are adjusted in the third stage to obtain the third medical question-and-answer model. The medical question-and-answer sample set includes medical questions and positive and negative response information for medical questions.

[0010] According to another aspect of the present invention, a training apparatus for a medical question-answering model is provided, comprising:

[0011] The initial medical question-answering model acquisition module is used to acquire the initial medical question-answering model to be trained.

[0012] The first-stage training module is used to adjust the model parameters of the initial medical question-and-answer model based on medical theory question-and-answer pairs, so as to obtain the first medical question-and-answer model.

[0013] The second-stage training module is used to adjust the model parameters of the first medical question-answering model based on medical record sample data to obtain the second medical question-answering model.

[0014] The third-stage training module is used to adjust the model parameters of the second medical question-answering model in the third stage based on the medical question-answering sample group to obtain the third medical question-answering model. The medical question-answering sample group includes medical questions and positive and negative answer information for medical questions.

[0015] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0016] At least one processor; and

[0017] A memory that is communicatively connected to at least one processor; wherein,

[0018] The memory stores a computer program that can be executed by at least one processor, such that the at least one processor is able to execute the training method of the medical question-answering model according to any embodiment of the present invention.

[0019] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the training method of the medical question-answering model provided in any embodiment of the present invention.

[0020] According to another aspect of the present invention, a method for training a medical question-answering model as provided in any embodiment of the present invention is provided, which includes a computer program that, when executed by a processor, implements a method for training such a model when executing a computer program.

[0021] The technical solution of this invention provides a model foundation for subsequent analysis and processing by acquiring an initial medical question-and-answer model to be trained. Based on medical theoretical question-and-answer pairs, the initial medical question-and-answer model undergoes a first-stage parameter adjustment to obtain a first medical question-and-answer model, achieving the first stage of training and improving the accuracy of the first medical question-and-answer model in the medical entity recognition dimension. Based on medical record sample data, the first medical question-and-answer model undergoes a second-stage parameter adjustment to obtain a second medical question-and-answer model, achieving the second stage of model training and improving the accuracy of the second medical question-and-answer model in the case reasoning dimension. Based on a medical question-and-answer sample set, the second medical question-and-answer model undergoes a third-stage parameter adjustment to obtain a third medical question-and-answer model. The medical question-and-answer sample set includes medical questions and positive and negative responses to those questions, achieving multi-stage training of the medical question-and-answer model. This solves the problem of low accuracy in existing medical question-and-answer models and improves the accuracy of the medical question-and-answer model while ensuring that the output of the medical question-and-answer model complies with medical safety and ethical constraints.

[0022] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a training method for a medical question-answering model provided in Embodiment 1 of the present invention;

[0025] Figure 2 This is a flowchart of a training method for a medical question-answering model provided in Embodiment 2 of the present invention;

[0026] Figure 3 This is a schematic diagram of the structure of a training device for a medical question-and-answer model provided in Embodiment 3 of the present invention;

[0027] Figure 4 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] Example 1

[0031] Figure 1 This is a flowchart of a training method for a medical question-answering model provided in Embodiment 1 of the present invention. This embodiment is applicable to situations involving multi-stage training of a medical question-answering model. The method can be executed by a training device for the medical question-answering model, which can be implemented in hardware and / or software. This training device can be configured in the electronic device provided in this embodiment of the invention. The electronic device can be a server, computer, or mobile terminal, such as a mobile phone or tablet computer. Figure 1 As shown, the training method for this medical question-answering model specifically includes the following steps:

[0032] S110. Obtain the initial medical question-answering model to be trained.

[0033] This invention is applicable to the training of initial medical question-answering models in medical scenarios. The initial medical question-answering model is a model with basic medical knowledge question-answering capabilities. Initial medical question-answering models include, but are not limited to, large language models; for example, an initial medical question-answering model could be an Llama model. The initial medical question-answering model to be trained can be obtained from an initial medical question-answering model database. Users can search the database according to their needs to obtain the initial medical question-answering model to be trained. The database can store multiple different initial medical question-answering models. The initial medical question-answering model to be trained can also be constructed by the user. Obtaining the initial medical question-answering model to be trained provides a model foundation for subsequent analysis and processing.

[0034] The learning dimensions of the model differ at different training stages. In this invention, the first stage's learning dimension is the medical entity recognition dimension from a medical theory perspective; the second stage's learning dimension is the case reasoning dimension; and the third stage's learning dimension is the medical safety and ethical constraints dimension. These three dimensions will be explained in detail below.

[0035] To prevent data leakage during the training of medical question-answering models, an isolated environment can be deployed. The medical question-answering model resides in this isolated environment during multiple training phases, which can effectively reduce data leakage and improve the security of data processing.

[0036] It should be noted that different training data are used in different training stages. Before training the model, the training data format of each stage can be converted into a preset data format, which realizes the uniformity of the training data format in the same training stage, so as to ensure that subsequent tasks can be carried out efficiently and in an orderly manner.

