Ageing-suitable multimedia data generation method and system based on medical treatment and transboundary fusion

Abstract

The invention is suitable for the technical field of artificial intelligence, and provides a medical and cross-boundary fusion-based multimedia data generation method and system suitable for aging, and the method comprises the steps: constructing and training an adaptation degree evaluation model which takes health knowledge features and cross-boundary content features as input and takes the adaptation degree between the health knowledge features and the cross-boundary content features as output; obtaining target health knowledge data and a candidate cross-border content data set, and determining an adaptation degree predicted value of each candidate cross-border content data in the candidate cross-border content data set based on the adaptation degree evaluation model; and determining target cross-border content data based on the adaptation degree predicted value of each candidate cross-border content data, and generating aging-suitable multimedia data based on the target health knowledge data, the target cross-border content data and an adaptation degree evaluation model. According to the technical scheme provided by the invention, the accuracy of health knowledge and cross-border content suitability evaluation is improved, the quality of the multimedia data suitable for aging and the matching degree with the elderly user are improved, and the transmission effect of the multimedia data suitable for aging is improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a method and system for generating age-friendly multimedia data based on medical and cross-disciplinary integration. Background Technology

[0002] With the acceleration of population aging, the health management needs of the elderly are increasing. As an important carrier for improving the health literacy of the elderly, the quality and suitability of health science popularization content directly affect the effectiveness of health knowledge transmission.

[0003] Traditional health education content production technologies often adopt a "knowledge-oriented" one-way output model, with content design focusing on the completeness of medical knowledge. For example, standardized science texts are generated by matching medical knowledge bases with preset templates, but the differences in interests among different age groups are not considered. This results in a monotonous format of science content, making it difficult to attract the attention of elderly users and leading to low efficiency in the effective delivery of health knowledge.

[0004] To enhance the appeal of health education content, some fields have attempted to develop "medical + cross-disciplinary" content. However, the creation of such content relies heavily on the creators' subjective experience and lacks a systematic adaptation evaluation mechanism. This makes it difficult to balance the accuracy of medical knowledge with the engaging nature of the content, severely impacting the user's viewing experience and knowledge absorption. Furthermore, existing solutions for "age-friendly" health education content only optimize superficial parameters such as visual and auditory elements, like simply enlarging the font, increasing the volume, and adding speech speed prompts. They fail to consider core needs such as the cultural background (e.g., familiarity with traditional opera) and cognitive habits (e.g., narrative rhythm preferences) of the elderly. As a result, while the generated content may possess basic age-friendly attributes, its appeal remains insufficient, failing to effectively increase elderly users' willingness to actively watch. In addition, the evaluation of existing health education content mainly relies on macro-level dissemination indicators such as play counts, likes, and shares, making it impossible to accurately quantify and evaluate key quality indicators such as "accuracy of knowledge delivery," "naturalness of narrative integration," and "age-friendliness," leading to poor age-friendly dissemination effects.

[0005] Therefore, improving the accuracy of age-friendly multimedia data suitability assessment to enhance the effectiveness of age-friendly communication has become an urgent problem to be solved. Summary of the Invention

[0006] This application provides a method and system for generating age-friendly multimedia data based on the integration of medical and cross-disciplinary technologies, aiming to improve the accuracy of age-friendly multimedia data suitability assessment and enhance the effectiveness of age-friendly communication.

[0007] In a first aspect, this application provides a method for generating age-friendly multimedia data based on the integration of medical and cross-disciplinary technologies, the method comprising: S1. Construct and train an adaptation evaluation model that takes health knowledge features and cross-border content features as inputs and the degree of adaptation between the two as output. S2, acquire target health knowledge data and candidate cross-border content data set, and determine the predicted fit value of each candidate cross-border content data in the candidate cross-border content data set based on the fit evaluation model. S3. Based on the predicted fit values ​​of each candidate cross-border content data, determine the target cross-border content data, and generate age-friendly multimedia data based on the target health knowledge data, the target cross-border content data, and the fit evaluation model.

