Method for training large language model for health notification

Abstract

The embodiment of the invention discloses a method for training a large language model for health notification. The scheme can comprise the following steps: acquiring initial training data; the initial training data comprises sample data and label data for the sample data; the sample data comprises a user input sample and a first health notification text; according to the incidence relation data, knowledge extension information for the user input sample is generated; the association relationship data is used for reflecting an association relationship among a plurality of pieces of data extracted from a preset text and basic knowledge of a field to which the preset text belongs; the preset text at least comprises a first health notification text; generating training sample data based on the initial training data and the knowledge extension information; and performing supervised fine tuning training on the pre-trained large model by using the training sample data to obtain a large language model for health notification.

Description

Technical Field

[0001] The present application relates to the field of computer data processing technology, and in particular to a method for training a large language model for health notification, an analysis method for health notification, an apparatus for training a large language model for health notification, an analysis apparatus for health notification, a computer device, and a computer-readable storage medium. Background Art

[0002] When purchasing insurance, users need to review the health disclosure requirements issued by the insurance company. Users can determine whether they meet the health disclosure requirements based on their own health status and, therefore, whether they can be insured. However, the health disclosure requirements issued by insurance companies are often technical and redundant, and the health disclosures cover a large number of diseases. Users are often unable to accurately determine whether they meet the health disclosure requirements, which makes it difficult for users to make accurate judgments and may result in the inability to obtain health insurance.

[0003] Therefore, how to train a large language model for health information disclosure that can provide users with accurate insurance advice based on their inquiries to protect their rights and interests is a technical problem that needs to be solved urgently. Summary of the Invention

[0004] The embodiments of this specification provide a method for training a large language model for health information, so that the trained large language model for health information can provide users with accurate insurance recommendations based on user consultation questions to protect user rights and interests.

[0005] To solve the above technical problems, an embodiment of this specification provides a method for training a large language model for health notification, including: obtaining initial training data; the initial training data includes sample data and label data for the sample data; the sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample belongs to the description scope of the first health notification text; the second label data indicates that the user input sample does not belong to the description scope of the first health notification text; based on the association relationship data, generating knowledge extension information for the user input sample; the association relationship data is used to reflect the association relationship between multiple data extracted from a preset text and the basic knowledge of the field to which the preset text belongs; the preset text at least includes the first health notification text; based on the initial training data and the knowledge extension information, generating training sample data; using the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain the large language model for health notification.

[0006] The embodiments of this specification also provide a method for analyzing health notifications, including: obtaining user input text; Obtain a target health notification text corresponding to the user input text in a preset text; generate target knowledge extension information for the user input text based on the association relationship data; the association relationship data is data used to reflect the association relationship between multiple data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; input the target health notification text, the user input text and the target knowledge extension information into a large language model for health notification to obtain target analysis result information; the target analysis result information includes first target result information or second target result information; the first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

[0007] An embodiment of this specification also provides a device for training a large language model for health notification, including: an initial training data acquisition module, used to acquire initial training data; the initial training data includes sample data and label data for the sample data; the sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample belongs to the description scope of the first health notification text; the second label data indicates that the user input sample does not belong to the description scope of the first health notification text; a knowledge extension information generation module, used to generate knowledge extension information for the user input sample based on association relationship data; the association relationship data is used to reflect the association relationship between multiple data extracted from a preset text and the basic knowledge of the field to which the preset text belongs; the preset text at least includes the first health notification text; a training sample data generation module, used to generate training sample data based on the initial training data and the knowledge extension information; a fine-tuning training module, used to use the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain the large language model for health notification.

[0008] An embodiment of this specification also provides an analysis device for health notification, including: a user input text acquisition module for acquiring user input text; a health notification text acquisition module for acquiring a target health notification text corresponding to the user input text in a preset text; a target knowledge extension information generation module for generating target knowledge extension information for the user input text based on association relationship data; the association relationship data is data used to reflect the association relationship between multiple data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; an analysis module for inputting the target health notification text, the user input text and the target knowledge extension information into a large language model for health notification to obtain target analysis result information; the target analysis result information includes first target result information or second target result information; the first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

[0009] An embodiment of this specification also provides a computer device, comprising: 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, and the instructions are executed by the at least one processor to enable the at least one processor to implement a method for training a large language model for health notification or an analysis method for health notification.

[0010] An embodiment of this specification also provides a computer-readable storage medium having a computer program or instructions stored thereon, which can be executed by a processor to implement a method for training a large language model for health notification or steps of an analysis method for health notification.

[0011] At least one embodiment of this specification can achieve the following beneficial effects: by obtaining initial training data including user input samples, a first health notification text, and label data, and generating knowledge extension information for the user input sample based on association relationship data that reflects the association relationship between multiple data extracted from a preset text and basic knowledge in the field to which the preset text belongs, training sample data can be generated based on the initial training data and the knowledge extension information, and then the pre-trained large model can be supervised and fine-tuned using the training sample data to obtain a large language model for health notification. Therefore, without the need for the user to have professional knowledge, the trained large language model for health notification can provide the user with accurate insurance advice based on the information input by the user to protect the user's rights and interests. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some of the embodiments described in this application. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0013] Figure 1 This is a schematic diagram of the overall solution architecture of a method for training a large language model for health notification provided in an embodiment of this specification; Figure 2 This is a flowchart of a method for training a large language model for health notification provided in an embodiment of this specification; Figure 3 This is a schematic diagram of a processing flow of a sample health notification text provided in an embodiment of this specification; Figure 4 This is a flow chart of determining association relationship data of preset texts provided by an embodiment of this specification; Figure 5 This is a schematic diagram of association relationship data provided in an embodiment of this specification; Figure 6 This is a flow chart of processing user input text using a large language model for health notification provided by an embodiment of this specification; Figure 7 This is a flowchart of an analysis method for health notification provided in an embodiment of this specification; Figure 8 This specification provides the corresponding embodiment Figure 2 A schematic diagram of a device for training a large language model for health information; Figure 9 This specification provides the corresponding embodiment Figure 7 A schematic structural diagram of an analysis device for health notification; Figure 10 This is a structural diagram of a computer device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0014] To make the purpose, technical solutions, and advantages of one or more embodiments of this specification more clear, the technical solutions of one or more embodiments of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of one or more embodiments of this specification.

[0015] The terms used in one or more embodiments of the present application are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of the present application. The singular forms "a", "the" and "the" used in one or more embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of the present application refers to and includes any or all possible combinations of one or more associated listed items.

[0016] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present application, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of one or more embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0017] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0018] In order to facilitate understanding of the various embodiments in this specification, some terms are explained below.

[0019] Query: refers to the natural language query text entered by the user, which contains consultation content related to health conditions such as diseases, symptoms, examinations, and treatments.

[0020] Supervised Fine-Tuning (SFT) is a training method in natural language processing and machine learning that uses labeled data to further train a pre-trained model, making it more adaptable to data from a specific task or domain. Specifically, the parameters of the pre-trained model are used as the initial parameters, and further training is performed using labeled data. Through optimization methods such as backpropagation and gradient descent, the model parameters are adjusted to minimize the loss function for the specific task. This allows the large model to gradually adapt to the data distribution of the specific task, thereby improving its performance on that task.

[0021] Few shots: These are examples of tasks that the large model is provided with. For example, N-shot training provides N examples. Zero-shot training allows the large model to complete the task directly, without providing any examples.