[0037] S120. Based on medical theory question-and-answer pairs, the model parameters of the initial medical question-and-answer model are adjusted in the first stage to obtain the first medical question-and-answer model.

[0038] The medical theory question-and-answer pairs consist of medical theory questions and corresponding medical theory answers. The sources of these pairs include, but are not limited to, medical textbooks, drug instructions, and medical literature. For example, medical theory question-and-answer pairs can be obtained by parsing medical textbooks, drug instructions, and medical literature using a natural language processing model. Based on these pairs, the first stage of model parameter adjustment for the initial medical question-and-answer model can be as follows: input the medical theory questions from the pairs into the initial model for processing to obtain predicted medical theory answers. Then, determine whether the initial model has reached convergence based on the predicted and actual medical theory answers. If the initial model has not reached convergence, adjust its parameters until it does. The first medical question-and-answer model is a model with systematic medical theory question-and-answer capabilities. The first model and the initial model have the same model structure, but their model parameters differ.

[0039] Specifically, the medical theory questions in the medical theory question-and-answer pairs are input into the initial medical question-and-answer model for processing to obtain the predicted medical theory answers. Based on the medical theory answers and the predicted medical theory answers, it is determined whether the initial medical question-and-answer model has reached the convergence condition. If the initial medical question-and-answer model has not reached the convergence condition, the model parameters of the initial medical question-and-answer model are adjusted until the initial medical question-and-answer model reaches the convergence condition, thus obtaining the first medical question-and-answer model. This realizes the training of the initial medical question-and-answer model and helps to improve the accuracy of the first medical question-and-answer model in the dimension of medical entity recognition.

[0040] It should be noted that the first stage of model parameter adjustment for the initial medical question-answering model can be full parameter training of the initial medical question-answering model, that is, adjusting all the model parameters in the initial medical question-answering model.

[0041] S130. Based on the medical record sample data, the model parameters of the first medical question-answering model are adjusted in the second stage to obtain the second medical question-answering model.

[0042] The medical record sample data consists of data documenting clinical practice. This data includes historical medical record data and theoretical analysis link data. Historical medical record data records actual clinical data, while theoretical analysis link data represents the decision-making process based on the actual clinical data. For example, theoretical analysis link data might sequentially include symptoms, associated pathological mechanisms, exclusion diagnoses, final diagnosis, and treatment plan. To prevent the leakage of sensitive information, the medical record sample data can be anonymized. For example, names, ages, and unique identifiers can be removed from the medical record sample data. Based on the medical record sample data, the second stage of model parameter adjustment for the first medical question-and-answer model can be performed as follows: historical medical record data is input into the first medical question-and-answer model for processing to obtain training analysis link data. The theoretical analysis link data and training analysis link data are used to determine whether the first medical question-and-answer model has reached convergence. If the first medical question-and-answer model has not reached convergence, its model parameters are adjusted until convergence is achieved. The training analysis link data represents the decision-making process of the clinical data generated by the first medical question-and-answer model. The second medical question-answering model possesses systematic medical theoretical question-answering capabilities and case-based reasoning abilities. The second medical question-answering model shares the same structure as the first medical question-answering model, but its model parameters differ from those of the first and initial medical question-answering models.

[0043] Specifically, historical medical record data is input into the first medical question-answering model for processing to obtain training analysis link data. Based on the theoretical analysis link data and the training analysis link data, it is determined whether the first medical question-answering model has reached the convergence condition. If the first medical question-answering model has not reached the convergence condition, the model parameters of the first medical question-answering model are adjusted until the first medical question-answering model reaches the convergence condition, thus obtaining the second medical question-answering model. This realizes the second stage of model training and helps to improve the accuracy of the second medical question-answering model in the case reasoning dimension.

[0044] Optionally, based on medical record sample data, a second stage of model parameter adjustment is performed on the first medical question-answering model to obtain a second medical question-answering model. This includes: acquiring historical medical record data and theoretical analysis link data from the medical record sample data; inputting the historical medical record data into the first medical question-answering model to obtain the training analysis link data output by the first medical question-answering model; generating a third loss function based on the theoretical analysis link data and the training analysis link data, and performing a second stage of model parameter adjustment on the first medical question-answering model to obtain the second medical question-answering model.