[0008] In one possible implementation, step S2 includes: S21, Based on the target health knowledge data, extract the target health knowledge features; S22, Based on the candidate cross-border content data set, extract the candidate cross-border content features of each candidate cross-border content data; S23, the target health knowledge features and each of the candidate cross-border content features are input into the fit evaluation model for processing, and the fit prediction value corresponding to each of the candidate cross-border content data is determined.

[0009] In one possible implementation, step S21 includes: S211, extract the medical entity features, knowledge point complexity features and knowledge attribute features of the target health knowledge data. The medical entity features are used to indicate the medical entity vectors in the health knowledge data. The knowledge point complexity features are used to indicate the rarity of medical entities in the health knowledge data. The knowledge attribute features are used to indicate the knowledge level and applicable scenarios of the health knowledge data. S212, the medical entity features, the knowledge point complexity features, and the knowledge attribute features are concatenated and projected onto the shared space to obtain the target health knowledge features.

[0010] In one possible implementation, step S22 includes: S221, determine the cross-domain to which each of the candidate cross-domain content data belongs, including any one or more of history, opera and cooking; S222, based on the feature extraction method corresponding to the cross-border domain to which each candidate cross-border content data belongs, extract the sub-features corresponding to each candidate cross-border content data, and after feature splicing each sub-feature, project it to the shared space to obtain the candidate cross-border content features corresponding to each candidate cross-border content data. Among them, the sub-features in the field of history include any one or more of the following: dynastic features, features of personal relationships, and features of cultural symbols; the sub-features in the field of opera include any one or more of the following: genre classification features, vocal style and emotional features, and repertoire features; and the sub-features in the field of cooking include any one or more of the following: ingredient attribute features, cooking scene features, and solar term association features.

[0011] In one possible implementation, step S23, the processing of the fitness evaluation model, includes: S231, Input the target health knowledge feature and the candidate cross-border content feature to be tested, wherein the candidate cross-border content feature to be tested is any of the candidate cross-border content features; S232, Based on the cross-attention mechanism, bidirectional feature interaction is performed on the target health knowledge features and the cross-border content features to be tested to obtain fused attention-enhanced features; S233, Perform bidirectional temporal feature fusion on the fused attention enhancement features to extract narrative temporal features; S234, Based on the fused attention-enhanced features, extract global features; S235, Dynamically fuse the narrative temporal features and the global features to determine the final fused features; S236, Based on the final fusion features, a prediction is made to obtain the predicted fit between the target health knowledge features and the cross-border content features to be tested.

[0012] In one possible implementation, step S233 includes: S2331, Perform the first bidirectional temporal feature fusion on the fused attention enhancement features to obtain the first positive hidden state sequence and the first negative hidden state sequence, and splice them together to obtain the shallow temporal feature sequence; S2332, perform a second bidirectional temporal feature fusion on the shallow temporal feature sequence to obtain a second forward hidden state sequence and a second reverse hidden state sequence, and concatenate the last hidden state in the second forward hidden state sequence and the first hidden state in the second reverse hidden state sequence to obtain the narrative temporal feature.

[0013] In one possible implementation, step S3 includes: S31, based on the adaptation prediction value of each candidate cross-border content data, determine the candidate cross-border content data with the highest preset number of adaptation prediction values ​​as the target cross-border content data; S32, Input the target health knowledge data and all the target cross-border content data into the fit evaluation model for processing to obtain the fusion fit; S33, Based on the target health knowledge data, the target cross-border content data, and the fusion adaptation degree, generate the age-friendly multimedia data.

[0014] In one possible implementation, step S33 includes: S331. Based on the target health knowledge data, the cross-border content data of each target, and the preset age-appropriate parameters, determine the structured outline. The preset age-appropriate parameters include any one or more of the preset speech rate, preset font size, and preset color contrast. S332, Based on the structured outline and the fusion adaptability, generate the age-friendly multimedia data.