[0022] Retrieval-Augmented Generation (RAG) is a natural language processing method that combines retrieval and generation techniques. This method enhances the performance and accuracy of generative models by retrieving relevant information from external knowledge sources. RAG utilizes a retrieval system to find relevant snippets from large amounts of text data. These snippets are then fed into the model as additional text, enabling it to generate more accurate and informative text.

[0023] Chain of Thought (CoT) breaks down complex text into a series of step-by-step reasoning steps to better understand and solve problems. This can be used to guide large models in generating more coherent and logical outputs. In tasks requiring multi-step reasoning, CoT improves the accuracy and interpretability of large models' responses by breaking down complex questions into step-by-step reasoning steps.

[0024] The In-Context Learning (ICL) paradigm is a machine learning paradigm that allows models to learn and reason using contextual information without the need for training or adjusting internal parameters. By learning context, models can understand the structure and rules of a task, enabling them to make reasonable predictions even on unseen data. Similar to how humans learn new skills by observing examples, ICL relies on contextual examples and cues to guide the model's learning process.

[0025] The technical solutions provided by the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0026] Before purchasing insurance, users need to check the health disclosure requirements corresponding to the insurance they are about to purchase. However, the content of the health disclosure requirements is highly professional and redundant, which makes it difficult for users to judge whether they meet the health disclosure requirements. As a result, users are unable to determine whether they can purchase insurance and are unable to protect their own rights and interests.

[0027] Related technologies include an underwriting assistant developed using generative AI technology and open to underwriters. This assistant addresses the health disclosure requirements for both life and health insurance. Based on the underwriter's questions and the corresponding health disclosure requirements, the assistant generates a summary for the underwriter. The summary provided by the assistant is professional, and underwriters possess certain insurance or medical expertise, understanding how to correctly ask questions and the model's output based on those questions. Because policyholders lack relevant expertise, they cannot directly ask questions through the assistant or determine whether to purchase insurance based on the assistant's output. This decision must be made by the underwriter. Furthermore, if the underwriter's expertise is limited, inaccurate underwriting conclusions can easily result, potentially compromising the rights and interests of policyholders. Therefore, training a large language model for health disclosure that can provide accurate pre-underwriting conclusions even for policyholders without specialized knowledge is a pressing technical challenge.

[0028] In order to solve the defects in the prior art, this solution provides the following embodiments: Figure 1 This is a schematic diagram of the overall solution architecture of a method for training a large language model for health notification provided in an embodiment of this specification.

[0029] like Figure 1 As shown, the solution may include initial training data 1, server 2, and a large language model 3 for health notification. The initial training data 1 may contain sample data and label data for the sample data, and the sample data may include user input samples and the first health notification text. Server 2 may contain a database storing association data. Server 2 may obtain the initial training data 1 and generate knowledge expansion information for the user input sample based on the association data. Server 2 may process the initial training data and the knowledge expansion information to generate training sample data. Server 2 may use the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain a large language model 3 for health notification. The large language model 3 for health notification can give accurate insurance recommendations based on the questions input by the user, thereby protecting the rights and interests of the user.

[0030] Next, a method for training a large language model for health notification provided in an embodiment of the specification will be specifically described with reference to the accompanying drawings: Figure 2This is a flow chart of a method for training a large language model for health notification provided in an embodiment of this specification. From a program perspective, the execution subject of the process can be a program or application client installed on an application server. Figure 2 As shown, the method may include the following steps.

[0031] Step 202: Obtain initial training data.

[0032] The initial training data includes sample data and label data for the sample data; the sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample belongs to the description range described by the first health notification text; the second label data indicates that the user input sample does not belong to the description range described by the first health notification text.

[0033] In the embodiments of this specification, the user input sample can be based on the questions that the user has asked historically based on the first health notification text; or, it can be simulated and generated based on the first health notification text using a generative model; or, it can be manually written for the first health notification text, etc. There is no specific limitation on the user input sample here. The first health notification text can be the health notification requirements corresponding to the insurance that the user wants to purchase corresponding to the user input sample. The first health notification text can include text for describing the conditions under which insurance cannot be purchased, and can also include text for describing that insurance can be purchased even if certain conditions are met.

[0034] In an embodiment of the present specification, the label data may be conclusion information indicating whether the user is advised to take out insurance. The description range may indicate the uninsurable range described in the first health notification text. If the user input sample falls within the description range of the first health notification text, it may indicate that the user input sample meets certain diseases or symptoms or other situations in the first health notification text. It can be determined that the user input sample falls within the uninsurable range, and then the first label data can be determined to be a label that does not recommend insurance. If the user input sample does not fall within the description range of the first health notification text, it may indicate that the user input sample does not meet certain diseases or symptoms or other situations in the first health notification text; or it may indicate that the user input sample meets the situation in the first health notification text where insurance can be taken out normally, for example, the user has cold symptoms but can still take out insurance normally; it can be determined that the user input sample does not fall within the uninsurable range, but falls within the insurable range, and then the second label can be determined to be a label that recommends insurance.

[0035] Step 204: Generate knowledge expansion information for the user input sample based on the association relationship data.

[0036] Among them, the association relationship data is used to reflect the association relationship between multiple data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; the preset text at least includes the first health notification text.

[0037] In the embodiments of this specification, the association relationship data can be structured data used to describe the association relationship between multiple data in the preset text, obtained by pre-processing the preset text using a large model; or it can be structured data used to describe the association relationship between multiple data in the basic knowledge obtained by pre-processing the basic knowledge of the field to which the preset text belongs using a large model; or it can be structured basic knowledge obtained directly from the data provider. The structure of the association relationship data can be a knowledge graph structure, or it can be an array, a linked list, a suffix tree, a suffix array, etc. The structure of the association relationship can be a structure that meets the model recognition requirements. The structure of the association relationship is not specifically limited here.

[0038] In an embodiment of the present specification, the knowledge extension information may be obtained by expanding the user input sample by using a large model to obtain data related to the user input sample from the association data based on the user input sample. In an embodiment of the present specification, the preset text may be a plurality of health notification texts; the user input sample may be the health status information input by the user for the first health notification text. The preset text may contain various health notification texts corresponding to various insurances that the platform, application, or insurance company can provide to the user. Health status information may be information that the user uses to describe his or her own physical condition, such as "I have a cold", "My blood pressure was a little high when I checked it last month", and other text information indicating the user's physical condition. If the preset text is a health notification text, the basic knowledge in the field to which the preset text belongs may be medical knowledge.

[0039] Step 206: Generate training sample data based on the initial training data and the knowledge expansion information.

[0040] In the embodiment of this specification, the server may concatenate the initial training data with the knowledge expansion information to obtain training sample data. Specifically, the initial training data may be concatenated before the knowledge expansion information to obtain training sample data, or the initial training data may be concatenated after the knowledge expansion information to obtain training sample data.

[0041] Step 208: Use the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain the large language model for health notification.

[0042] In the embodiments of this specification, the pre-trained large model can be obtained by pre-training the large model using a large number of samples. The large language model for health notification can process the text information input by the user and output insurance recommendations for the user, such as whether insurance is not recommended or recommended. Using labeled training sample data to perform supervised fine-tuning training on the pre-trained large model can improve the accuracy of the trained large language model for health notification in processing user input text.

