Old age health management education interaction method and device, computer device and product

Abstract

The application relates to the technical field of artificial intelligence generation, in particular to an old-age health management education interaction method and device, computer equipment and products. The old-age health management education interaction method comprises the following steps: receiving an education instruction; performing education task characteristic analysis on the education instruction, generating a modulation characteristic description based on the analysis, and the modulation characteristic description representing the deviation requirement of the education task for the knowledge organization path of the health education content; injecting the modulation characteristic description into the generation context of a large language model; and the large language model generates the health education content based on the generation context. The method provided by the application can more adaptively provide health guidance matched with the actual needs of old-age users.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence generation technology, and in particular to an interactive method, device, computer equipment and product for elderly health management education. Background Technology

[0002] With the increasing capabilities of large language models in natural language generation, their application in health management education has become a technological trend. In health management education for elderly users, existing technologies typically use large language models to receive health education-related instructions or questions and generate user-oriented health education text content.

[0003] To ensure that generated content meets the needs of elderly users, current practices typically involve requiring large language models to use simplified language expressions through instruction design or prompt word engineering, or setting up a dialogue management module external to the large language model to control the interaction flow and topic selection. In this process, existing systems usually extract health-related named information variables (such as disease diagnosis, medication regimens, and physiological indicator data) from the user's dialogue input. This extracted information is used as contextual input for the large language model to generate educational content, and information variables successfully extracted from user input are considered as known information already possessed by the user, adjusting the subsequent educational content generation strategy accordingly.

[0004] However, existing technologies still have some limitations, making it difficult for elderly users to obtain guidance that matches their actual needs from health education content generated by large language models. Therefore, existing interactive methods for elderly health management education still need improvement. Summary of the Invention

[0005] To address, or at least partially address, the aforementioned technical problems, this application provides an interactive method, device, computer equipment, and product for elderly health management education, which can provide health guidance that is more tailored to the actual needs of elderly users.

[0006] Firstly, this application provides an interactive method for elderly health management education, which includes the following steps: Receive educational instructions; The educational instructions are subjected to educational task feature analysis, and a modulation feature description is generated based on the analysis. The modulation feature description characterizes the offset requirement of the educational task on the knowledge organization path of health education content. The modulation feature description is injected into the generation context of the large language model; The large language model generates the health education content based on the generated context.

[0007] Optionally, the modulation feature description includes a center offset direction and an allowable domain. The center offset direction is inferred based on information available in the educational instructions, and the allowable domain defines an optional range of knowledge organization paths with the center offset direction as the anchor point. The large language model autonomously selects a knowledge organization path based on the knowledge structure encoded in its own parameter space within the scope defined by the permissible domain.

[0008] Optionally, the width of the allowable field is determined based on the information missing degree identification result; The information missing degree identification includes: The information variables in the educational instructions are evaluated for their status. The information variables include named information variables and unnamed information variables. The named information variables are those whose values ​​can be extracted from user input, and the unnamed information variables are those whose values ​​cannot be extracted from user input. The more unnamed information variables there are and the greater their influence weight on the corresponding modulation dimension, the wider the allowable domain will be. The more named information variables there are, the narrower the allowed domain becomes.

[0009] Optionally, the method further includes: Extract named information variables that are in a known state from the educational instructions, and evaluate the cognitive anchoring degree of the variable in the user's cognitive structure based on the relationship between the way the variable value is presented in the user input and the surrounding text. The contribution of the named information variable to the narrowing of the allowable region is obtained by weighting the cognitive anchoring degree. The higher the cognitive anchoring degree of the named information variable, the greater its contribution to the narrowing of the allowable region. The lower the cognitive anchoring degree of the named information variable, the smaller its contribution to the narrowing of the allowable region.

[0010] Optionally, the cognitive anchoring degree is determined based on a comprehensive analysis of the following characteristics: The contextual embedding method of the variable value in the user input, wherein the contextual embedding method represents whether the variable value is embedded in the user's own narrative structure; Does this variable have a user-initiated causal or empirical relationship with other variables? The precision of this variable's value matches the overall level of cognitive expression demonstrated by the user in other parts of the conversation.

[0011] Optionally, the information variables are distinguished into named information variables and unnamed information variables in the following ways: Based on the educational instructions, a list of expected variables related to the user's current consultation question is generated, and the user's input text is searched to see if each variable in the list of expected variables has a corresponding value. The variable whose value can be retrieved is the named information variable; The variable for which no value could be retrieved is the unnamed information variable, and the degree of missing information of the unnamed information variable is inferred based on the state and cognitive anchoring of the named information variable.

[0012] Optionally, the modulation feature description includes at least the following dimensions: Causal depth feature description, used to adjust the degree of expansion of the medical causal chain in the health education content; Behavioral anchoring feature descriptions are used to adjust the embedding density of behavioral management chain elements in the health education content; Information interweaving feature description, used to modulate the interweaving of causal explanations and behavioral guidance in the health education content; The path initiation feature description is used to adjust the initial unfolding direction of the large language model generation process.

[0013] In a second aspect, there is an interactive device for elderly health management education, the device comprising at least one module, the at least one module being used to perform any of the interactive methods for elderly health management education described in the first aspect.

[0014] Thirdly, a computer device, the computer device including a processor, the processor being configured to execute a computer program stored in a memory to implement the elderly health management education interactive method described in any of the first aspects.

[0015] Fourthly, a computer program product containing instructions, characterized in that, when the instructions are executed by a computer device, the computer device performs the elderly health management education interactive method as described in any of the first aspects.