[0045] The medical record sample data includes historical medical record data and theoretical analysis link data. Historical medical record data records real-world medical data. Theoretical analysis link data represents the decision-making process of real-world medical data. Training analysis link data represents the decision-making process of the medical data generated by the first medical question-answering model. Historical medical record data and theoretical analysis link data in the medical record sample data can be obtained from a medical record sample database. For example, the historical medical record data and theoretical analysis link data of a given medical record sample can be retrieved by searching the database based on its unique identification information. The medical record sample database can store historical medical record data and theoretical analysis link data corresponding to multiple medical record samples. A third loss function is used to adjust the model parameters in the second stage. The third loss function includes, but is not limited to, the cross-entropy loss function. The third loss function is the loss function of the first medical question-answering model in the second stage. The third loss function can be determined based on the theoretical analysis link data and the training analysis link data. For example, theoretical analysis link data and training analysis link data can be input into a pre-trained third loss function determination model for processing to obtain the third loss function. The third loss function determination model includes, but is not limited to, neural network models and mathematical models. The third loss function determination model is selected according to the requirements, and this invention does not impose any restrictions. The second stage of model parameter adjustment for the first medical question-answering model can be as follows: compare the third loss function with a first preset threshold. When the third loss function is greater than the first preset threshold, adjust the model parameters of the first medical question-answering model in the second stage, calculate a new third loss function, until the third loss function is less than or equal to the first preset threshold.

[0046] Specifically, based on the unique identification information corresponding to the medical record sample data, a search is performed in the medical record sample database to obtain the historical medical record data and theoretical analysis link data of the medical record sample data. The historical medical record data is input into the first medical question answering model for processing to obtain the training analysis link data output by the first medical question answering model. The theoretical analysis link data and the training analysis link data are input into the trained third loss function determination model for processing to obtain the third loss function. When the third loss function is greater than the first preset threshold, the model parameters of the first medical question answering model in the second stage are adjusted, and a new third loss function is calculated until the third loss function is less than or equal to the first preset threshold, thus obtaining the second medical question answering model. This completes the model training in the second stage, which is beneficial to improving the accuracy of the second medical question answering model in the case reasoning dimension.

[0047] Based on the above embodiments, the invention further includes: determining the corresponding medical knowledge graph based on the unique identification information of the medical record sample data, inputting the historical medical record data and the corresponding medical knowledge graph into the first medical question answering model, and obtaining the training analysis link data output by the first medical question answering model.

[0048] It should be noted that the second stage of model parameter adjustment for the first medical question-answering model can be to perform full parameter training on the first medical question-answering model, that is, to adjust all the model parameters in the first medical question-answering model.

[0049] S140. Based on the medical question-and-answer sample group, the model parameters of the second medical question-and-answer model are adjusted in the third stage to obtain the third medical question-and-answer model. The medical question-and-answer sample group includes medical questions and positive and negative answer information for medical questions.

[0050] The third medical question-answering model is a pre-trained medical question-answering model. To ensure that the output of the second medical question-answering model complies with medical safety and ethical constraints, a third stage of model training can be performed on the second medical question-answering model using a medical question-answering sample set. The sources of the medical question-answering sample set include, but are not limited to, medical ethics guidelines. The medical question-answering sample set includes medical questions and positive and negative responses to those questions. Medical questions are used to represent questions raised during the diagnostic process. Medical questions can include medical questions in different scenarios, including but not limited to rehabilitation, emergency, prevention, and medication scenarios. Positive responses represent responses that comply with medical safety and ethical constraints. Negative responses represent responses that pose medical risks or violate medical norms. The medical question-answering sample set can be determined according to medical ethics guidelines. For example, medical ethics guidelines can be input into a pre-trained medical question-answering sample set generation model for processing to obtain medical questions and positive and negative responses. The medical question-answering sample set generation model includes, but is not limited to, natural language processing models. The appropriate medical question-answering sample set generation model can be selected based on requirements; this invention does not impose any limitations. Based on the medical question-answering sample set, the third stage of model parameter adjustment for the second medical question-answering model can be carried out as follows: inputting medical questions, positive and negative responses to medical questions into the second medical question-answering model for processing to obtain the predicted response information output by the second medical question-answering model, calculating the semantic similarity between the predicted response information and the positive response information, and adjusting the model parameters of the second medical question-answering model in the third stage when the semantic similarity between the predicted response information and the positive response information is less than a preset similarity threshold, inputting new medical questions, positive and negative responses to medical questions into the second medical question-answering model for processing, calculating new semantic similarity, until the semantic similarity between the predicted response information and the positive response information is greater than or equal to the preset similarity threshold.

[0051] Specifically, medical questions, positive and negative responses to these questions are input into a second medical question-answering model for processing, resulting in predicted responses. The semantic similarity between the predicted and positive responses is calculated. When the semantic similarity is less than a preset similarity threshold, the model parameters of the second medical question-answering model are adjusted in the third stage. New medical questions, positive and negative responses are then input into the second medical question-answering model for processing, and new semantic similarities are calculated until the semantic similarity between the predicted and positive responses is greater than or equal to the preset similarity threshold, resulting in a third medical question-answering model. This multi-stage training of the medical question-answering model improves its accuracy while ensuring that the output complies with medical safety and ethical constraints.

[0052] Based on the above embodiments, a third stage of model parameter adjustment is performed on the second medical question-answering model based on the medical question-answering sample group to obtain a third medical question-answering model. This includes: inputting the medical question-answering sample group into the second medical question-answering model, obtaining the selection results of the second medical question-answering model on positive and negative response information in the medical question-answering sample group, generating a fourth loss function based on the selection results, and performing a third stage of model parameter adjustment on the second medical question-answering model based on the fourth loss function to obtain the third medical question-answering model.