[0043] In actual applications, the server can use a preset number of training sample data to perform preset rounds of supervised fine-tuning training on the pre-trained large model. The preset number can be 1000, 2000, etc.; the preset rounds can be 1, 5, 13, etc. In actual applications, after completing the supervised fine-tuning of the pre-trained large model, the fine-tuned large model can also be verified using the validation set. If the verification result does not meet the preset requirements, such as the loss value calculated based on the loss function is greater than or equal to the preset value, the training sample data can be continued to perform supervised fine-tuning training on the fine-tuned large model whose verification result does not meet the preset requirements, until the verification result of the fine-tuned large model meets the preset requirements, then the supervised fine-tuning training is stopped to obtain a large language model for health notification.

[0044] It should be understood that the order of some steps in the methods described in one or more embodiments of this specification can be interchanged according to actual needs, or some steps can be omitted or deleted.

[0045] Figure 2 The method obtains initial training data including user input samples, the first health notification text, and label data, and generates knowledge extension information for the user input sample based on association relationship data reflecting the association relationship between multiple data extracted from the preset text and the basic knowledge of the field to which the preset text belongs. Training sample data can be generated based on the initial training data and the knowledge extension information, and then the pre-trained large model is fine-tuned using the training sample data to obtain a large language model for health notification. Therefore, without the user having professional knowledge, the trained large language model for health notification can provide accurate insurance advice to the user based on the information input by the user to protect the user's rights and interests.

[0046] based on Figure 2The embodiment of this specification also provides some specific implementation methods of this method, which are described below. Before using the association relationship data, the preset text can be used to pre-construct the association relationship data to improve the efficiency of generating knowledge extension information. Optionally, before generating the knowledge extension information for the user input sample based on the association relationship data in the embodiment of this specification, it can also include: obtaining the preset text; obtaining target annotation data whose semantic similarity with the preset text meets the preset conditions; the target annotation data includes a second health notification text, a first sample association relationship data corresponding to the second health notification text, and a first reasoning data for reflecting the reasoning process of determining the first sample association relationship data based on the second health notification text; the preset text and the target annotation data are input into the first large model to obtain the association relationship data corresponding to the preset text output by the first large model.

[0047] In the embodiments of this specification, the preset text may be obtained from a data provider. The data provider may include a business entity providing insurance services. The target annotation data may be one or more pieces of data. The target annotation data may be the one piece of target annotation data obtained from a annotation database storing multiple annotation data using RAG technology, having the highest semantic similarity to the preset text; or the target annotation data may be one or more pieces of target annotation data obtained from a annotation database storing multiple annotation data using RAG technology, having a semantic similarity to the preset text greater than or equal to a preset similarity. The multiple annotation data in the annotation database may be processed based on a portion of the health notification text in the preset text. The multiple annotation data in the annotation database may serve as minority examples to guide health notification text in the preset text that is not in the annotation database. The second health notification text may also be one of the multiple health notification texts in the preset text. The preset condition may be any of the conditions described above. Failure to meet the preset condition may indicate that the semantic similarity between the annotation data in the annotation database and the preset text is less than the preset similarity; or that the semantic similarity between the annotation data and the preset text is not the maximum similarity among the semantic similarities between the multiple annotation data and the preset text.

[0048] In the embodiments of this specification, the second health notification text may be the same as the first health notification text, or the second health notification text may be different from the first health notification text. The first sample association relationship data may be obtained by annotating the second health notification text by an expert. The first inference data may be obtained using a large model, specifically, it may be the inference step information of obtaining the first sample association relationship data based on the second health notification text using a large model. The first inference data may guide the process of the first large model inferring the preset text, and may correct the inference steps in the inference process to improve the accuracy of the processing result of the first large model processing the preset text.

[0049] In the embodiments of this specification, the first large model can be a smaller-scale large model with fewer parameters in the large model, such as a strong base model with parameters of the order of 1B or 2B. The strong base model can be a general artificial intelligence model that has undergone large-scale pre-training and has strong generalization capabilities, such as strong base models such as GPT-4, Gemini1.5, and LLaMA3. The target annotation data can guide the first large model to process the preset text according to the format of the target annotation data or the generated data type to obtain association relationship data corresponding to the preset text.

[0050] In the embodiments of this specification, the association relationship data corresponding to the preset text can be in the form of a knowledge graph, and specifically can include entity information, relationship information, and attribute information. Entity information can be information describing an entity, such as a symptom entity or a result entity; relationship information can be information describing the relationship between entity information and attribute information, or information describing the relationship between two entities, such as the means used or the inclusion relationship; and attribute information can be information describing the attributes of an entity, such as a time attribute or a description content attribute corresponding to an entity.

[0051] As an implementation method, some health notification texts in the preset text can be processed in advance as examples for processing other health notification texts in the preset text, thereby improving the accuracy and efficiency of obtaining the association relationship data corresponding to the preset text. Optionally, in an embodiment of this specification, before obtaining the target annotation data whose semantic similarity with the preset text meets the preset conditions, it can also include: obtaining a third health notification text in the preset text; obtaining annotation information for the third health notification text to obtain first annotation data; the first annotation data includes the third health notification text and the second sample association relationship data corresponding to the third health notification text; the second sample association relationship data includes the annotation information; the first annotation data is input into the second large model to obtain the second annotation data obtained by further annotating the first annotation data with reference to the third health notification text; the second annotation data includes the target annotation data; the second annotation data includes the third health notification text, the second sample association relationship data corresponding to the third health notification text, and the second reasoning data for reflecting the reasoning process of determining the second sample association relationship data based on the third health notification text.

[0052] In the embodiments of this specification, the third health notification text may be any one of a plurality of health notification texts selected from a preset text set as a small sample example. The third health notification text may be the same as the second health notification text, or the third health notification text may be different from the second health notification text. The annotation information may be relationship information, attribute information, or other information annotated in the third health notification text based on an expert annotation method; or the second sample association relationship data annotated in the third health notification text based on an expert annotation method.

[0053] In the embodiments of this specification, the second largest model can be a larger model with a larger number of parameters within the large model, such as a strong base model with 10B or 20B parameters. The first largest model can be a model distilled from the second largest model. It is understood that the second largest model is the teacher model and the first largest model is the student model. The first largest model has fewer parameters but can process data in the same way as the second largest model.

[0054] In an embodiment of the present specification, if the third health notification text is the second health notification text mentioned above, the second annotation data may be the target annotation data. The second annotation data may be second reasoning data that adds the reasoning process of determining the second sample association relationship data through the third health notification text on the basis of retaining the first annotation data. The second reasoning data may be determined by the large model through the method of thinking chain. If the third health notification text is the second health notification text mentioned above, the second reasoning data may be the first reasoning data mentioned above. In this way, part of the health notification text in the preset text can be processed, and the annotation data obtained after processing can be added to the annotation database, so that when the first large model is used to process the remaining preset text, it can be used as a small sample to prompt the first large model, thereby improving the accuracy of the processing result of the first large model.

[0055] As an implementation method, in order to improve the accuracy of the annotation information of the third health notification text, the third health notification text can also be sliced ​​first, and then the sliced ​​data can be annotated after slicing to avoid errors in the recognition of the first large model and output of erroneous results. Optionally, in the embodiment of this specification, obtaining the annotation information for the third health notification text can specifically include: inputting the third health notification text into the third large model to obtain a number of slice data corresponding to the third health notification text output by the third large model; obtaining annotation information for a number of the slice data to obtain a number of sub-annotation data; one sub-annotation data includes one slice data, and sub-sample association relationship data corresponding to the slice data.

[0056] In the embodiments of this specification, the third model can be a Large Language Model (LLM). The server can input the third health notification text into the third model, which can then segment the third health notification text into sentences to generate a plurality of slices of data. The slices can be complete sentences in the third health notification text, or paragraphs corresponding to a heading in the third health notification text.