[0015] In one possible implementation, step S1 includes: S11, The fitness evaluation model is constructed based on a hybrid Transformer and a bidirectional gated loop unit; S12, acquire several health knowledge sample data and cross-border content sample data, and construct several health knowledge-cross-border content sample pairs; S13, perform fit labeling on each of the sample pairs to construct a training dataset with health knowledge sample data, cross-border content sample data and fit as the core dimensions. The fit includes a total score and four components. The total score is the comprehensive fit score. The components are the knowledge transfer accuracy score, interest matching score, narrative integration naturalness score and dissemination age-appropriateness score. S14. Based on the training dataset, the fitness evaluation model is trained using a loss function. The loss function includes a weighted sum of component losses and a regularization term. The component losses are used to characterize the mean squared error loss between the predicted value and the corresponding labeled value of each of the four components.

[0016] On the other hand, this application also provides a computer system including a memory and a processor; the memory stores program code executable by the processor; the program code is used to perform the method as described in the first aspect or any of the implementations thereof.

[0017] The beneficial effects of this application compared to existing technologies are as follows: It constructs and trains a fit evaluation model that takes health knowledge features and cross-border content features as input and the fit between them as output; it acquires a target health knowledge data set and a candidate cross-border content data set; based on the fit evaluation model, it determines the predicted fit value of each candidate cross-border content data set; based on the predicted fit values ​​of each candidate cross-border content data set, it determines the target cross-border content data; and based on the target health knowledge data, the target cross-border content data, and the fit evaluation model, it generates age-friendly multimedia data. By using the fit evaluation model, the fit between the target health knowledge data and each candidate cross-border content data is quantitatively evaluated, improving the accuracy of the fit evaluation. Furthermore, based on this fit, the target cross-border content data is selected from the candidate cross-border content data set, and combined with the target health knowledge data to generate age-friendly multimedia data. Compared to relying on the creator's subjective experience to generate age-friendly multimedia data, this effectively improves the quality of age-friendly multimedia data and its matching degree with elderly users, thereby enhancing the effect of age-friendly dissemination.

[0018] It is understood that the computer system provided in this application has the same beneficial effects as the above-mentioned method for generating age-friendly multimedia data based on medical and cross-disciplinary integration, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an age-friendly multimedia data generation method based on medical and cross-disciplinary integration, provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an adaptation evaluation model provided in an embodiment of this application. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of this application. However, those skilled in the art will understand that this application may be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0023] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] For ease of understanding, the technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating a method for generating age-friendly multimedia data based on medical and cross-disciplinary integration, as provided in an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The method provided in this embodiment includes the following steps: S1. Construct and train an adaptation evaluation model that takes health knowledge features and cross-border content features as inputs and the degree of adaptation between the two as outputs.

[0029] In one possible implementation, step S1 may include: S11, based on a hybrid Transformer and a bidirectional gated loop unit, constructs an adaptation evaluation model.

[0030] As an example, such as Figure 2 As shown, the fitness evaluation model includes an input unit, a cross-attention fusion unit, a bidirectional gated recurrent coding unit, a Transformer coding unit, a feature integration unit, and an output unit.

[0031] Specifically, the input of the cross-attention fusion unit is connected to the input unit, the input of the bidirectional gated cyclic coding unit and the Transformer coding unit is connected to the output of the cross-attention fusion unit, the input of the feature integration unit is connected to the output of the bidirectional gated cyclic coding unit and the Transformer coding unit, and the output unit is connected to the feature integration unit.

[0032] Preferably, the bidirectional gated cyclic coding unit adopts a stacked bidirectional structure, including a first bidirectional gated cyclic coding layer and a second bidirectional gated cyclic coding layer, each of which independently completes the bidirectional gated cyclic coding operation.

[0033] Preferably, the output unit includes one or more regression heads, with each regression head used to predict different fit components. For example, when the output unit includes one classification head, it predicts the comprehensive fit score between health knowledge features and cross-border content features; or, when the output unit includes four classification heads, it predicts the knowledge transfer accuracy score, interest matching score, narrative fusion naturalness score, and age-appropriateness score between health knowledge features and cross-border content features, respectively, and can further perform a weighted summation of the above four scores to obtain the comprehensive fit score.