[0016] The technical solution provided in this application has the following advantages compared with the prior art: One of its beneficial effects and its working principle is as follows: In health education dialogue systems targeting elderly users, the large language model, upon receiving behavioral guidance-based educational instructions, generates content that meets requirements in terms of language simplification and medical accuracy. However, the information arrangement still follows the clinical reasoning chain, progressing from etiology and pathological mechanisms to clinical evidence and recommendations, rather than the behavioral management chain logic required by the educational instructions. During the autoregressive generation process of the large language model, the conditional probability distribution of each token is shaped by the statistical distribution of the training corpus. Health and medical training texts are primarily clinical texts, and the conditional probability paths between medical concepts are fixed to the clinical reasoning chain direction during the training phase. This path locking means that while the health education content received by elderly users is linguistically appropriate, its information organization remains based on clinical professional logic. Elderly users find it difficult to extract operational guidance directly corresponding to their daily behaviors, resulting in insufficient practical enforceability of the educational content.

[0017] To address this issue, one direct approach is to pre-define a mandatory information arrangement path in the prompts, requiring the large language model to organize content in a fixed order from a description of the life scenario to specific operational steps and then to execution precautions. While this superficially alters the paragraph arrangement of the output text, its impact on the knowledge organization path is constrained in two ways. First, the arrangement path instructions are primarily treated as task format requirements by the large language model. The semantic signals carried in these instructions have low information density regarding the direction of knowledge organization, and the offset to the conditional probability distribution at each character position is insufficient to counteract the high-probability paths solidified by large-scale clinical corpus statistics in the training distribution. When medical concept characters are needed within a paragraph to explain the rationale for the action, their appearance activates clinical knowledge clusters and accumulates the traction of path locking. The semantic signals of the format instructions cannot offset this cumulative effect, and the information development direction within the paragraph slides back to the logic of the clinical reasoning chain. Second, any health education content inevitably needs to reference medical concepts to provide causal support for the action. Completely avoiding the clinical reasoning chain would compromise the integrity of the causal information needed for elderly users to make independent judgments in non-standard situations. The contradiction between path locking and causal integrity cannot be resolved by formatting instructions at the prompt word level.

[0018] The technical solution of this application does not directly specify the information arrangement path in the prompts. Instead, before the large language model generates content, it performs educational task feature analysis on the educational instructions and generates a modulation feature description based on the analysis results. This modulation feature description does not specify the exact arrangement path that the large language model must follow; instead, it is injected into the generation context of the large language model in the form of a semantic feature description that centrally represents the direction of knowledge organization offset. The semantic information about the direction of knowledge organization in the modulation feature description is extracted and concentrated from the educational instructions through educational task feature analysis, and its information density is higher than the sparse semantic signal attached to the format instructions. When the large language model generates each subsequent character, the attention mechanism continuously assigns weights to all characters in the preceding context. The high-density semantic signal carried in the modulation feature description continuously participates in the conditional probability calculation at each character position, interacting with the knowledge path encoded in the parameter space. This provides a offset bonus sufficient to counteract the traction force of the clinical reasoning chain path for behavior management paths with lower conditional probabilities. This offset effect extends to the conditional probability distribution at each character position, rather than just affecting the paragraph-level arrangement order.

[0019] The knowledge organization logic of health education content can thus be adjusted between the clinical reasoning chain and the behavior management chain according to the nature of the educational task. Educational content aimed at behavior change starts with life scenarios and unfolds along the main line of operational steps, while educational content aimed at causal understanding can still retain the causal development ability of the clinical reasoning chain. Different parts of the same educational content can be biased towards different organizational logics, and the bias method is dynamically determined by the modulation feature description, maintaining the integrity of causal information for elderly users while avoiding path locking.

[0020] Its second beneficial effect and its working principle are as follows: The aforementioned modulation feature description injects a generation context before generation to offset the knowledge organization path. This offset process requires determining a specific offset direction for each modulation dimension. In actual elderly health education dialogues, the same educational instruction may be addressed to elderly users in different cognitive states. Users encountering the health topic for the first time and needing to establish basic causal cognition differ from users with years of management experience who only need direct behavioral guidance. The educational instruction text does not contain information distinguishing these two types of users. The optimal knowledge organization method is not determined by the educational instruction alone, but by a joint function of the educational instruction and the user's current cognitive state. Under the condition of missing user cognitive state information, if the educational task feature analysis outputs a deterministic offset direction value for each modulation dimension, the output selection is driven not by decision information but by the training distribution bias of the analysis process itself. The modulation configuration is locked to a single point in the configuration space. Under this single-point constraint, only one of the multiple knowledge organization methods originally encoded in the large language model parameter space for different educational topics is activated, while the rest are excluded from the reachable path.

[0021] In the technical solution of this application, the modulation feature description comprises two components: a center offset direction and an allowable domain. The center offset direction is inferred based on information available in the educational instructions, and the allowable domain defines the selectable range of knowledge organization paths with the center offset direction as the anchor point. The width of the allowable domain is determined based on the information missingness identification results; the more key variables in the missing or low-anchored state, the wider the allowable domain; the more key variables in the known state, the narrower the allowable domain. After the modulation feature description is injected into the generation context, the center offset direction provides the basic offset trend for the knowledge organization path of the large language model, and the allowable domain defines the range of paths that the large language model can autonomously choose. Within the range defined by the allowable domain, the large language model autonomously selects a specific knowledge organization path based on the knowledge structure encoded in its own parameter space.

[0022] In educational scenarios where user cognitive state information is lacking, the existence of the permissive domain activates the implicit knowledge about knowledge organization methods in the parameter space of the large language model. The large language model's autonomous path selection within the constrained space leverages its implicit understanding of knowledge organization methods for different educational topics, compensating for the lack of precise specification provided by external analysis due to information deficiency. The synergy between the central direction and the permissive domain ensures the effectiveness of the entire modulation mechanism within the continuously changing range of information conditions. When information is lacking, the permissive domain provides a directional space for autonomous exploration; when information is abundant, the narrow permissive domain provides control precision with a tendency towards determinism.