[0057] In the embodiments of this specification, the annotation information of the slice data may include relationship information and attribute information annotated for the slice data. After obtaining a plurality of sub-annotation data, the plurality of sub-annotation data containing annotation information and slice data belonging to the same health notification text may be concatenated according to the order of the slice data in the third health notification text to obtain first annotation data, and the first annotation data may be stored in the annotation database.

[0058] In actual applications, the server can directly store the sub-annotation data corresponding to the slice data in the annotation database, and use the sub-annotation data as a sample example to guide the first large model to process the preset text. This can reduce the amount of text input into the first large model while providing the first large model with reliable small sample examples, thereby ensuring the accuracy of the first large model in processing the text while improving the rate at which the first large model processes the preset text.

[0059] As an embodiment, the server selects a small amount of health notification text from the preset text as health notification samples to construct an annotation database, so as to provide a small number of sample examples for the subsequent processing of the remaining health notification text in the preset text. In the process of constructing the small number of sample examples, the health notification text can be sliced ​​and the small number of sample examples can be constructed based on the slice data. The method may also include: inputting each health notification sample into the third largest model to obtain a plurality of slice data corresponding to each health notification sample output by the third largest model; determining a preset number of sample slice data as samples from the plurality of slice data; annotating the preset number of sample slice data to obtain sample annotation data; the sample annotation data may include relationship information, attribute information and sample slice data; determining target slice data that does not belong to the sample slice data in the plurality of slice data; inputting the preset number of sample annotation data and the target slice data into the second largest model to obtain slice annotation data corresponding to the target slice data output by the second largest model; the slice annotation data may include the target slice data, the associated relationship data corresponding to the target slice data and the inference data of the inference process of the second largest model determining the associated relationship data from the target slice data.

[0060] As an implementation method, the above-mentioned acquisition of target annotation data whose semantic similarity with the preset text meets preset conditions may specifically include: calculating the semantic similarity between the preset text and each slice annotation data; taking the slice annotation data corresponding to the maximum semantic similarity as the target annotation data; or calculating the semantic similarity of each slice annotation data in the preset text, and taking several slice annotation data corresponding to the semantic similarity greater than or equal to the preset similarity as the target annotation data.

[0061] In practical applications, the data input into the first large model also includes prompt words for prompting the first large model to complete the task. In order to facilitate understanding of the embodiments of this specification, an example of prompt words for inputting the first large model is provided. The following example: As an underwriting expert, please extract relationships and attributes from the following health declaration as required. #Relationship Extraction: Contains: Contains relationships, including XXX, containing XXX. Examinations: Medical examinations, including CT, MRI, blood pressure, blood sugar, etc.

[0062] #Attribute extraction: Time: Time range, within XXX months or XXX years. Frequency: XX consecutive normal tests. Duration: Duration, more than XXX weeks.

[0063] #Example: Input: In the past 6 months, any of the following conditions has occurred: prolonged fever (lasting 2 weeks), blood in the stool (non-hemorrhoidal bleeding), white blood cell count (WBC) less than 1.5*10^9 / L or greater than 15*10^9 / L, or elevated tumor marker test results (indicators exceeding the normal reference range).

[0064] Relationship: [blood in stool, excluding hemorrhoidal bleeding], [white blood cell count, examination, less than 1.5x10^9 / L], [white blood cell count, examination, more than 15x10^9 / L], [tumor marker, examination, high result (referring to exceeding the normal reference range)].

[0065] Attributes: [Long-term fever, lasting for more than 2 weeks], [Long-term fever, duration, within the last 6 months].

[0066] Expert Explanation: The health notice specifies that the listed conditions are not covered within the past six months, so the timeframe is: past six months. Furthermore, the notice states that a prolonged fever is limited to more than two weeks; bloody stools do not include hemorrhoidal bleeding. Furthermore, white blood cell counts and tumor markers are considered part of the physical examination.

[0067] The input health notification text XXX needs to be processed by the large model.

[0068] In practical applications, the prompt words input into the first and second models can be generated in the format of the prompt word examples above. Both the first and second models can refer to the examples to determine the relationship information and attribute information from the health notification text or slice data to be processed. The first model has fewer parameters than the second model, and the first model can fuse the attribute information and relationship information extracted from the health notification text to obtain association relationship data with a preset structure, such as data in a knowledge graph structure.

[0069] In order to clearly understand the process of building a labeling database in the embodiments of this specification, Figure 3 This is a schematic diagram of the processing flow of a sample health notification text provided in the embodiment of this specification. Figure 3As shown, the server can input sample health notification text into the third-largest model. The third-largest model slices the sample health notification text, dividing it into semantically indivisible segments to obtain slice data. A small amount of slice data is obtained from the slice data through expert annotation and annotated to obtain sample annotated data. Based on the remaining slice data, the server can obtain a small amount of sample annotated data from the sample annotated data as a fewshot example. The remaining slice data and the fewshot example are organized to obtain prompt words, which are then input into the second-largest model. The prompt words may include task prompts that prompt the second-largest model to perform relationship extraction on the remaining slice data and task prompts that prompt the second-largest model to perform attribute extraction on the remaining slice data. The second-largest model can extract relationship information and attribute information from the remaining slice data based on the contextual learning paradigm and reference the fewshot example. The second-largest model can correct and interpret the extraction process based on the fewshot example and output the annotated data for each slice in the sample health notification text. The slice annotated data contains not only relationship information and attribute information, but also inference data representing the second-largest model's reasoning process on the remaining slice data. The server can build an annotation database based on the obtained multiple slice annotation data.

[0070] Accordingly, Figure 4 This is a flow chart of determining the association relationship data of a preset text provided in the embodiment of this specification. Figure 4 As shown, the server can use RAG technology to obtain a small number of examples from the annotation database whose semantic similarity with the preset text meets the preset conditions. The small number of examples and the preset text can be sorted to obtain prompt words. The prompt words can include task prompts that prompt the first large model to perform relationship extraction on the remaining slice data, and task prompts that prompt the first large model to perform attribute extraction on the remaining slice data. The prompt words are input into the first large model. The first large model can also complete the relationship information extraction and attribute information extraction respectively based on the context learning paradigm. The first large model can fuse the extracted relationship information and attribute information, and output the associated relationship data corresponding to the preset text that meets the preset structure.

[0071] To clearly illustrate the structure of the associated data, we take the following input as an example: "The insured has had abnormal test results in the past year: blood sugar, liver function, thyroid function, kidney function (urea, creatinine, cystatin C), myocardial enzymes, tumor markers, urine test, electrocardiogram, electroencephalogram, imaging, endoscopy, pathology." Figure 5 This is a schematic diagram of association relationship data provided in the embodiment of this specification. Figure 5As shown, the second largest model can extract the relationship information and attribute information in the example and fuse them to obtain two sets of knowledge graphs. The knowledge graph can include information such as disease entities, relationships, attributes, and result entities. Knowledge graph 1 can include the disease entity "blood sugar", the relationship "examination", the attribute "time: within the past year", and the result entity "abnormal result". Knowledge graph 1 can include the disease entity "renal function", the relationship "examination", the attribute "time: within the past year", "content: urea, creatinine, cystatin C", and the result entity "abnormal result". Figure 5 It is only for illustration and can serve as a specific limitation of the structure of the association relationship data. In addition to the above-mentioned knowledge graph, it can also be other structures that can be recognized by the big model, such as data tables and other structures.