[0034] S12, acquire several health knowledge sample data and cross-border content sample data, and construct several health knowledge-cross-border content sample pairs.

[0035] As an example, we obtain several health knowledge sample data h. i (such as "Autumn Moisturizing", "Dietary Taboos for Hypertension", etc.) and sample data of several cross-disciplinary contents f i (e.g., Chapter 45 of *Dream of the Red Chamber*, excerpts from the Peking Opera *The Drunken Beauty*, Mid-Autumn Festival mooncake making tutorials, etc.), constructing 1200 sample pairs (h i ,f i ).

[0036] S13. Adaptability labeling is performed on each sample pair to construct a training dataset with health knowledge sample data, cross-border content sample data and adaptability as the core dimensions. Adaptability includes a total score and four components; the total score is the comprehensive adaptability score; the components are the knowledge transfer accuracy score, interest matching score, narrative integration naturalness score, and dissemination age-appropriateness score.

[0037] Optionally, each sample pair can be labeled with its suitability. If a sample is suitable for elderly users, it can be labeled as a suitability score. If it is not suitable for the elderly, it will be marked as suitable. To obtain the training dataset .

[0038] Preferably, the labeled dataset is formed by weighting and summing the following scores for each sample: accuracy of knowledge transfer (A), interest matching (I), naturalness of narrative integration (N), and age-friendliness of dissemination (C), as well as the comprehensive fit score (S) obtained by weighting and summing A, I, N, and C. .

[0039] For example, the comprehensive adaptation score is a core indicator for measuring the integration effect of health knowledge and cross-border content. It is defined as a weighted sum of four dimensions, and the formula for calculating the comprehensive adaptation score is as follows: ; in, The score represents the accuracy of knowledge delivery, with a preferred weight α of 0.3, used to quantify the completeness and accessibility of health knowledge points. The score represents the interest matching degree, with a preferred weight β of 0.3, used to quantify the degree of fit between cross-border content and the preferences of elderly users; The narrative integration naturalness score is represented by a weight γ, which is preferably 0.25, and is used to quantify the closeness of the plot connection between health knowledge points and cross-border content. The score represents the age-friendliness of the content, with a preferred weight δ of 0.15, used to quantify the suitability of the content format for the elderly population; α+β+γ+δ=1, and the weights are determined by the Analytic Hierarchy Process (AHP).

[0040] For example, knowledge transfer accuracy score Based on "completeness of knowledge point retention" And the degree of popularization of medical terminology Weighted composition, calculated based on health knowledge data. ;in, Core medical entities refer to entities that are crucial to understanding health knowledge points (such as "lung yin" and "dryness evil" in "autumn dryness injures the lungs"). A simplified definition is "understanding rate ≥ 80% among users aged 60 and above".

[0041] For example, the interest matching score is calculated as follows: ;in, This represents the familiarity score of elderly users with cross-domain fields, queried through a pre-constructed elderly interest graph (e.g., familiarity with "Dream of the Red Chamber" is set to 9.2 / 10, and familiarity with modern street dance is set to 3.5 / 10, normalized to [0,30]). This indicates the fit between content format preferences, such as preferences for opera excerpts or historical stories. It is calculated using the cosine similarity between the user profile vector and the content format vector. ; For the preference vector of elderly users, For content-form feature vectors, .

[0042] For example, the narrative integration naturalness score is calculated as follows: ;in, The relevance score is calculated using the pointwise mutual information (PMI) of co-occurring concept pairs between health knowledge data and cross-domain content data. , This is a concept within health knowledge data. For cross-boundary content, M represents the corresponding concepts, and M represents the number of related concept logarithms. ; The transition naturalness score is measured by the rate of change of the hidden states in the Long Short-Term Memory (LSTM) network of the narrative chain. , The smaller the value, the more natural the transition. Normalized to [0, 25 × 0.4].