[0023] Its third beneficial effect and its working principle are as follows: The width of the aforementioned permissible domain is determined based on the information missingness identification results. This identification process requires determining whether the key variables in the educational instructions are known or missing. In actual health education dialogues for the elderly, digital tools make it extremely convenient for elderly users to copy and forward professional information. Users paste paragraphs of health science articles from WeChat groups into the chat box, input test reports via text recognition, and forward medication information found by their children or caregivers. This information from external sources exhibits highly professional and structured textual features. If information missingness identification is based solely on whether variable values ​​can be successfully extracted from the text to determine whether they are known or missing, all variables in the above scenarios can be successfully extracted and marked as known. Their contribution to narrowing the permissible domain is the same as that of variables written by users based on their own cognitive understanding. The textual extractability of variable values ​​reflects the professional level of external information sources, not the user's own cognitive depth. Binary judgment cannot distinguish whether variables are superficially present or deeply anchored in the user's cognition, causing the adjustment direction of the permissible domain width to deviate from the user's actual cognitive needs.

[0024] In the technical solution of this application, for each named information variable in a known state, the cognitive anchoring degree of the variable in the user's cognitive structure is evaluated based on the relationship between the variable value's presentation in user input and the surrounding text. The evaluation of anchoring degree is based on a comprehensive analysis of three features: contextual embedding, spontaneous association between variables, and expression precision matching. Contextual embedding assesses whether the variable value is embedded in the user's own narrative structure; personal experience details and expressions of uncertainty are byproducts of cognitive processing, and their existence indicates that the information has been processed in the user's cognition. Spontaneous association between variables assesses whether the user spontaneously establishes causal or experiential connections with other variables when mentioning the variable; isolated variables suggest that the user is transcribing an external list rather than describing their own understanding of the knowledge structure. Expression precision matching assesses whether the professional precision of the variable value is consistent with the user's overall cognitive expression level in other parts of the dialogue; precision jumps suggest that high-precision segments come from external sources different from the user's own cognition. The contribution of known state variables to the narrowing of the permissible domain is weighted by cognitive anchoring degree; the higher the anchoring degree, the greater the narrowing contribution, and the lower the anchoring degree, the smaller the narrowing contribution.

[0025] The generation of the permissive domain is thus no longer based solely on the number of variables that can be extracted from the text, but rather on the number of variables that are operational in the user's cognition. For elderly users whose input text contains a large amount of externally copied information, the narrowing contribution of their low-anchored known variables is reduced, and the permissive domain is correspondingly widened. The large language model gains accessibility to the organizational path that includes causal explanations, compensating for the user's insufficient understanding of the information's meaning. For elderly users who provide limited but cognitively deep information based solely on their own understanding, their high-anchored known variables maintain a large narrowing contribution, and the permissive domain is narrowed to match the user's cognitive level, without introducing redundant basic explanations. The evaluation of anchoring is completed simultaneously during the variable extraction stage, allowing the permissive domain to be adjusted based on the estimate of cognitive depth during the user's first round of input. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating an application scenario of the interactive health management education method for the elderly provided in this application embodiment; Figure 2 This is one of the flowcharts illustrating the interactive method for elderly health management education provided in this application embodiment; Figure 3 A second flowchart illustrating the interactive method for elderly health management education provided in this application embodiment; Figure 4 This is the third flowchart illustrating the interactive method for elderly health management education provided in this application embodiment. Detailed Implementation

[0027] The technical solutions in this application will now be described with reference to the accompanying drawings.

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0029] Before providing a detailed explanation of the embodiments of this application, let's first introduce the application scenarios involved in the embodiments of this application.

[0030] Figure 1 This is a schematic diagram illustrating an application scenario of the interactive health management education method for the elderly provided in this application. For example... Figure 1 As shown, an interactive health management education device for elderly users is provided in the health management education service, and a large language model is deployed in the interactive device. Elderly users submit educational commands to the interactive device via a terminal device. The interactive device executes the method described in the following embodiments to process the educational commands, generate health education content matching the nature of the educational task, and return the health education content to the terminal device for the elderly user to refer to and perform corresponding health management actions.

[0031] The method of the elderly health management education interaction method provided in this application embodiment can be loaded and executed by an elderly health management education interaction device. The device includes at least one module, which is used to execute the elderly health management education interaction method described in any of the following embodiments.

[0032] Figure 2 This is one of the flowcharts illustrating the interactive method for elderly health management education provided in this application. (Refer to...) Figure 2 As shown in the embodiments of this application, the interactive method for elderly health management education includes the following steps: S201: Receive educational instructions.

[0033] Specifically, in this embodiment, the educational instruction consists of two parts. The first part is a preset system prompt in the health education dialogue device. This system prompt contains task guidance information for health management education for elderly users, instructing the large language model to process subsequent user input with health education as the goal. The second part is a natural language prompt submitted by the elderly user through the terminal device, the content of which is the user's expressed health management-related consultation needs. The system prompt and the user prompt together constitute the educational instruction described in this application. The receipt of the educational instruction can be achieved through a text input interface commonly used in the field, which belongs to human-computer interaction technology known in the field.

[0034] S202: Perform educational task feature analysis on the educational instruction, and generate a modulation feature description based on the analysis. The modulation feature description characterizes the offset requirement of the educational task on the knowledge organization path of health education content.

[0035] Figure 3 This is a second flowchart illustrating the interactive method for elderly health management education provided in this application. (Refer to...) Figure 3 As shown, specifically, in this embodiment of the application, the process of generating modulation feature descriptions based on the analysis in S202 includes: S301: Infer the center offset direction for each modulation dimension. The center offset direction is inferred based on information available in the educational instructions, expressing the most likely offset direction for that dimension.

[0036] Specifically, in this embodiment, the inference of the center offset direction is achieved by guiding the large language model to reason about the educational task nature of the educational instruction through prompt words. The educational instruction text received in S201 is used as input, and the center offset direction for each modulation dimension is output. The large language model performing this inference can be the same model instance as the one used to generate health education content in S204, invoked at a different inference stage, or it can be another independently deployed large language model instance; this application does not limit this. The center offset direction is presented in natural language form, and its content is not a format instruction for the large language model, but a semantic description of the offset direction of the educational task in each dimension of the knowledge organization path. This center offset direction is an intermediate product in the modulation feature description generation process; the final generation of the modulation feature description is completed in subsequent steps.