[0072] As an implementation mode, optionally, the generation of knowledge extension information for the user input sample based on the association relationship data as described in the embodiments of this specification may specifically include: determining first association data from the association relationship data based on the user input sample; if the first association data is entity data extracted from the first health notification text, then determining the first association data as the knowledge extension information.

[0073] In the embodiments of this specification, the first associated data may be data in the associated relationship data that is highly correlated with or similar to the user input sample. The entity data may be the disease entity mentioned in the first health notification text. The knowledge expansion information may indicate information explaining how the entity data is determined based on the user input sample.

[0074] In the embodiments of this specification, the first associated data is determined from the associated relationship data based on the user input sample. Specifically, the user input sample is input into an extended model. The extended model can query a database storing associated relationship data for data with the highest similarity to the user input sample, or a similarity greater than or equal to a preset value, and use that data as the first associated data. The extended model can be a strong base model, deployed on a server. Alternatively, it can be determined using RAG technology.

[0075] In practical applications, in a database storing relational data, if the relational data is pre-set text, a corresponding health notification identifier can be added to each relational data. The health notification identifier is used to uniquely identify each health notification text. When a user consults about insurance or trains a model, the server can use the extended model to quickly determine the relational data corresponding to the health notification text corresponding to the insurance policy the user is consulting from the database based on the health notification identifier. This eliminates the need to perform similarity calculations with the relational data for each health notification text, thereby improving the efficiency of the strong base model in processing text. If the relational data is basic knowledge, there is no need to identify the relational data.

[0076] In actual applications, among the multiple preset structured association data stored in the database storing association data, the same data may exist in data belonging to different association combinations. For example, the entity data "leukemia" exists in the association combination corresponding to the health notification text 1, the entity data "leukemia" may also exist in the association combination corresponding to the health notification text 2, and the association data "leukemia" may also exist in the association combination corresponding to the basic knowledge.

[0077] As another implementation, optionally, the method described in the embodiment of this specification may also include: if the first associated data is data extracted from the basic knowledge, then obtaining second associated data that has an associated relationship with the first associated data from the associated relationship data; if the second associated data is entity data extracted from the first health notification text, then determining the first associated data and the second associated data as the knowledge extension information.

[0078] In an embodiment of the present specification, the second association relationship data may be both data extracted from the basic knowledge and entity data extracted from the first health notification text. Specifically, if the server fails to obtain association data whose similarity with the user input sample meets the preset conditions from the association relationship data corresponding to the first health notification text in the database using the extended model, the server may obtain the first association data whose similarity with the user input sample meets the preset conditions from the association relationship data corresponding to the basic knowledge, and may further determine the second association data having an association relationship with the first association data from the association relationship combination where the first association data is located; if the second association data is also included in the association relationship data corresponding to the first health notification text, it may be determined that the second association relationship data also belongs to the data extracted from the first health notification text. In an embodiment of the present specification, the knowledge extension information may be generated based on the first association data, the second association data and the user input sample. Specifically, the knowledge extension information may be used to explain the process of determining the first association data from the user input sample and further determining the second association data.

[0079] In practical applications, after the server receives a user input sample, it can be fed into an expansion model, which can then output knowledge expansion information corresponding to the user input sample. Alternatively, the server can retrieve the user input sample from a database using RAG technology and generate knowledge expansion information based on the retrieval results. This can generate training sample data containing knowledge expansion information, improving the accuracy of text processing results from a large language model trained using the training sample data for health notification.

[0080] By generating training sample data containing factor information and explanatory information, the model obtained after training this type of training sample data can be based on user consultation questions and can also output corresponding factor information and explanatory information for easy user understanding. Optionally, the method described in the embodiment of this specification may also include: generating factor information and explanatory information based on the knowledge expansion information and the first health notification text; the factor information represents the entity information belonging to the first health notification text in the knowledge expansion information; the explanatory information is used to explain the reasoning process of the pre-trained large model to determine the label data based on the user input sample and the factor information; generating training sample data based on the initial training data and the knowledge expansion information, which may specifically include: generating the training sample data based on the initial training data, the knowledge expansion information, the factor information and the explanatory information.

[0081] In the embodiment of this specification, the factor information may also be entity information with attribute information and relationship information in the first health notification text, such as abnormal blood sugar index results found within one year. Explanation information may be generated based on user input samples, factor information, and knowledge extension information. Specifically, the explanation information may include the process of determining the factor information represented in the knowledge extension information based on the user input sample, and the reason for obtaining the label data based on the factor information. For example, if the factor information is "malignant tumor" and the user input sample is "cholecystectomy", the explanation information may be "the indications for cholecystectomy include malignant tumors, so malignant tumors may be involved, and malignant tumors are clearly not insurable in the health notification." In the embodiment of this specification, the training sample data may be obtained by splicing the initial training data, knowledge extension information, factor information, and explanation information. There is no limitation on the order of splicing here, and it can be set based on actual needs.

[0082] To facilitate understanding of the training sample data in the embodiments of this specification, the following is an example of training sample data provided in the embodiments of this specification.

[0083] You are an underwriting expert. Based on the product health notice, the user's input, and the relevant extended knowledge, give one of the following conclusions: [Fail, Likely Fail, Likely Pass, Pass, No health notice required]. Return the relevant symptoms and provide an explanation.

[0084] Health notification text.

[0085] User input: Gallbladder removal.

[0086] Knowledge expansion: Cholecystectomy is a surgical procedure mainly used for gallstones, cholecystitis, gallbladder polyps and gallbladder cancer. Gallbladder cancer is a type of malignant tumor.

[0087] Please input in the following format: Conclusion: May not pass.

[0088] Associated symptomatic factors: malignant tumors.

[0089] Explanation: The indications for cholecystectomy include malignant tumors, so malignant tumors may be involved, but malignant tumors are clearly not insurable in the health declaration.

[0090] The text input by the user is processed by the large language model for health notification, so that the user can get accurate underwriting advice even without professional knowledge. Optionally, the method described in the embodiment of this specification may also include: obtaining the user input text; obtaining the fourth health notification text corresponding to the user input text in the preset text; generating target knowledge extension information for the user input text based on the association relationship data; inputting the fourth health notification text, the user input text and the target knowledge extension information into the large language model for health notification to obtain analysis result information; the analysis result information includes first result information or second result information; the first result information indicates that the user input text belongs to the description scope of the fourth health notification text; the second result information indicates that the user input text does not belong to the description scope of the fourth health notification text.

[0091] In the embodiments of this specification, the user input text may be a text description of the user's health condition. The fourth health notification text may be the health notification text corresponding to the insurance policy the user wishes to purchase. The target knowledge extension information may be extended information obtained by processing the user input text based on an extension model using association data.

[0092] In an embodiment of this specification, the analysis result information may include insurance recommendation information, such as recommending insurance or not recommending insurance. The description range of the fourth health notification text may indicate the uninsurable range described in the fourth health notification text. The first result information may indicate that the user input text falls within the uninsurable range, and a result of not recommending insurance is obtained. The first result information may also include factor information that the user input text hits the uninsurable range described in the fourth health notification text, as well as explanatory information explaining why insurance is not recommended.

[0093] In the embodiment of this specification, the second result information may indicate that the user input text does not fall within the scope of non-insurable items, and a recommended insurance result is obtained. The second result information may also contain explanatory information explaining the recommended insurance, and may not contain factor information. If the fourth health notification text describes some insurable health abnormalities, and the user input text hits one of them, the second result information may also contain factor information. For example, the health notification text describes "The following conditions can be insured, respiratory department: colds, rhinitis"; the user input text is "I had a cold this week", then a recommended insurance can be output, the factor information is "cold", and the explanatory information is "cold is an insurable condition in the health notification, so insurance is recommended."