[0053] The selection results are used to characterize the response preferences of the second medical question-answering model. The selection results may also include probability data corresponding to positive and negative responses. For example, the selection results could be a probability of 0.6 for positive responses and 0.4 for negative responses. The fourth loss function is used to characterize the degree of deviation between the response preferences of the second medical question-answering model and medical safety and ethical constraints. The fourth loss function includes, but is not limited to, the cross-entropy loss function. Generating the fourth loss function based on the selection results can be achieved by inputting the probability data of positive responses into the fourth loss function determination model for processing, thereby obtaining the corresponding fourth loss function. The fourth loss function determination model includes, but is not limited to, neural network models and mathematical models; the model is determined according to requirements, and this invention is not limited to this. The third stage of model parameter adjustment for the second medical question-answering model based on the fourth loss function can be achieved by adjusting the model parameters in the third stage when the fourth loss function is greater than a second preset threshold, inputting a new medical question-answering sample group into the second medical question-answering model, recalculating the fourth loss function, until the fourth loss function is less than or equal to the second preset threshold.

[0054] Specifically, the medical question-and-answer sample set is input into the second medical question-and-answer model to obtain the selection results of positive and negative response information in the sample set. The probability data of positive response information is input into the fourth loss function determination model for processing to obtain the corresponding fourth loss function. When the fourth loss function is greater than the second preset threshold, the model parameters of the second medical question-and-answer model in the third stage are adjusted. A new medical question-and-answer sample set is input into the second medical question-and-answer model, and the fourth loss function is recalculated until the fourth loss function is less than or equal to the second preset threshold, thus obtaining the third medical question-and-answer model. This achieves multi-stage training of the medical question-and-answer model, improving the accuracy of the model while ensuring that its output complies with medical safety and ethical constraints.

[0055] Based on the above embodiments, the method further includes: training the second medical question-answering model in the third stage based on a low-rank matrix: inserting the low-rank matrix into the second medical question-answering model, freezing a preset number of model parameters, and training the remaining unfrozen model parameters. For example, inserting the low-rank matrix into the second medical question-answering model yields a frozen second medical question-answering model; inputting a medical question-answering sample group into the frozen second medical question-answering model yields the selection results of the frozen second medical question-answering model for positive and negative response information in the medical question-answering sample group; generating a fourth loss function based on the selection results; and adjusting the model parameters in the third stage based on the fourth loss function to obtain a third medical question-answering model. This method is beneficial for improving the training efficiency of the second medical question-answering model while retaining the learning capabilities of the model in the medical entity recognition and case reasoning dimensions from the first two stages.

[0056] It should be noted that the third stage of model parameter adjustment for the second medical question-and-answer model can be a fine-tuning of some parameters of the second medical question-and-answer model, that is, adjusting a small number of parameters in the second medical question-and-answer model.

[0057] The technical solution of this embodiment provides a model foundation for subsequent analysis and processing by acquiring an initial medical question-and-answer model to be trained. Based on medical theoretical question-and-answer pairs, the initial medical question-and-answer model undergoes a first-stage parameter adjustment to obtain a first medical question-and-answer model, thus achieving training of the initial model and improving its accuracy in the medical entity recognition dimension. Based on medical record sample data, the first medical question-and-answer model undergoes a second-stage parameter adjustment to obtain a second medical question-and-answer model, achieving second-stage model training and improving its accuracy in the case reasoning dimension. Based on a medical question-and-answer sample set, the second medical question-and-answer model undergoes a third-stage parameter adjustment to obtain a third medical question-and-answer model. This third-stage model includes medical questions and positive and negative responses to those questions, achieving multi-stage training of the medical question-and-answer model. While ensuring that the output of the medical question-and-answer model complies with medical safety and ethical constraints, it improves the accuracy of the medical question-and-answer model.

[0058] Example 2

[0059] Figure 2 This is a flowchart of a training method for a medical question-answering model provided in Embodiment 2 of the present invention. This embodiment is a refinement of the above embodiments. Based on the foregoing embodiments, it provides a detailed explanation of the first-stage model parameter adjustment of the initial medical question-answering model based on medical theory question-answer pairs to obtain a first medical question-answering model. For specific implementation details, please refer to the technical solution of this embodiment. Technical terms that are the same as or corresponding to those in the above embodiments will not be repeated here. Figure 2 As shown, the training method for this medical question-answering model specifically includes the following steps:

[0060] S210. Obtain the initial medical question-answering model to be trained.

[0061] S220. Obtain the actual entities and actual entity attributes in the medical theory question-and-answer pair. The actual entity attributes include entity location and entity type. Input the medical theory question-and-answer pair into the initial medical question-and-answer model to obtain the predicted entities and predicted entity attributes in the process of processing the medical theory question-and-answer pair by the initial medical question-and-answer model. Based on the actual entities and actual entity attributes, as well as the predicted entities and predicted entity attributes, generate a first loss function to perform the first stage of model parameter adjustment on the initial medical question-and-answer model to obtain the first medical question-and-answer model.