[0037] The modulation feature description includes at least four dimensions: causal depth, behavioral anchoring, information interweaving, and path initiation. The following is an example of a cue word composition used to guide a large language model in inferring the center offset direction of each dimension. The cue word contains a task instruction part and an educational instruction part to be analyzed. The content of the task instruction part is as follows: "Please perform an educational task characteristic analysis on the following educational instructions and output the center offset direction from the following four dimensions."

[0038] Causal Depth: Analyze whether the health education topic covered by the educational instruction requires the user to develop a causal understanding in order to make autonomous behavioral judgments. If so, the instruction should maintain a high degree of causal chain expansion; if the instruction points to specific daily behavioral norms, the instruction should compress the degree of causal chain expansion.

[0039] Behavior anchoring: Analyze the complexity of the behavior management tasks involved in the educational instruction. If it involves multi-step, multi-time-point behavior management tasks, the instruction should increase the embedding density of the behavior management chain elements; if it involves single-action behavior guidance, the instruction should decrease the embedding density.

[0040] Information Interweaving: Analyze the dependency between the behavioral actions and causal understanding involved in the educational instruction. If the correct execution of the behavior depends on the user's understanding of the causal relationship, the instruction closely interweaves the causal explanation and behavioral guidance within the same paragraph; if the behavior can be performed independently of causal understanding, the instruction separates the two into independent paragraphs.

[0041] Path Initiation: Analyze the core objective type of this educational instruction. If the core objective is causal understanding, set the starting anchor point to a medical concept; if the core objective is behavior change, set the starting anchor point to a description of a real-life scenario.

[0042] Please output the center offset directions of the above four dimensions in natural language. The educational instruction portion to be analyzed is the educational instruction text received by S201. After receiving the above prompt words, the large language model outputs natural language text containing four-dimensional center offset directions through the inference process.

[0043] The judgment logic for each dimension is further explained below.

[0044] Causal depth feature description is used to adjust the degree of expansion of the medical causal chain in the health education content. Specifically, in this embodiment, the large language model analyzes whether the health education topic involved in the educational instruction requires the user to establish a causal understanding in order to make an autonomous behavioral judgment. When the large language model determines that the topic involved in the educational instruction requires the user to understand causal relationships, such as scenarios where the educational instruction involves the user to autonomously adjust the medication time or diet plan based on changes in their own indicators, the causal depth feature description instructs to maintain a high degree of causal chain expansion, so that the generated content includes the causal deduction process from etiology to mechanism to behavioral basis. When the large language model determines that the educational instruction points to specific daily behavioral norms, such as fixed medication time reminders or standardized dietary restrictions, the causal depth feature description instructs to compress the degree of causal chain expansion, so that the generated content focuses on the behavioral operation itself rather than the underlying medical principles.

[0045] Behavioral anchoring feature descriptions are used to adjust the embedding density of behavioral management chain elements in the health education content. Specifically, in this embodiment, behavioral management chain elements include life scenario triggering conditions, specific operation steps, immediate and perceptible execution feedback, and repetitive execution cues. A large language model analyzes the complexity of the behavioral management tasks involved in the educational instructions. When the large language model determines that the educational instruction involves multi-step, multi-time-point behavioral management tasks, such as a blood glucose management process requiring different operations at multiple time points throughout the day, the behavioral anchoring feature description instructs for increasing the embedding density, ensuring that behavioral management chain elements appear frequently in the output text and permeate the entire document. When the large language model determines that the educational instruction involves single-action behavioral guidance, the behavioral anchoring feature description instructs for decreasing the embedding density.

[0046] Information interweaving feature description is used to regulate the interweaving of causal explanations and behavioral guidance in the health education content. Specifically, in the embodiments of this application, the large language model analyzes the dependency relationship between the behavioral operation involved in the educational instruction and the causal understanding. When the large language model determines that the correct execution of the behavior depends on the user's understanding of the causal relationship, for example, when the educational instruction involves a scenario where the user needs to adjust the exercise period according to the blood pressure fluctuation pattern, the correctness of the behavioral operation depends on the user's causal understanding of the relationship between blood pressure and exercise time. The information interweaving feature description instructs that the causal explanation and behavioral guidance be closely interwoven within the same paragraph, so that the user can simultaneously obtain the causal basis of the behavior while reading the behavioral guidance. When the large language model determines that the behavior can be executed independently of the causal understanding, the information interweaving feature description instructs that the causal explanation and behavioral guidance be separated into independent paragraphs.

[0047] The path initiation feature description is used to adjust the initial expansion direction of the large language model generation process. Specifically, in this embodiment, the path initiation feature description determines the expansion direction of the first few characters when the large language model generates health education content. The initial direction of the generation process has a guiding effect on the global path, and the path dependency characteristic of the autoregressive process means that the choice of the initial character affects the expansion direction of the entire subsequent sequence. The large language model analyzes the core objective type of the educational instruction. Educational instructions with causal understanding as the core objective set the starting anchor point as a medical concept, so that the generation process deduces from medical knowledge to the behavioral level; educational instructions with behavior change as the core objective set the starting anchor point as a description of a life scenario, so that the generation process introduces behavioral operation guidance from the user's daily life context.

[0048] The four dimensions described above constitute the content dimensions of the modulation feature description, and the center offset direction of each dimension is the most likely offset direction for that dimension. In this embodiment, the final generation of the modulation feature description also requires determining the allowable domain for each dimension. The allowable domain expresses the acceptable range of offset variation for that dimension, and its width is determined based on the state evaluation of information variables in the user input. The following steps describe the process of determining the allowable domain.

[0049] S302: Determine the permissible domain for each modulation dimension. The permissible domain defines the selectable range of the knowledge organization path for that dimension, anchored by the center offset direction. The width of the permissible domain is determined based on the information missingness identification result. The information missingness identification includes a state assessment of the information variables in the educational instruction. The information variables include named information variables and unnamed information variables. Named information variables are those whose values ​​can be extracted from user input, and unnamed information variables are those whose values ​​cannot be extracted from user input. The more unnamed information variables there are and the greater their influence weight on the corresponding modulation dimension, the wider the permissible domain; the more named information variables there are, the narrower the permissible domain.