[0094] If the user input text lacks some insurance advice judgments, the expansion model can also be used to interact with the user to obtain more information, so as to improve the accuracy of generating the target knowledge expansion information. Optionally, the generation of target knowledge expansion information for the user input text based on the association relationship data described in the embodiments of this specification may specifically include: determining from the association relationship data first target association data whose similarity to the user input text meets the similarity condition; generating a model query text based on second target association data associated with the first target association data; obtaining the user's answer text input by the user for the model query text; and generating the target knowledge expansion information based on the model query text and the user answer text.

[0095] In an embodiment of the present specification, the first target associated data may be one or more associated data in the basic knowledge category or the fourth health notification text category in the associated relationship data, whose similarity with the user input text is greater than or equal to a preset value, or a set number of associated data with the greatest similarity. The second target associated data may be data in the same associated relationship data combination as the first target associated data. For example, an associated relationship combination includes a symptom entity, a relationship information, and an attribute information. If the first target associated relationship data is a symptom entity, the second target associated data may be at least one of the relationship information and the attribute information.

[0096] In the embodiments of this specification, the model query text is a descriptive text generated by the server after processing the second target associated data using the extended model, used to ask the user a question. For example, if the user inputs the text "My blood sugar index value is a bit abnormal," the server identifies the "blood sugar" and "abnormal results" in the health notice of the user's insurance policy that are highly similar to the user input text; and based on "blood sugar," the server finds the relationship information "examination" and the attribute information "time: within the past year." The generated model query text may be "Did you get the examination within the past year?" This model query text may be displayed in the interface for interacting with the user.

[0097] In the embodiments of this specification, the user answer text can be the text of the user's corresponding reply to the model query text; if the user does not understand the model query text, he or she can also enter text indicating that he or she does not understand. After receiving the user's answer information, the server can also input it into the extended model. The extended model regenerates text that is easy for the user to understand, and then displays it to the user to obtain the user's answer text.

[0098] In practice, the server can input the model query text and the user's response text into the extended model. The extended model can then generate target knowledge extension information based on the user's response text, the user input text, and the target-related data. Continuing with the "blood sugar" example above, if the user's response text is "yes," the target knowledge extension information would be "The user has had abnormal blood sugar results in the past year."

[0099] In order to clearly explain the use of the large language model for health notification in the embodiment of this specification, Figure 6 This is a flow chart of a method for processing user input text using a large language model for health notification provided in an embodiment of this specification. Figure 6As shown, the Query can be text information entered by the user. The server can use the knowledge expansion method to obtain a set number of target related data with the highest similarity to the user-entered Query, or target related data with a similarity greater than or equal to a preset value, from the related data of the basic knowledge type contained in the association relationship data or the related data of the health declaration corresponding to the insurance policy purchased by the user. The knowledge expansion method can be implemented using an expansion model. The expansion model can interact with the user in multiple rounds based on the target related data found by the user-entered Query, and perform thinking and reasoning based on the results of each interaction to generate target knowledge expansion information. For example, if a user enters the query "I've had a cholecystectomy before," the extended model will perform knowledge expansion on "cholecystectomy." Specifically, the associated data "cholecystectomy" is found in the database. The reasoning is confirmed to be correct by considering "cholecystectomy is a surgical procedure used to remove the patient's gallbladder." Further, based on the data related to cholecystectomy, "gallbladder cancer" and "cholecystitis," the reasoning is confirmed to be correct by considering "cholecystectomy is a surgical procedure primarily used to treat gallstones, cholecystitis, gallbladder polyps, and gallbladder cancer." The model generates the query text "Did you have a cholecystectomy due to gallbladder cancer or cholecystitis?" If the user answers "gallbladder cancer," the data related to gallbladder cancer, "malignant tumor," is found in the database. The reasoning is confirmed to be correct by considering "cholecystectomy is a surgical procedure primarily used to treat gallstones, cholecystitis, gallbladder polyps, and gallbladder cancer, and gallbladder cancer is a type of malignant tumor." Since malignant tumors are explicitly mentioned in the health notification text, malignant tumors are identified as health notification factors. Target knowledge expansion information is then generated based on this thought process. The server can input the knowledge expansion information, health notification text, and user-entered queries into a large language model for health notifications. The large language model for health notifications processes the input data and outputs analysis results. The analysis results can include pre-underwriting results, factor information, and explanation information.

[0100] Figure 7 This is a flow chart of an analysis method for health notification provided in an embodiment of this specification. From a program perspective, the execution subject of the process can be a program installed on an application server or an application client.

[0101] like Figure 7 As shown, the method may include the following steps.

[0102] Step 702: Obtain user input text.

[0103] In the embodiments of this specification, the user input text can be a query text entered by the user regarding the insurance policy they wish to purchase, or a text indicating their health status. The user can select the insurance policy they wish to purchase on the terminal interface and, after clicking on the policy, can choose to enter a consultation interface; alternatively, after clicking on the policy, the user can directly jump to the consultation interface, where the user can enter a consultation text regarding the policy.

[0104] Step 704: Obtain the target health notification text corresponding to the user input text in the preset text.

[0105] In the embodiments of this specification, the preset text can be multiple health notification texts; the user input text can be the health status information entered by the user for the target health notification text. The target health notification text can be a corresponding health notification text determined based on the insurance policy clicked by the user, or the user can proactively send a link to the insurance policy they want to consult in the consultation interface, and the server can determine the corresponding target health notification text based on the insurance policy link.

[0106] Step 706: Generate target knowledge expansion information for the user input text based on the association relationship data.

[0107] The association relationship data is data used to reflect the association relationship between multiple pieces of data extracted from the preset text and basic knowledge in the field to which the preset text belongs.

[0108] In an embodiment of this specification, the association relationship data may be structured data that associates multiple data items after pre-processing a preset text using a large model. The association relationship data may also be basic knowledge with a structure that can be recognized by the model. In an embodiment of this specification, the server may obtain from the association relationship data one or more association data items whose similarity to the user input text is greater than or equal to a preset value, or obtain the top N association data items with the largest similarity values, and generate target knowledge expansion information based on the one or more association data items. The target knowledge expansion information may include information about the reasoning process or thought process used to determine the association data from the user input text.

[0109] In actual applications, the server can also use the extension model to interact with the user to improve the accuracy of the generated target knowledge extension information. The specific implementation method can refer to the above embodiment and will not be described in detail here.

[0110] Step 708: Input the target health notification text, the user input text, and the target knowledge expansion information into the large language model for health notification to obtain target analysis result information.

[0111] Among them, the target analysis result information includes first target result information or second target result information; the first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

[0112] In the embodiments of this specification, the large language model for health notification has the function of outputting a pre-underwriting conclusion for the user based on the input data. The large language model for health notification can be obtained by fine-tuning the pre-trained large model using the training data corresponding to the business scenario. The specific training method can be referred to the above description of model training, and will not be elaborated here.

[0113] In the embodiments of this specification, the target analysis result information may include pre-underwriting conclusion information, factor information determined from the target health notification text for the user input text, and explanatory information used to explain how the pre-underwriting conclusion was obtained based on the input data. The description range may be the uninsurable range described in the target health notification text. The first target result information may indicate that the pre-underwriting was not passed; the second target result information may indicate that the pre-underwriting was passed.