[0062] In this context, "actual entities" refers to the object entities within the medical theory question-and-answer pair, specifically those related to medical theory questions. Actual entity attributes characterize the features of these object entities. Entity attributes include entity location and entity type. Entity location characterizes the position of the object entity within the medical theory question. Different actual entities correspond to different entity locations. Entity type characterizes the semantic classification of the object entity within the medical theory question. Entity types include medical entities and non-medical entities. Medical entities include professional concepts in the medical field, including but not limited to drug names, symptoms, and physiological indicators. Non-medical entities are not medical terms and include, but are not limited to, time, people, and locations. Actual entities and their attributes can be determined based on the medical theory question-and-answer pair. For example, the medical theory question-and-answer pair can be input into a trained word segmentation model for word segmentation processing to obtain the actual entities and their attributes within the pair. Word segmentation models include, but are not limited to, natural language processing models.

[0063] Optionally, the entity type can be determined by: matching the actual entity in a medical entity set to obtain the matching result, and then determining the entity type based on the matching result.

[0064] The medical entity set comprises entity data consisting of professional concepts in the medical field. This set includes, but is not limited to, drug names, symptoms, and physiological indicators. The medical entity set can be pre-constructed. The matching result indicates whether an actual entity belongs to the medical entity set. The matching result can be a successful match or a failed match. The matching result can be determined based on the actual entity and the medical entity set. For example, the matching result can be obtained by matching the actual entity within the medical entity set based on its unique identifier. Determining the entity type based on the matching result can be as follows: when the matching result is successful, the corresponding entity type in the medical entity set is used as the entity type of the actual entity; when the matching result is a failed match, the entity type of the actual entity is classified as a non-medical entity.

[0065] Specifically, the system matches the actual entity in the medical entity set based on the unique identifier of the actual entity to obtain the matching result. When the matching result is successful, the entity type corresponding to the medical entity set is taken as the entity type of the actual entity. When the matching result is unsuccessful, the entity type of the actual entity is a non-medical entity. This achieves accurate determination of the entity type and provides accurate data support for subsequent non-analysis and processing.

[0066] The predicted entity is the object entity obtained by the initial medical question-answering model during the processing of medical theory question-answer pairs. Predicted entity attributes characterize the features of the object entities obtained by the initial medical question-answering model during the processing of medical theory question-answer pairs. Predicted entity attributes include predicted entity location and predicted entity type. Predicted entities and predicted entity attributes can be obtained from medical theory question-answer pairs. For example, medical theory question-answer pairs can be input into the initial medical question-answering model for processing to obtain predicted entities and predicted entity attributes from the initial medical question-answering model's processing of medical theory question-answer pairs. The first loss function characterizes the degree of offset between the actual entity and the predicted entity, and between the actual entity attributes and the predicted entity attributes. The first loss function includes, but is not limited to, the cross-entropy loss function. The first loss function can be determined based on the actual entity and actual entity attributes, as well as the predicted entity and predicted entity attributes. For example, the consistency between entity location in actual entity attributes and predicted entity location in predicted entity attributes, as well as the consistency between entity type in actual entity attributes and predicted entity type in predicted entity attributes, is calculated. When both of these consistencies are greater than or equal to the corresponding preset consistency thresholds, the actual entity and actual entity attributes, as well as the predicted entity and predicted entity attributes, are input into a pre-trained first loss function determination model for processing to obtain the first loss function. The first loss function determination model includes, but is not limited to, neural network models and mathematical models. The generation of the first loss function performs the first stage of model parameter adjustment on the initial medical question-answering model. This can be achieved by: when the first loss function is greater than a third preset threshold, adjusting the model parameters of the initial medical question-answering model in the first stage, inputting new medical theory question-answer pairs, and recalculating a new first loss function until the first loss function is less than or equal to the third preset threshold.

[0067] Specifically, the medical question-and-answer pairs are input into a trained word segmentation model for word segmentation processing to obtain the actual entities and actual entity attributes in the medical question-and-answer pairs. The consistency between the entity position in the actual entity attributes and the predicted entity position in the predicted entity attributes, as well as the consistency between the entity type in the actual entity attributes and the predicted entity type in the predicted entity attributes, are calculated. When both of these consistency values ​​are greater than or equal to the corresponding preset consistency thresholds, the actual entities and actual entity attributes, as well as the predicted entities and predicted entity attributes, are input into a trained first loss function determination model for processing to obtain the first loss function. When the first loss function is greater than a third preset threshold, the model parameters of the initial medical question-and-answer model in the first stage are adjusted, new medical question-and-answer pairs are input, and a new first loss function is recalculated until the first loss function is less than or equal to the third preset threshold, thus obtaining the first medical question-and-answer model. This completes the training of the initial medical question-and-answer model and helps improve the accuracy of the first medical question-and-answer model in the dimension of medical entity recognition.