[0043] For example, the communication age-friendliness score integrates age-friendliness characteristics from three aspects: language, visual, and cultural. The calculation formula for the communication age-friendliness score is as follows: ;in, To indicate the age-appropriateness of language rhythm, the deviation between the speech rate (words / minute) and the speech rate adapted to older users is quantified: , The optimal speaking speed is 120 words / minute (the average speaking speed that elderly users are comfortable with). The age-friendliness of visual elements is indicated by factors such as font size (normalized to [0,1]) and color contrast (1 point for conforming to the industry standard (WCAG 2.1AA) in Web Accessibility International Standard 2.1, otherwise 0 points). =15 × (Font score × 0.5 + Contrast score × 0.5) × 0.3; The age-friendliness of cultural symbols is indicated by the frequency of occurrence of symbols familiar to the elderly, such as traditional festivals and classic opera excerpts. .

[0044] In addition, when labeling the fit of each sample pair, a labeling team can be formed. This team can consist of 3 geriatric medicine experts (A rating), 3 gerontology scholars (I rating), 3 science popularization creators (N rating), and 3 age-friendly design experts (C rating). Each expert independently scores the sample pair, and the intraclass correlation coefficient (ICC) is calculated to test consistency. ;in, The variance of the true difference in the sample. For the variance of raters, Let be the variance of the random error. (Selected) Scores greater than 0.85 (significant consistency) are used as annotation information for sample pairs.

[0045] S14. Based on the training dataset, a fitness evaluation model is trained using a loss function. The loss function includes a weighted sum of the component losses and a regularization term. The component losses are used to characterize the mean squared error loss between the predicted value of each component and the corresponding labeled value of the component.

[0046] As an example, the training parameters used in the training process are as follows: the training dataset is divided into a training set (840 samples, 70%), a validation set (180 samples, 15%), and a test set (180 samples, 15%). AdamW is used as the optimizer, and the initial learning rate is... The system employs a linear warm-up of 1000 steps followed by cosine decay; an early stopping strategy is adopted, where the training stops if the mean absolute error (MAE) of the validation set does not improve for 10 consecutive epochs; the batch size is 16, and the gradient clipping threshold is 1.0.

[0047] For example, the loss function is calculated as follows: ;in, , , and Let S represent the mean squared error loss (MSE) corresponding to the knowledge transfer accuracy score, interest matching score, narrative integration naturalness score, and age-appropriateness score. The formula for calculating the mean squared error loss is: , Indicates the labeled value. Indicates the predicted value. For all parameters of the model, The L2 regularization coefficients, α, β, γ, and δ, are the same as those used in the formula for calculating the overall fit score.

[0048] S2, acquire the target health knowledge data and candidate cross-border content data set, and determine the predicted fit value of each candidate cross-border content data in the candidate cross-border content data set based on the fit evaluation model.

[0049] For example, a candidate cross-border content dataset (such as "Dream of the Red Chamber", Peking Opera, Mid-Autumn Festival cooking, etc.).

[0050] In one possible implementation, step S2 may include: S21. Based on the target health knowledge data, extract the target health knowledge features.

[0051] As an example, step S21 may optionally include: S211, extract the medical entity features, knowledge point complexity features, and knowledge attribute features of the target health knowledge data. The medical entity features are used to indicate the medical entity vectors in the health knowledge data, the knowledge point complexity features are used to indicate the rarity of the medical entities in the health knowledge data, and the knowledge attribute features are used to indicate the knowledge level and applicable scenarios of the health knowledge data.

[0052] Optionally, BioBERT can be used to extract medical entity vectors from the target health knowledge data, with each entity represented as... Take the average vector of the top-5 core entities. As a feature vector of medical entities.

[0053] Optional, knowledge point complexity features , ,in For entities The rarity (ranging from 0 to 1, representing the reciprocal of the frequency of occurrence in the medical dictionary), where M is the total number of entities.

[0054] Optional, knowledge attribute features It adopts 1-hot encoding for knowledge levels (including two dimensions: basic and advanced) and applicable scenarios (including three dimensions: prevention, conditioning, and treatment).

[0055] S212 involves concatenating the medical entity features, knowledge point complexity features, and knowledge attribute features, and then projecting them onto a shared space to obtain the target health knowledge features.