[0050] Specifically, in this embodiment, the center offset direction inferred for each modulation dimension in S301 represents the most likely offset direction for that dimension. However, the optimal offset configuration cannot be uniquely determined solely by the information available in the educational instruction text. The optimal knowledge organization method is determined not only by the educational instruction itself but also by the user's current cognitive state. Under the condition of incomplete user cognitive state information, if a deterministic offset direction is specified for each dimension, the modulation configuration will be locked at a single point in the configuration space, excluding the reachability of the large language model to other paths that may be more suitable for the current user.

[0051] The introduction of the permissive domain means that the offset requirement for each modulation dimension is no longer expressed as a deterministic direction, but rather as an optional range anchored to the central offset direction. The permissive domain expands outwards from the central offset direction, which provides the basic offset trend for that dimension. The boundary constraints of the permissive domain ensure that the offset does not deviate from the basic intent of the educational instruction. The width of the permissive domain is determined by the state evaluation results of the information variables. Specifically, the more unnamed information variables in the user input for which no value can be extracted, the greater the uncertainty of the user's cognitive state, and the wider the permissive domain, resulting in a larger autonomous choice space for the large language model in that dimension. Conversely, the more named information variables in the user input for which values ​​can be extracted, the more sufficient the information available to determine the modulation configuration, resulting in a narrower permissive domain and a more deterministic modulation.

[0052] The state assessment process for information variables comprises two levels: the first level involves the extraction and classification of information variables, determining which variables are named information variables and which are unnamed information variables; the second level involves assessing the cognitive anchoring of named information variables, further refining their contribution to the allowable domain width. The following steps describe the assessment process at these two levels.

[0053] Figure 4 This is the third flowchart illustrating the interactive method for elderly health management education provided in this application. (Refer to...) Figure 4 As shown, specifically, in this embodiment of the application, the state evaluation of the information variable includes: S401: Generate a list of expected variables related to the user's current consultation question based on the educational instruction. Search the user's input text to see if each variable in the list of expected variables has a corresponding value. Variables that can be retrieved are marked as named information variables, and variables that cannot be retrieved are marked as unnamed information variables.

[0054] Specifically, in the embodiments of this application, the generation of the expected variable list and the retrieval of values ​​are performed in two stages.

[0055] The first stage involves generating a list of expected variables: using the educational instruction text as input, the large language model is guided by prompts to infer the information variables needed to answer the user's current health management question, outputting a list of expected variables. Each entry in the expected variable list is a variable category rather than a specific value, such as current medication regimen, recent blood pressure data, daily dietary habits, and exercise frequency. Each variable category describes a type of information related to the user's health management question, and its specific value varies from user to user. The following is an example of the prompts used to guide the large language model in generating the expected variable list. The prompts include a task instruction part and an educational instruction part to be analyzed. The content of the task instruction part is as follows: "Please analyze the following user's health management inquiry and deduce the information variables that need to be considered when generating appropriate health education content for this user. The information variables are output as variable categories, each describing a type of user health information that influences the knowledge organization and content depth of the education content, such as current medication regimen, recent blood pressure data, daily dietary habits, exercise frequency, past medical history, and comorbidities. This method involves the following four modulation dimensions: causal depth, behavioral anchoring, information interweaving, and path initiation. Please output a list of expected variable categories and indicate which modulation dimensions each variable category is related to." After receiving the above prompt words, the large language model outputs a list of expected variable categories through the reasoning process. Each entry contains the variable category name and the associated modulation dimension label.

[0056] The second stage is variable value retrieval: using the expected list of variable categories and the historical context of the current dialogue as input, it searches the input text of each variable category in each round of the user's dialogue to see if a corresponding specific value exists. The retrieval process uses attribute matching rules to complete the matching through a large model. That is, each variable category is treated as an attribute name, and the specific content in the user input that semantically matches that attribute is searched. For example, the variable category "current medication regimen" can match the specific drug name and dosage information in the user input (such as amlodipine 5mg, valsartan 80mg), and the variable category "recent blood pressure data" can match the specific measurement value in the user input (such as blood pressure measured in the morning 140 / 90). A variable category may match multiple specific values, and each matched specific value is treated as a named information variable instance under that variable category. Variable categories that can match at least one specific value are marked as named information variables, and variable categories that cannot match any specific value are marked as unnamed information variables. The final output is a list of variables with named or unnamed labels, where the named information variables also carry the matched specific values. The above-mentioned variable value retrieval and classification are routine applications of information extraction techniques in this field.

[0057] The degree of missing information for unnamed variables cannot be directly observed and must be inferred based on the state and cognitive anchoring of named variables. This inference is achieved through the reasoning process of a large language model. Using the list of named variables and their values ​​output by S401, the cognitive anchoring values ​​of each named variable output by S402, and the list of unnamed variables as input, the large language model is guided by prompt words to infer the degree of missing information for each unnamed variable. The following is an example of the prompt word structure used to guide the large language model in performing this inference. The task instruction section is as follows: "Based on the state and cognitive anchoring values ​​of the following named information variables, infer the degree of missing information for each unnamed information variable."

[0058] The inference is based on the following: Each unnamed information variable describes a cognitive state dimension of the user. This dimension cannot be directly extracted from the user's input, but it can be indirectly inferred through the state and anchoring degree of named information variables. Please independently determine the degree of association between each named information variable and each unnamed information variable, and make inferences based on this judgment. If you determine that the anchoring degree of named information variables associated with a certain unnamed information variable is high, it indicates that the user has a knowledge base based on their own understanding in this cognitive dimension, and the degree of lack of that unnamed information variable is low. If you determine that the anchoring degree of associated named information variables is low or most of them are in an unnamed state, it indicates that the user lacks their own understanding in this cognitive dimension, and the degree of lack of that unnamed information variable is high.