[0114] In the embodiments of this specification, user input text can be processed by combining knowledge expansion with a large language model for health notification, and a pre-underwriting conclusion can be output to the user, allowing the user to obtain accurate underwriting advice even without professional knowledge, thereby protecting the user's rights and interests. Knowledge expansion can also be used to provide more input data for the large language model for health notification, alleviating knowledge forgetting in the large language model for health notification and eliminating large model illusions.

[0115] Through the above method, on the one hand, basic knowledge and multiple health notification texts can be used to build a database corresponding to the associated relationship data, which can be used as reference data for knowledge expansion of the user input text to eliminate the illusion of the existence of the large model and improve the accuracy of the large model in processing the text. On the other hand, any health notification text can be disassembled into associated relationship data in the above manner to facilitate data search and large model understanding. On the third hand, the large language model used for health notification can automatically understand the text input by the user, and combine the knowledge expansion method to output the health notification content that may be involved in the user's consultation on the disease or condition, which can help users quickly decide whether to take out insurance.

[0116] Based on the same idea, the embodiments of this specification also provide a device corresponding to the above method. Figure 8 The embodiments of this specification provide corresponding Figure 2A schematic diagram of a device for training a large language model for health information. Figure 8 As shown, the device may include: Initial training data acquisition module 802 is used to acquire initial training data; the initial training data includes sample data and label data for the sample data; the sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample falls within the description range of the first health notification text; the second label data indicates that the user input sample does not fall within the description range of the first health notification text; The knowledge extension information generation module 804 is configured to generate knowledge extension information for the user input sample based on association relationship data; the association relationship data is configured to reflect association relationships between multiple pieces of data extracted from a preset text and basic knowledge in the field to which the preset text belongs; the preset text includes at least the first health notification text; A training sample data generating module 806 is configured to generate training sample data based on the initial training data and the knowledge expansion information; The fine-tuning training module 808 is used to use the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain the large language model for health notification.

[0117] based on Figure 8 The embodiment of this specification also provides some specific implementation plans of the method, which are described below. Optionally, the device may also include an association relationship data generation module, which can be specifically used to: obtain the preset text; obtain target annotation data whose semantic similarity with the preset text meets the preset conditions; the target annotation data includes a second health notification text, a first sample association relationship data corresponding to the second health notification text, and a first reasoning data for reflecting the reasoning process of determining the first sample association relationship data based on the second health notification text; input the preset text and the target annotation data into the first large model to obtain the association relationship data corresponding to the preset text output by the first large model.

[0118] Optionally, the association relationship generation module can also be used to: obtain the third health notification text in the preset text; obtain the annotation information for the third health notification text to obtain first annotation data; the first annotation data includes the third health notification text and the second sample association relationship data corresponding to the third health notification text; the second sample association relationship data includes the annotation information; input the first annotation data into the second large model to obtain the second annotation data obtained by further annotating the first annotation data by the second large model with reference to the third health notification text; the second annotation data includes the target annotation data; the second annotation data includes the third health notification text, the second sample association relationship data corresponding to the third health notification text, and the second reasoning data used to reflect the reasoning process of determining the second sample association relationship data based on the third health notification text.

[0119] Optionally, the association relationship generation module can also be used to: input the third health notification text into the third largest model to obtain several slice data corresponding to the third health notification text output by the third largest model; obtain annotation information for several slice data to obtain several sub-annotation data; one sub-annotation data includes one slice data and sub-sample association relationship data corresponding to the slice data.

[0120] Optionally, the knowledge extension information generation module can be specifically used to: determine first associated data from the associated relationship data based on the user input sample; if the first associated data is entity data (gallbladder cancer) extracted from the first health notification text, then determine the first associated data as the knowledge extension information.

[0121] Optionally, the knowledge extension information generation module can be specifically used to: if the first associated data is data extracted from the basic knowledge, then obtain second associated data that has an associated relationship with the first associated data from the associated relationship data; if the second associated data is entity data extracted from the first health notification text, then determine the first associated data and the second associated data as the knowledge extension information.

[0122] Optionally, the device can also be used to: generate factor information and explanatory information based on the knowledge expansion information and the first health notification text; the factor information represents the entity information belonging to the first health notification text in the knowledge expansion information; the explanatory information is used to explain the reasoning process of the pre-trained large model to determine the label data based on the user input sample and the factor information; generate training sample data based on the initial training data and the knowledge expansion information, specifically including: generating the training sample data based on the initial training data, the knowledge expansion information, the factor information and the explanatory information.

[0123] Optionally, the device may also include a model processing module, which can be specifically used to: obtain user input text; obtain a fourth health notification text corresponding to the user input text in the preset text; generate target knowledge extension information for the user input text based on the association relationship data; input the fourth health notification text, the user input text and the target knowledge extension information into the large language model for health notification to obtain analysis result information; the analysis result information includes first result information or second result information; the first result information indicates that the user input text belongs to the description scope of the fourth health notification text; the second result information indicates that the user input text does not belong to the description scope of the fourth health notification text.

[0124] Optionally, the device may also include a model processing module, which can be specifically used to: determine the first target association data whose similarity with the user input text meets the similarity condition from the association relationship data; generate a model inquiry text based on the second target association data associated with the first target association data; obtain the user answer text input by the user for the model inquiry text; and generate the target knowledge extension information based on the model inquiry text and the user answer text.

[0125] Optionally, the preset text is a plurality of health notification texts; and the user input sample is health status information input by the user for the first health notification text.

[0126] Based on the same idea, the embodiments of this specification also provide a device corresponding to the above method. Figure 9 The embodiments of this specification provide corresponding Figure 7 A schematic diagram of the structure of an analysis device for health information. Figure 9 As shown, the device may include: A user input text acquisition module 902 is used to acquire user input text; The health notification text acquisition module 904 is used to acquire a target health notification text corresponding to the user input text from a preset text; The target knowledge extension information generating module 906 is configured to generate target knowledge extension information for the user input text based on the association relationship data; the association relationship data is data used to reflect the association relationship between multiple pieces of data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; Analysis module 908 is used to input the target health notification text, the user input text and the target knowledge expansion information into the large language model for health notification to obtain target analysis result information; the target analysis result information includes first target result information or second target result information; the first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

[0127] based on Figure 9 The present specification also provides some specific implementation plans of the method, which are described below.

[0128] Optionally, the preset text is a plurality of health notification texts; and the user input text is health status information input by the user for the target health notification text.

[0129] Based on the same idea, the embodiments of this specification also provide devices corresponding to the above methods. Figure 10 This is a schematic diagram of the structure of a computer device provided in the embodiment of this specification. Figure 10 As shown, the device 1000 may include: at least one processor 1010; and a memory 1030 communicatively connected to the at least one processor; wherein the memory 1030 stores instructions 1020 that can be executed by the at least one processor 1010, and the instructions are executed by the at least one processor 1010 to enable the at least one processor 1010 to implement the method for training a large language model for health notification described in any of the above embodiments or the analysis method for health notification described in any of the above embodiments.