[0068] Optionally, the training method of the medical question-answering model further includes: identifying the weight value corresponding to each actual entity based on the entity weight configuration information with preset values; generating a second loss function based on the actual entity and actual entity attributes, the predicted entity and predicted entity attributes, and the weight value corresponding to the actual entity to perform the first stage of model parameter adjustment on the initial medical question-answering model, thereby obtaining the first medical question-answering model.

[0069] The entity weight configuration information is used to characterize the mapping relationship of the importance of actual entities. This information includes, but is not limited to, a mapping table between actual entities and their weight values. Different actual entities correspond to different weight values. Medical entities have higher weight values ​​than non-medical entities. Weight values ​​characterize the importance of an actual entity. The weight value corresponding to an actual entity can be determined based on preset entity weight configuration information. For example, the weight value corresponding to an actual entity can be obtained by matching its unique identifier in the mapping table between actual entities and their weight values. The second loss function is used to characterize the deviation of the initial medical question-answering model in the medical entity recognition dimension. This second loss function includes, but is not limited to, the cross-entropy loss function and its variants. The second loss function can be determined based on the actual entity and its attributes, the predicted entity and its attributes, and the weight value corresponding to the actual entity. For example, the consistency between the entity location in the actual entity attribute and the predicted entity location in the predicted entity attribute, as well as the consistency between the entity type in the actual entity attribute and the predicted entity type in the predicted entity attribute, is calculated. When the above two consistency values ​​are greater than or equal to the corresponding preset consistency thresholds, the medical entity in the actual entity attribute is determined. The unique identifier information of the medical entity is matched in the mapping relationship table between the actual entity and the weight value to obtain the weight value corresponding to the medical entity. The weight value corresponding to the medical entity is input into the trained second loss function determination model for processing to obtain the second loss function. The second loss function determination model includes, but is not limited to, neural network models and mathematical models.

[0070] Specifically, the consistency between the entity location in the actual entity attributes and the predicted entity location in the predicted entity attributes, as well as the consistency between the entity type in the actual entity attributes and the predicted entity type in the predicted entity attributes, is calculated. When the above two consistency values ​​are greater than or equal to the corresponding preset consistency thresholds, the medical entity in the actual entity attributes is determined. Based on the unique identifier information of the medical entity, it is matched in the mapping relationship table between the actual entity and the weight value to obtain the weight value corresponding to the medical entity. The weight value corresponding to the medical entity is input into the trained second loss function to determine the model for processing, and the second loss function is obtained. When the second loss function is greater than the fourth preset threshold, the model parameters of the initial medical question answering model in the first stage are adjusted, and the initial medical question answering model is trained again until the obtained second loss function is less than or equal to the fourth preset threshold, thus obtaining the first medical question answering model. This realizes the training of the initial medical question answering model, which is conducive to improving the accuracy of the initial medical question answering model in the medical entity recognition dimension in medical theory.

[0071] S230. Based on the medical record sample data, the model parameters of the first medical question-answering model are adjusted in the second stage to obtain the second medical question-answering model.

[0072] S240. Based on the medical question-and-answer sample group, the model parameters of the second medical question-and-answer model are adjusted in the third stage to obtain the third medical question-and-answer model. The medical question-and-answer sample group includes medical questions and positive and negative answer information for medical questions.

[0073] The technical solution of this embodiment provides a model foundation for subsequent analysis and processing by acquiring an initial medical question-answering model to be trained; it acquires the actual entities and actual entity attributes in the medical theoretical question-answering pairs, with entity attributes including entity location and entity type; it inputs the medical theoretical question-answering pairs into the initial medical question-answering model to obtain the predicted entities and predicted entity attributes in the processing of the medical theoretical question-answering pairs by the initial medical question-answering model; based on the actual entities and actual entity attributes, as well as the predicted entities and predicted entity attributes, it generates a first loss function to perform the first stage of model parameter adjustment on the initial medical question-answering model, thus obtaining the first medical question-answering model. This achieves the training of the initial medical question-answering model and is beneficial to improving the first medical question-answering model. The accuracy of the model in the medical entity recognition dimension was improved. Based on medical record sample data, the second stage of model parameter adjustment was performed on the first medical question-answering model to obtain the second medical question-answering model, realizing the second stage of model training, which is conducive to improving the accuracy of the second medical question-answering model in the case reasoning dimension. Based on the medical question-answering sample group, the second medical question-answering model was adjusted in the third stage to obtain the third medical question-answering model. The medical question-answering sample group includes medical questions and positive and negative answer information for medical questions, realizing multi-stage training of the medical question-answering model. While ensuring that the output of the medical question-answering model complies with medical safety and ethical constraints, the accuracy of the medical question-answering model is improved.