[0059] For each unnamed information variable, output a value between 0 and 1 indicating the degree of information missing, where 0 represents that the information for that cognitive state dimension is fully available, and 1 represents that the information for that cognitive state dimension is completely missing. S402: For the named information variable, based on the relationship between the way the variable value is presented in the user input and the surrounding text, evaluate the cognitive anchoring degree of the variable in the user's cognitive structure.

[0060] Specifically, in this embodiment, cognitive anchoring is a continuous numerical value that measures the extent to which the variable value is anchored within the user's own cognitive structure rather than merely floating on the surface of the input text. The evaluation of cognitive anchoring does not analyze the information content of the variable value itself, but rather the relationship between the information content and its presentation.

[0061] The assessment of cognitive anchoring is achieved through the reasoning process of a large language model. Using the historical context of the current dialogue, the user's input text in each round of dialogue, and the list of named information variables marked in S401 as input, the large language model is guided by prompts to analyze each named information variable from three features. The results of the analysis of these three features are then combined to output the cognitive anchoring value for that variable. The introduction of the dialogue historical context allows the large language model to observe the user's overall expression patterns in multi-round dialogues, rather than making judgments based solely on a single round of input. The large language model performing the cognitive anchoring assessment can be the same model instance used in S204 to generate health education content, invoked at different inference stages, or it can be another independently deployed large language model instance; this application does not limit this. The same applies to the large language model performing center offset direction inference in S301.

[0062] The following is an example of the cue word composition used to guide a large language model in performing cognitive anchoring assessment. The cue word includes a task instruction portion, the historical context of the current dialogue, the user's input text in each round of dialogue, and a list of named information variables output by S401.

[0063] The task instructions section contains the following: "Please evaluate the cognitive anchoring degree for each named information variable in the following user input. Cognitive anchoring degree measures the extent to which the variable's value is anchored within the user's own cognitive structure rather than merely existing on the surface of the text. Please analyze from the following three characteristics."

[0064] Contextual embedding method: Whether the variable value is embedded in the user's own narrative structure. If the variable value is wrapped in personal experience details, perceptual descriptions, or expressions of uncertainty, the anchoring degree is high; if the variable value is presented in an independent format and has no trace of personal experience, the anchoring degree is low.

[0065] Spontaneous associations between variables: Whether users spontaneously establish causal or empirical relationships between a variable and other variables when mentioning it. If user-generated associations exist, the anchoring degree is high; if variables are listed in isolation, the anchoring degree is low.

[0066] Accuracy of Expression Match: Whether the accuracy of this variable's expression is consistent with the user's overall level of cognitive expression in other parts of the conversation. If consistent, the anchoring degree is high; if there are significant discrepancies, the anchoring degree is low.

[0067] Based on the above three features, output a cognitive anchoring value for each variable, ranging from 0 to 1. 0 indicates that the variable value is completely superficial and unprocessed by the user, while 1 indicates that the variable value is fully anchored within the user's own cognitive structure. After receiving the above prompt words, the large language model outputs a cognitive anchoring value for each named information variable through the reasoning process.

[0068] The judgment logic for each feature is further explained below.

[0069] The first feature is the contextual embedding of the variable value in the user input. This contextual embedding characterizes whether the variable value is embedded within the user's own narrative structure. The large language model analyzes the presentation of the variable value in the user input: when the variable value is wrapped in details of personal experience, perceptual descriptions, and expressions of uncertainty, it indicates that the information has undergone cognitive processing and has a high degree of anchoring. For example, when a user inputs "I take that little white pill every morning, I think it's called amlodipine," the variable value "amlodipine" is embedded in the user's personal medication experience narrative. When the variable value is presented in an independent, formatted manner, it indicates that the information has not undergone cognitive processing and has a low degree of anchoring. For example, when a user inputs "amlodipine 5mg qd," the variable value exists in a professional abbreviation and standard dosage format without any trace of personal experience.

[0070] The second characteristic is whether there are user-initiated causal or experiential associations between the variable and other variables. The large language model analyzes whether users spontaneously associate the variable with other variables when mentioning it: when users establish causal or temporal associations between the variable and other variables at the level of personal experience, it indicates that the user has a cognitive understanding of the relationship between the variables, with a high degree of anchoring. For example, if a user enters "My blood pressure has been much more stable since I switched to this medication," there is a user-initiated experiential association between the medication variable and the blood pressure variable. When variables are listed in isolation, it suggests that the user is transcribing an external list rather than describing their own understanding of the knowledge structure, with a low degree of anchoring. For example, if a user enters "Amlodipine 5mg, Valsartan 80mg, Metformin 500mg," there is no description of a relationship between the variables.

[0071] The third characteristic is the degree of match between the expressive precision of the variable value and the user's overall cognitive expression level in other parts of the dialogue. The large language model judges the overall expressive precision of all the user's input text in the current dialogue and compares the expressive precision of the variable value with the user's overall level: when the two are consistent, it indicates that the information source is homogeneous and the anchoring degree is high. For example, the user uses everyday colloquial language to express health information, such as "I take the blood pressure medication in the morning" and "My blood pressure is about 130 or 140". The expressive precision of each variable is consistent with the user's overall expression level. When there is a jump in the expressive precision of the variable value and the user's overall level, it indicates that the variable value comes from an external source different from the user's own cognition and the anchoring degree is low. For example, when describing daily symptoms, the user uses vague everyday terms such as "sometimes dizzy" and "my legs are a little swollen", but when describing a medication regimen, they use professional expressions such as "valsartan 80mg bid", which shows a jump in expressive precision.

[0072] Based on a comprehensive analysis of the three features mentioned above, the large language model outputs a cognitive anchoring value for each named information variable. The contribution of each named information variable to the narrowing of the allowable region is obtained by weighting the cognitive anchoring values. The higher the cognitive anchoring value of a named information variable, the greater its contribution to the narrowing of the allowable region; conversely, the lower the cognitive anchoring value of a named information variable, the smaller its contribution to the narrowing of the allowable region.