[0056] Optionally, the aforementioned medical entity features, knowledge point complexity features, and knowledge attribute features are concatenated into a 775-dimensional (768+2+5) structure, and then projected onto a 512-dimensional shared space through a linear layer to obtain the target health knowledge features. ,in , These are the parameters for the projection layer.

[0057] S22, Based on the candidate cross-border content data set, extract the candidate cross-border content features of each candidate cross-border content data.

[0058] As an example, step S22 may optionally include: S221, determine the cross-domain to which each candidate cross-domain content data belongs, including any one or more of history, opera and cooking.

[0059] It should be noted that the cross-disciplinary fields, including history, opera and cooking, are just one example. The cross-disciplinary fields in this application embodiment may also include calligraphy, folk music and other fields, which will not be listed here.

[0060] S222: Based on the feature extraction method corresponding to the cross-border domain to which each candidate cross-border content data belongs, extract the sub-features corresponding to each candidate cross-border content data, and after feature splicing, project the sub-features onto the shared space to obtain the candidate cross-border content features corresponding to each candidate cross-border content data.

[0061] Among them, the sub-features in the field of history include any one or more of the following: dynastic features, features of personal relationships, and features of cultural symbols; the sub-features in the field of opera include any one or more of the following: genre classification features, vocal style and emotional features, and repertoire features; and the sub-features in the field of cooking include any one or more of the following: ingredient attribute features, cooking scene features, and solar term association features.

[0062] For example, in the historical domain, the feature extraction method involves extracting sub-features such as dynastic features, interpersonal relationship features, and cultural symbol features, and then concatenating these sub-features before projecting them into a shared space. Among these, dynastic features... This includes dynasty embedding vectors (preferably generated through a pre-trained historical knowledge model) and normalized time features, including the starting year. Year of termination and duration The formula for calculating the characteristics of a dynasty is: , .

[0063] Among them, the characteristics of character relationships The preferred method is to extract the relationship from the person-relationship graph using the GraphSAGE model, and the calculation formula is as follows: , This is a graph of relationships between historical figures (nodes represent people, and edges represent relationship types).

[0064] Among them, cultural symbol characteristics Preferably, cultural symbols (such as "Mid-Autumn Festival" and "Imperial Examination") in historical texts are encoded using Bidirectional Encoder Representations from Transformers (BERT). The formula for calculating cultural symbol features is as follows: (Description of cultural symbols).

[0065] Furthermore, the sub-features such as dynastic characteristics, interpersonal relationship characteristics, and cultural symbol characteristics are spliced ​​together and projected onto a shared space to obtain cross-disciplinary content features in the field of history. .

[0066] For example, in the field of traditional Chinese opera, the feature extraction method involves extracting sub-features such as opera genre classification features, vocal style and emotional features, and repertoire features. These sub-features are then concatenated and projected into a shared space. Among these, the opera genre classification features... The preferred encoding method is 1-hot encoding, which includes 10 major opera genres such as Peking Opera, Yue Opera, and Huangmei Opera.

[0067] Among them, the emotional characteristics of singing style First, candidate cross-disciplinary content data belonging to the opera field is preprocessed by extracting Mel-Frequency Cepstral Coefficients (MFCC) sequences and spectrogram sequences, which are then concatenated to form a high-dimensional temporal feature sequence. This high-dimensional temporal feature sequence is then input into an LSTM network, where its internal gating mechanisms (forget gate, input gate, output gate) are used to model the long-distance temporal dependencies. Finally, the hidden state of the LSTM network at the last time step is extracted as the vocal emotion feature of the candidate cross-disciplinary content data, calculated using the following formula: .

[0068] Among them, the characteristics of the play The preferred encoding method is BERT-encoded text describing the plot of the play, and the calculation formula is: (Synopsis).

[0069] Furthermore, the sub-features such as the above-mentioned genre classification characteristics, vocal style emotional characteristics, and repertoire characteristics are spliced ​​together and projected onto a shared space to obtain cross-border content features in the field of opera. .

[0070] For example, the feature extraction method in the culinary field is to extract the characteristics of ingredients, cooking scene, and solar term, and then project the sub-features into a shared space after feature splicing.