[0074] Example 3

[0075] Figure 3 This is a schematic diagram of the structure of a training device for a medical question-and-answer model provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes an initial medical question-and-answer model acquisition module 310, a first-stage training module 320, a second-stage training module 330, and a third-stage training module 340.

[0076] The system includes: an initial medical question-and-answer model acquisition module 310, used to acquire an initial medical question-and-answer model to be trained; a first-stage training module 320, used to adjust the model parameters of the initial medical question-and-answer model in the first stage based on medical theory question-and-answer pairs, to obtain a first medical question-and-answer model; a second-stage training module 330, used to adjust the model parameters of the first medical question-and-answer model in the second stage based on medical record sample data, to obtain a second medical question-and-answer model; and a third-stage training module 340, used to adjust the model parameters of the second medical question-and-answer model in the third stage based on a medical question-and-answer sample group, to obtain a third medical question-and-answer model. The medical question-and-answer sample group includes medical questions and positive and negative answer information for medical questions.

[0077] The technical solution of this embodiment obtains an initial medical question-and-answer model for training through an initial medical question-and-answer model acquisition module, providing a model foundation for subsequent analysis and processing. Through a first-stage training module, based on medical theoretical question-and-answer pairs, the initial medical question-and-answer model undergoes first-stage model parameter adjustment to obtain a first medical question-and-answer model, achieving first-stage training and improving the accuracy of the first medical question-and-answer model in the medical entity recognition dimension. Through a second-stage training module, based on medical record sample data, the first medical question-and-answer model undergoes second-stage model parameter adjustment to obtain a second medical question-and-answer model, achieving second-stage model training and improving the accuracy of the second medical question-and-answer model in the case reasoning dimension. Through a third-stage training module, based on a medical question-and-answer sample set, the second medical question-and-answer model undergoes third-stage model parameter adjustment to obtain a third medical question-and-answer model. The medical question-and-answer sample set includes medical questions and positive and negative responses to those questions, achieving multi-stage training of the medical question-and-answer model. This improves the accuracy of the medical question-and-answer model while ensuring that its output complies with medical safety and ethical constraints.

[0078] Based on the above embodiments, optionally, the first-stage training module 320 is further configured to: obtain the actual entities and actual entity attributes in the medical theory question-and-answer pair, wherein the entity attributes include entity location and entity type; input the medical theory question-and-answer pair into the initial medical question-and-answer model to obtain the predicted entities and predicted entity attributes in the process of the initial medical question-and-answer model processing the medical theory question-and-answer pair; and generate a first loss function based on the actual entities and actual entity attributes as well as the predicted entities and predicted entity attributes to perform a first-stage model parameter adjustment on the initial medical question-and-answer model to obtain the first medical question-and-answer model.

[0079] Optionally, the first-stage training module 320 is also used to: identify the weight value corresponding to each actual entity based on the entity weight configuration information with preset values; and generate a second loss function based on the actual entity and actual entity attributes, the predicted entity and predicted entity attributes, and the weight value corresponding to the actual entity to perform the first-stage model parameter adjustment of the initial medical question answering model, thereby obtaining the first medical question answering model.

[0080] Optionally, entity types include medical entities and non-medical entities.

[0081] Optionally, the weight value of medical entities is greater than the weight value of non-medical entities.

[0082] Optionally, the first-stage training module 320 is also used to: match the actual entities in the medical entity set, obtain the matching results, and determine the entity type based on the matching results.

[0083] Optionally, the second-stage training module 330 is also used to: acquire historical medical record data and theoretical analysis link data from the medical record sample data; input the historical medical record data into the first medical question answering model to obtain the training analysis link data output by the first medical question answering model; generate a third loss function based on the theoretical analysis link data and the training analysis link data, and perform second-stage model parameter adjustment on the first medical question answering model to obtain the second medical question answering model.

[0084] Optionally, the third-stage training module 340 is also used to: input the medical question-and-answer sample group into the second medical question-and-answer model, obtain the selection results of the second medical question-and-answer model on the positive and negative response information in the medical question-and-answer sample group, generate a fourth loss function based on the selection results, and adjust the model parameters of the second medical question-and-answer model in the third stage based on the fourth loss function to obtain the third medical question-and-answer model.

[0085] The training device for a medical question-answering model provided in this embodiment of the invention can execute the training method for a medical question-answering model provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0086] Example 4

[0087] Figure 4 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 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 (such as helmets, glasses, watches, etc.), 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 invention described and / or claimed herein.

[0088] like Figure 4As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0089] Multiple components in electronic device 10 are connected to input / output (I / O) interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of monitors, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0090] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 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 processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a training method for a medical question-answering model.

[0091] In some embodiments, a training method for a medical question-answering model can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via read-only memory (ROM) 12 and / or communication unit 19. When the computer program is loaded into random access memory (RAM) 13 and executed by processor 11, one or more steps of the training method for a medical question-answering model described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute a training method for a medical question-answering model by any other suitable means (e.g., by means of firmware).