[0073] S303: Based on the state evaluation results of the information variables, perform information missing degree identification and determine the allowable domain width for each modulation dimension.

[0074] Specifically, in this embodiment, the input for information missingness identification includes three parts: the classification results of named and unnamed information variables output by S401, the cognitive anchoring values ​​of each named information variable output by S402, and the missingness values ​​of each unnamed information variable output by S401. Information missingness identification independently calculates the tolerance range width for each modulation dimension.

[0075] The allowable width is determined through the reasoning process of the large language model. Using the three sets of data mentioned above and the list of modulation dimensions determined in S301 as input, prompt words guide the large language model to determine the allowable width for each modulation dimension. The following is an example of the prompt word structure used to guide the large language model in determining the allowable width. The task instruction section is as follows: "Based on the state evaluation results of the following information variables, determine the allowable domain width for each modulation dimension."

[0076] The input data includes: a list of named information variables and their cognitive anchoring values, and a list of unnamed information variables and their missing values.

[0077] The established rules are as follows: For each modulation dimension, the influence of each information variable on the dimension is determined by comprehensively referring to the dimension relationships marked in S401 and independently assessing the degree of influence of each information variable on the dimension. The narrowing contribution of named information variables to the allowable domain of the dimension is obtained by weighting their cognitive anchoring degree; the higher the anchoring degree, the greater the narrowing contribution, and the lower the anchoring degree, the smaller the narrowing contribution. The broadening contribution of unnamed information variables to the allowable domain of the dimension is determined by their missing value; the higher the missing value, the greater the broadening contribution. Combining the narrowing and broadening contributions of all relevant information variables, an allowable domain width value is output for this dimension.

[0078] Please output a tolerance width value between 0 and 1 for each modulation dimension. 0 indicates that the tolerance is empty, i.e., completely deterministic specification, and 1 indicates the widest tolerance, i.e., the maximum autonomous choice space.

[0079] S304: Generate the modulation feature description based on the center offset direction and permissible domain of each modulation dimension. The modulation feature description is presented in natural language text form, and each dimension includes a description of the center offset direction and permissible domain of that dimension.

[0080] Specifically, in this embodiment, the generation of modulation feature descriptions is achieved through the inference process of a large language model. Using the center offset directions of each dimension output by S301 and the allowable domain width values ​​of each dimension output by S303 as input, prompt words guide the large language model to generate natural language text containing center direction and allowable domain descriptions for each dimension. The following is an example of the prompt word composition used to guide the large language model in generating modulation feature descriptions. The task instruction section is as follows: "Based on the center offset direction and allowable domain width values ​​for each of the following modulation dimensions, generate a modulation feature description for each dimension. The allowable domain width value ranges from 0 to 1, where 0 indicates an empty allowable domain (i.e., a completely deterministic specification), and 1 indicates the widest allowable domain (i.e., the maximum autonomous choice space). The description for each dimension should include: a description of the center offset direction for that dimension, and a description of the range within which the offset direction can vary under the current allowable domain width. The closer the allowable domain width value is to 1, the wider the range of variation described; the closer the allowable domain width value is to 0, the narrower the range of variation described, until it approaches a deterministic specification." After receiving the aforementioned prompt words, the large language model outputs natural language text containing modulation feature descriptions of various dimensions through an inference process. This text is the modulation feature description described in this application, and is the final output of "generating modulation feature descriptions based on the analysis" in S202.

[0081] The following is an example of the output text for a modulation feature description. Assuming the user submits the educational instruction "I have high blood pressure, what should I eat?", the center offset direction of each dimension analyzed in S301 is biased towards the behavior management chain, and the allowable domain width values ​​determined in S303 are: causal depth 0.6, behavior anchoring 0.3, information interweaving 0.5, and path initiation 0.1, then the output text of the modulation feature description would look like this: "Causal depth: The main direction is to compress the causal chain, but a brief description of the blood pressure regulation mechanism directly related to dietary behavior can be retained in appropriate places. The degree of expansion can be adjusted between compression and moderate retention."

[0082] Behavioral anchoring: The direction is based on a higher embedding density of behavioral management chain elements, and the embedding density can be slightly adjusted around this direction.

[0083] Information Interweaving: The main direction is to closely interweave causal explanations and behavioral guidance within the same paragraph, but the two can be appropriately separated in some paragraphs. The degree of interweaving can be adjusted between close interweaving and moderate separation.

[0084] Starting point: Use a description of a real-life scenario as the initial anchor, but if necessary, you can delve into medical concepts closely related to that scenario. In the output text above, the description of each dimension includes a center offset direction (with ... as the primary direction) and a tolerance description (which can be adjusted between ...). Dimensions with larger tolerance width values ​​(such as causal depth 0.6) describe a wider range of variability, dimensions with smaller tolerance width values ​​(such as behavior anchoring 0.3) describe a narrower range of variability, and dimensions with tolerance width values ​​close to 0 (such as path start 0.1) tend to have a deterministic specification.

[0085] S203: Inject the modulation feature description into the generation context of the large language model.

[0086] Specifically, in the embodiments of this application, the modulation feature description output by S304 exists in the form of natural language text, which together with the original educational instructions constitutes the generation context of the large language model.

[0087] The injection method involves placing the modulated feature description text in a preceding position in the generation context, so that it participates in the attention weight calculation as a preceding context when the large language model starts generating its first character. The injection process does not modify the model parameters of the large language model, adjust the sampling strategy, or change hyperparameters such as the inference temperature or sampling threshold; it only affects the conditional probability distribution in the subsequent generation process by changing the text content of the generation context. This text injection operation falls within the scope of common prompting engineering techniques in this field.

[0088] S204: The large language model generates the health education content based on the generation context.

[0089] Specifically, in this embodiment, after receiving a generation context containing modulation feature descriptions and educational instructions, the large language model generates health education content character by character in an autoregressive manner. When generating each character, the large language model's attention mechanism assigns weights to all preceding characters in the generation context, and the semantic signal about the knowledge organization direction carried in the modulation feature description continuously participates in the conditional probability calculation at each character position.