[0071] Among them, the food attribute features can be selected as a food attribute matrix. The options include the nutritional properties of 20 core ingredients (such as calories, vitamin content, etc.) and 1-hot ingredient types (including grains, fruits and vegetables, meat, etc.).

[0072] Among them, cooking scene characteristics It is preferred to use 1-hot encoded cooking scenarios, with 20 categories of cooking scenarios available, including home cooking, banquets, and festivals.

[0073] Among them, the characteristics related to solar terms It includes 24 solar term correlation vectors (with values ​​ranging from 0 to 1) and 368-dimensional cooking knowledge embeddings (preferably generated by pre-trained cooking BERT).

[0074] Furthermore, the sub-features, such as the ingredient attribute matrix, cooking scene features, and solar term correlation features, are concatenated and projected onto a shared space to obtain cross-disciplinary content features in the culinary field. .

[0075] S23, input the target health knowledge features and the features of each candidate cross-border content into the fit evaluation model for processing, and determine the fit prediction value corresponding to each candidate cross-border content data.

[0076] In one possible implementation, combining Figure 2 The fitness assessment model shown, in step S23, the processing procedure of the fitness assessment model may optionally include: S231, Input the target health knowledge features and the candidate cross-border content features to be tested. The candidate cross-border content features to be tested can be any candidate cross-border content features.

[0077] Optionally, the target health knowledge features and candidate cross-boundary content features to be tested can be input through the input unit. Cross-border content features of candidates to be tested .

[0078] S232, based on the cross-attention mechanism, performs bidirectional feature interaction between the target health knowledge features and the cross-border content features of the candidate to be tested, and obtains the fused attention-enhanced features.

[0079] Optionally, the cross-attention fusion unit first uses the target health knowledge features as the query vector (Q), and the candidate cross-border content features to be tested as the key vector (K) and value vector (V). Calculate the positive attention weights from health knowledge to cross-disciplinary content. and first attention enhancement features Similarly, the cross-attention fusion unit uses the cross-boundary content features of the candidate to be tested as the query vector (Q'), and the target health knowledge features as the key vector (K') and value vector (V'), to calculate the reverse attention weight from the cross-boundary content to the health knowledge. Second attention enhancement features ,in For attention parameters, For feature dimension, , Finally, the first attention enhancement feature and the second attention enhancement feature are fused together to obtain the fused attention enhancement feature. .

[0080] S233 performs bidirectional temporal feature fusion on the attention-enhanced features to extract narrative temporal features.

[0081] As an example, step S233 may optionally include: S2331, the first bidirectional temporal feature fusion is performed on the attention enhancement features to obtain the first positive hidden state sequence and the first negative hidden state sequence, and then they are spliced ​​together to obtain the shallow temporal feature sequence.

[0082] S2332, perform a second bidirectional temporal feature fusion on the shallow temporal feature sequence to obtain the second forward hidden state sequence and the second reverse hidden state sequence, and concatenate the last hidden state in the second forward hidden state sequence and the first hidden state in the second reverse hidden state sequence to obtain the narrative temporal feature.

[0083] Optionally, step S2331 is executed by the first bidirectional gated cyclic coding layer in the bidirectional gated cyclic coding unit, and step S2332 is executed by the second bidirectional gated cyclic coding layer.

[0084] For example, the first bidirectional gated recurrent coding layer fuses the attention-enhanced feature z of the input. att Split temporal features [z] att,1 ,z att,2 ,...,z att,T Simultaneously, a forward-gated loop is executed (from t=1 to t=T). And a reverse-gated loop (from t=T to t=1) Output the first positive hidden state sequence [h] 1,1 h 1,2 , ..., h 1,T ] and the first reverse hidden state sequence [h ’ 1,T , ..., h ’ 1,2 h ’ 1,1The shallow temporal feature sequence is concatenated and input into the second bidirectional gated recurrent encoding layer. The forward and reverse gated loops are executed again to capture deeper temporal dependencies, resulting in the second forward hidden state sequence [h]. 2,1 h 2,2 , ..., h 2,T ] and the second reverse hidden state sequence [h ’ 2,T , ..., h ’ 2,2 h ’ 2,1 ], the last hidden state h in the second forward hidden state sequence 2,T The first hidden state h' of the second reverse hidden state sequence 2,T The narrative temporal features are obtained by piecing them together. The first bidirectional gated cyclic coding layer and the second bidirectional gated cyclic coding layer each have 256 hidden units, and the narrative temporal feature obtained after bidirectional splicing is 512-dimensional.