[0092] 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), system-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 transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0093] Computer programs for implementing a training method of a medical question-answering model according to the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can 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.

[0094] Example 5

[0095] Embodiment 5 of the present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute a training method for a medical question-answering model. The training method for the medical question-answering model includes:

[0096] An initial medical question-answering model to be trained is obtained; based on medical theoretical question-answer pairs, the model parameters of the initial medical question-answering model are adjusted in the first stage to obtain the first medical question-answering model; based on medical record sample data, the model parameters of the first medical question-answering model are adjusted in the second stage to obtain the second medical question-answering model; based on a medical question-answering sample group, the model parameters of the second medical question-answering model are adjusted in the third stage to obtain the third medical question-answering model. The medical question-answering sample group includes medical questions and positive and negative answer information for medical questions.

[0097] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. 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 fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0098] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. 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).

[0099] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users 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., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0100] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0101] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from ROM 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of the embodiments of the present invention.

[0102] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0103] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. 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 invention should be included within the scope of protection of this invention.

Claims

1. A training method for a medical question-answering model, characterized in that, include: Obtain the initial medical question-answering model to be trained; Based on medical theoretical question-and-answer pairs, the initial medical question-and-answer model is adjusted in the first stage to obtain the first medical question-and-answer model; Based on medical record sample data, the model parameters of the first medical question-answering model are adjusted in the second stage to obtain the second medical question-answering model; Based on the medical question-and-answer sample set, the model parameters of the second medical question-and-answer model are adjusted in the third stage to obtain the third medical question-and-answer model. The medical question-and-answer sample set includes medical questions and positive and negative response information for the medical questions.

2. The method according to claim 1, characterized in that, The first stage of model parameter adjustment for the initial medical question-and-answer model, based on medical theory question-and-answer pairs, yields a first medical question-and-answer model, including: Obtain the actual entities and actual entity attributes in the medical theory question-and-answer pair, wherein the actual entity attributes include entity location and entity type; The medical theory question-and-answer pair is input into the initial medical question-and-answer model to obtain the predicted entities and predicted entity attributes in the process of processing the medical theory question-and-answer pair by the initial medical question-and-answer model. Based on the actual entity and its attributes, as well as the predicted entity and its attributes, a first loss function is generated to adjust the model parameters of the initial medical question-answering model in the first stage, thereby obtaining the first medical question-answering model.

3. The method according to claim 2, characterized in that, The method further includes: Based on the entity weight configuration information with preset values, identify the weight value corresponding to each actual entity; Based on the actual entity and its attributes, the predicted entity and its attributes, and the weight value corresponding to the actual entity, a second loss function is generated to adjust the model parameters of the initial medical question-answering model in the first stage, thereby obtaining the first medical question-answering model.

4. The method according to claim 3, characterized in that, The entity types include medical entities and non-medical entities; the weight value of medical entities is greater than the weight value of non-medical entities; The method for determining the entity type includes: The actual entity is matched against a set of medical entities to obtain a matching result, and the entity type is determined based on the matching result.

5. The method according to claim 1, characterized in that, The second stage of adjusting the model parameters of the first medical question-answering model based on medical record sample data to obtain a second medical question-answering model includes: Acquire historical medical record data and theoretical analysis link data from the medical record sample data; The historical medical record data is input into the first medical question-answering model to obtain the training and analysis link data output by the first medical question-answering model; Based on the theoretical analysis link data and the training analysis link data, a third loss function is generated, and the model parameters of the first medical question answering model are adjusted in the second stage to obtain the second medical question answering model.

6. The method according to claim 1, characterized in that, The process of adjusting the model parameters of the second medical question-answering model in the third stage, based on the medical question-answering sample group, to obtain the third medical question-answering model includes: The medical question-and-answer sample group is input into the second medical question-and-answer model to obtain the selection results of the positive and negative response information in the medical question-and-answer sample group by the second medical question-and-answer model. A fourth loss function is generated based on the selection results. The model parameters of the second medical question-and-answer model are adjusted in the third stage based on the fourth loss function to obtain the third medical question-and-answer model.

7. A training device for a medical question-and-answer model, characterized in that, include: The initial medical question-answering model acquisition module is used to acquire the initial medical question-answering model to be trained. The first-stage training module is used to adjust the model parameters of the initial medical question-and-answer model based on medical theory question-and-answer pairs to obtain the first medical question-and-answer model. The second-stage training module is used to adjust the model parameters of the first medical question-answering model in the second stage based on medical record sample data, so as to obtain the second medical question-answering model. The third-stage training module is used to adjust the model parameters of the second medical question-answering model in the third stage based on the medical question-answering sample group to obtain the third medical question-answering model. The medical question-answering sample group includes medical questions and positive and negative response information for the medical questions.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method of the medical question-answering model according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the training method of the medical question-answering model according to any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the training method for the medical question-answering model according to any one of claims 1-6.

Images