[0130] Based on the same idea, the embodiments of this specification also provide a computer-readable storage medium corresponding to the above method. A computer-readable storage medium stores a computer program or instruction, which can be executed by a processor to implement the method for training a large language model for health notification described in any of the above embodiments or the steps of the analysis method for health notification described in any of the above embodiments. The various embodiments in this specification are described in a progressive manner, and the same and similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for Figure 10As for the device shown, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0131] In the 1990s, technological improvements could be clearly distinguished as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with the advancement of technology, many process flow improvements can now be considered direct improvements to hardware circuit structures. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using physical hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly performed using software called a "logic compiler." This is similar to the software compilers used during program development. Before compilation, the original code must be written in a specific programming language, called a Hardware Description Language (HDL). There are many types of HDL, including ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that simply by programming a method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0132] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the memory control logic. Those skilled in the art will also appreciate that, in addition to implementing the controller purely in computer-readable program code, the controller can also be implemented in the form of logic gates, switches, an application-specific integrated circuit, a programmable logic controller, an embedded microcontroller, etc. by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the means for implementing the various functions included therein can also be considered as structures within the hardware component. Alternatively, the means for implementing the various functions can be considered both a software module implementing the method and a structure within the hardware component.

[0133] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0134] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0135] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0136] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0139] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in a computer-readable medium, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0140] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented using any method or technology for information storage. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0141] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0142] Those skilled in the art will appreciate that embodiments of the present application may be provided as methods, systems, or computer program products. Thus, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code. The present application may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In distributed computing environments, program modules may be located in local and remote computer storage media, including storage devices.

[0143] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for training a large language model for health information, comprising: Get initial training data; The initial training data includes sample data and label data for the sample data; The sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample falls within the description scope of the first health notification text; the second label data indicates that the user input sample does not fall within the description scope of the first health notification text; Generating knowledge expansion information for the user input sample according to the association relationship data; The association relationship data is used to reflect the association relationship between multiple pieces of data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; the preset text at least includes the first health notification text; generating training sample data based on the initial training data and the knowledge expansion information; The pre-trained large model is supervised and fine-tuned using the training sample data to obtain the large language model for health notification.

2. The method according to claim 1, before generating the knowledge expansion information for the user input sample based on the association relationship data, further comprising: Obtaining the preset text; Acquire target annotation data whose semantic similarity with the preset text meets preset conditions; The target annotation data includes a second health notification text, first sample association relationship data corresponding to the second health notification text, and first reasoning data for reflecting a reasoning process of determining the first sample association relationship data based on the second health notification text; The preset text and the target annotation data are input into a first large model to obtain association relationship data corresponding to the preset text output by the first large model.

3. The method according to claim 2, before obtaining target annotation data whose semantic similarity with the preset text meets a preset condition, further comprising: Obtaining a third health notification text from the preset text; Acquire annotation information for the third health notification text to obtain first annotation data; The first annotation data includes the third health notification text and second sample association relationship data corresponding to the third health notification text; the second sample association relationship data includes the annotation information; The first annotated data is input into the second large model to obtain second annotated data obtained by further annotating the first annotated data by the second large model with reference to the third health notification text; the second annotated data includes the target annotated data; the second annotated data includes the third health notification text, second sample association relationship data corresponding to the third health notification text, and second inference data used to reflect the inference process of determining the second sample association relationship data based on the third health notification text.

4. The method according to claim 3, obtaining annotation information for the third health notification text, specifically comprising: Inputting the third health notification text into the third large model to obtain a plurality of slice data corresponding to the third health notification text output by the third large model; The labeling information of a plurality of the slice data is acquired to obtain a plurality of sub-labeling data; one sub-labeling data includes one slice data and sub-sample association relationship data corresponding to the slice data.

5. The method according to claim 1, wherein generating knowledge expansion information for the user input sample based on the association relationship data specifically comprises: Determining first associated data from the associated relationship data according to the user input sample; If the first associated data is entity data extracted from the first health notification text, the first associated data is determined as the knowledge extension information.

6. The method according to claim 5, further comprising: If the first associated data is data extracted from the basic knowledge, obtaining second associated data having an associated relationship with the first associated data from the associated relationship data; If the second associated data is entity data extracted from the first health notification text, the first associated data and the second associated data are determined as the knowledge extension information.

7. The method according to claim 1, further comprising: Generate factor information and explanation information based on the knowledge expansion information and the first health notification text; The factor information represents entity information belonging to the first health notification text in the knowledge extension information; The explanation information is used to explain the reasoning process of the pre-trained large model determining the label data based on the user input sample and the factor information; Generating training sample data based on the initial training data and the knowledge expansion information specifically includes: The training sample data is generated based on the initial training data, the knowledge expansion information, the factor information, and the explanation information.

8. The method according to claim 1, further comprising: Get user input text; Obtaining a fourth health notification text corresponding to the user input text from the preset text; generating target knowledge expansion information for the user input text according to the association relationship data; The fourth health notification text, the user input text and the target knowledge expansion information are input into the large language model for health notification to obtain analysis result information; the analysis result information includes first result information or second result information; the first result information indicates that the user input text belongs to the description scope of the fourth health notification text; the second result information indicates that the user input text does not belong to the description scope of the fourth health notification text.

9. The method according to claim 8, wherein generating target knowledge expansion information for the user input text based on the association relationship data specifically comprises: Determining first target associated data from the associated relationship data, the similarity of which with the user input text satisfies a similarity condition; generating a model query text based on second target-related data associated with the first target-related data; Obtaining a user response text input by the user in response to the model query text; The target knowledge expansion information is generated based on the model query text and the user answer text.

10. A method for analyzing health information, comprising: Get user input text; Obtaining a target health notification text corresponding to the user input text in a preset text; Generating target knowledge expansion information for the user input text according to the association relationship data; The association relationship data is data used to reflect the association relationship between multiple pieces of data extracted from the preset text and basic knowledge in the field to which the preset text belongs; Inputting the target health notification text, the user input text, and the target knowledge expansion information into a large language model for health notification to obtain target analysis result information; The target analysis result information includes first target result information or second target result information; The first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

11. A device for training a large language model for health information, comprising: An initial training data acquisition module, used to acquire initial training data; The initial training data includes sample data and label data for the sample data; the sample data includes a user input sample and a first health notification text; the label data includes first label data or second label data; the first label data indicates that the user input sample falls within the description scope of the first health notification text; the second label data indicates that the user input sample does not fall within the description scope of the first health notification text; A knowledge extension information generation module, configured to generate knowledge extension information for the user input sample based on the association relationship data; The association relationship data is used to reflect the association relationship between multiple pieces of data extracted from the preset text and the basic knowledge of the field to which the preset text belongs; the preset text at least includes the first health notification text; A training sample data generating module, configured to generate training sample data based on the initial training data and the knowledge expansion information; The fine-tuning training module is used to use the training sample data to perform supervised fine-tuning training on the pre-trained large model to obtain the large language model for health notification.

12. An analysis device for health notification, comprising: User input text acquisition module, used to obtain user input text; A health notification text acquisition module is used to acquire a target health notification text corresponding to the user input text in a preset text; A target knowledge extension information generation module, configured to generate target knowledge extension information for the user input text based on the association relationship data; The association relationship data is data used to reflect the association relationship between multiple pieces of data extracted from the preset text and basic knowledge in the field to which the preset text belongs; an analysis module, configured to input the target health notification text, the user input text, and the target knowledge expansion information into a large language model for health notification to obtain target analysis result information; The target analysis result information includes first target result information or second target result information; The first target result information indicates that the user input text belongs to the description scope of the target health notification text; the second target result information indicates that the user input text does not belong to the description scope of the target health notification text.

13. A computer device comprising: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed 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 implement the method for training a large language model for health notification described in any one of claims 1 to 9 or the analysis method for health notification described in claim 10.

14. A computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the method for training a large language model for health notification described in any one of claims 1 to 9 or the steps of the analysis method for health notification described in claim 10.

Images