[0090] Each dimension of the modulation feature description contains two semantic signals: a center offset direction and an admissibility domain. The semantic signal of the center offset direction provides a basic offset trend for the path selection of the large language model at each character position, causing the conditional probability distribution to deviate from the default clinical inference chain direction in the training distribution. The semantic signal of the admissibility domain defines the acceptable range for the large language model's path selection. Within this range, the large language model autonomously selects a specific knowledge organization path based on the knowledge structure encoded in its own parameter space. The autonomous selection process does not require additional search or optimization steps. The character selection of the large language model in autoregressive generation is guided within the constraint domain by the feasible path with the highest conditional probability in the parameter space. This path reflects the implicit judgment of the large language model based on its own training regarding the most appropriate knowledge organization method for the current educational topic.

[0091] The center offset direction and the allowable domain work together during the generation process. The center offset direction ensures that the generation process does not slip back onto the default clinical inference chain path in the training distribution. The allowable domain ensures that the path selection of the large language model does not deviate from the basic intent of the educational instructions, while preserving the flexibility of the large language model to adapt to different users' cognitive states using its own knowledge structure. When the allowable domain width is large, the large language model has a larger autonomous choice space, and the generated health education content reflects more of the implicit judgment of the large language model in terms of knowledge organization. When the allowable domain width is small, the autonomous choice space of the large language model is limited, and the generated health education content is closer to the deterministic configuration specified by the center offset direction in terms of knowledge organization.

[0092] Health education content generated by the large language model is returned to elderly users via terminal devices.

[0093] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0094] In the various embodiments of this application, unless otherwise specified or logically conflicting, the terminology and / or descriptions between different embodiments are consistent and can be referenced mutually. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. In the textual description of the embodiments of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. In this application, "first," "second," and various numerical designations are only for ease of description and are not used to limit the scope of the embodiments of this application. For example, they are used to distinguish different messages, rather than to describe a specific order or sequence.

[0095] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application. The order of the process numbers does not imply the order of execution; the execution order of each process should be determined by its function and internal logic.

[0096] Finally, it should be noted that the above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An interactive method for elderly health management education, characterized in that, The interactive method for elderly health management education includes the following steps: Receive educational instructions; The educational instructions are subjected to educational task feature analysis, and a modulation feature description is generated based on the analysis. The modulation feature description characterizes the offset requirement of the educational task on the knowledge organization path of health education content. The modulation feature description is injected into the generation context of the large language model; The large language model generates the health education content based on the generated context.

2. The interactive method for elderly health management education according to claim 1, characterized in that, The modulation feature description includes a center offset direction and an allowable domain. The center offset direction is inferred based on information available in the educational instructions, and the allowable domain defines the selectable range of knowledge organization paths with the center offset direction as the anchor point. The large language model autonomously selects a knowledge organization path based on the knowledge structure encoded in its own parameter space within the scope defined by the permissible domain.

3. The interactive method for elderly health management education according to claim 2, characterized in that, The width of the allowable domain is determined based on the information missing degree identification result; The information missing degree identification includes: The information variables in the educational instructions are evaluated for their status. The information variables include named information variables and unnamed information variables. The named information variables are those whose values ​​can be extracted from user input, and the unnamed information variables are those whose values ​​cannot be extracted from user input. The more unnamed information variables there are and the greater their influence weight on the corresponding modulation dimension, the wider the allowable domain will be. The more named information variables there are, the narrower the allowed domain becomes.

4. The interactive method for elderly health management education according to claim 3, characterized in that, The method further includes: Extract named information variables that are in a known state from the educational instructions, and evaluate the cognitive anchoring degree of the variable in the user's cognitive structure based on the relationship between the way the variable value is presented in the user input and the surrounding text. The contribution of the named information variable to the narrowing of the allowable region is obtained by weighting the cognitive anchoring degree. The higher the cognitive anchoring degree of the named information variable, the greater its contribution to the narrowing of the allowable region. The lower the cognitive anchoring degree of the named information variable, the smaller its contribution to the narrowing of the allowable region.

5. The interactive method for elderly health management education according to claim 4, characterized in that, The cognitive anchoring degree is determined based on a comprehensive analysis of the following characteristics: The contextual embedding method of the variable value in the user input, wherein the contextual embedding method represents whether the variable value is embedded in the user's own narrative structure; Does this variable have a user-initiated causal or empirical relationship with other variables? The precision of this variable's representation matches the user's overall level of cognitive expression in other parts of the conversation.

6. The interactive method for elderly health management education according to claim 4, characterized in that, The information variables are distinguished into named information variables and unnamed information variables in the following ways: Based on the educational instructions, a list of expected variables related to the user's current consultation question is generated, and the user's input text is searched to see if each variable in the list of expected variables has a corresponding value. The variable whose value can be retrieved is the named information variable; The variable for which no value could be retrieved is the unnamed information variable, and the degree of missing information of the unnamed information variable is inferred based on the state and cognitive anchoring of the named information variable.

7. The interactive method for elderly health management education according to claim 1, characterized in that, The modulation feature description includes at least the following dimensions: Causal depth feature description, used to adjust the degree of expansion of the medical causal chain in the health education content; Behavioral anchoring feature descriptions are used to adjust the embedding density of behavioral management chain elements in the health education content; Information interweaving feature description, used to modulate the interweaving of causal explanations and behavioral guidance in the health education content; The path initiation feature description is used to adjust the initial unfolding direction of the large language model generation process.

8. An interactive device for elderly health management education, characterized in that, The device includes at least one module for performing the elderly health management education interactive method according to any one of claims 1-7.

9. A computer device, characterized in that, The computer device includes a processor for executing a computer program stored in a memory to implement the elderly health management education interactive method according to any one of claims 1-7.

10. A computer program product containing instructions, characterized in that, When the instructions are executed by a computer device, the computer device performs the elderly health management education interactive method as described in any one of claims 1-7.

Images