[0085] S234 extracts global features based on fused attention-enhanced features.

[0086] Optionally, the Transformer coding unit incorporates attention-enhanced features based on the input. The global structure is modeled using a 4-layer Transformer encoder (8-head attention). The [CLS] token output from the last layer is taken as the global feature.

[0087] S235, based on narrative temporal features and global features, dynamically fuses them to determine the final fused features.

[0088] Optionally, the feature integration unit employs a gating mechanism to dynamically fuse narrative temporal features. (512-dimensional) and global features (1024 dimensions): ; ; in, , ; This indicates that the global features are truncated in dimensionality, and the first 512 dimensions are taken and combined with the narrative temporal features. Alignment; The final fused features are matched with the input dimensions of the subsequent output layer.

[0089] S236. Based on the final fusion features, a prediction is made to obtain the fit prediction value between health knowledge features and cross-border content features to be tested.

[0090] Optionally, the output unit is based on the final fused features. The predicted values ​​of knowledge transfer accuracy scores were determined using four regression heads. Interest matching score prediction value Narrative integration naturalness score prediction value and the predicted value of age-friendliness score for communication ,in This is a linear layer with a 512-dimensional input and a 1-dimensional output. Combined with the comprehensive fit score calculation formula, the predicted fit value is calculated. .

[0091] S3. Based on the predicted fit values ​​of each candidate cross-border content data, determine the target cross-border content data, and generate age-friendly multimedia data based on the target health knowledge data, the target cross-border content data, and the fit evaluation model.

[0092] In one possible implementation, step S3 may include: S31, based on the adaptation prediction value of each candidate cross-border content data, determine the candidate cross-border content data with the highest preset number of adaptation prediction values ​​as the target cross-border content data.

[0093] Optionally, if the preset quantity is 3, then the 3 candidate cross-border content data with the highest fit prediction value are selected as the target cross-border content data. .

[0094] S32, input the target health knowledge data and all target cross-border content data into the fit evaluation model for processing, and obtain the fusion fit.

[0095] Optionally, the target health knowledge features of the target health knowledge data and the candidate cross-border content features corresponding to the three target cross-border content data are simultaneously input into the fit evaluation model for processing to obtain the fusion fit between the target health knowledge data and the three target cross-border content data, including the fusion predicted value of knowledge transfer accuracy score, the fusion predicted value of interest matching score, the fusion predicted value of narrative fusion naturalness score, the fusion predicted value of dissemination age-appropriateness score, and the fusion predicted value of comprehensive fit score.

[0096] It should be noted that the processing procedure for the fit evaluation model can be referred to in step S23, and will not be repeated here.

[0097] S33 generates age-friendly multimedia data based on target health knowledge data, cross-border content data of various targets, and integration and adaptation.

[0098] As an example, step S33 may optionally include: S331. Based on target health knowledge data, cross-border content data of each target, and preset age-appropriate parameters, determine the structured outline. The preset age-appropriate parameters include any one or more of preset speech rate, preset font size, and preset color contrast.

[0099] Optionally, a multimodal creation model can be constructed based on GPT-4V, and aging-appropriate parameters can be set and normalized to [0,1] to obtain preset aging-appropriate parameters. , To preset the speaking speed, Preset font size, This indicates the preset color contrast. It combines the target health knowledge data H and the cross-cutting content data from the three targets. The system inputs preset age-appropriate parameters into a multimodal creation model to generate a plot outline and knowledge point distribution, and outputs a structured outline. .

[0100] S332 generates age-appropriate multimedia data based on structured outlines and fusion adaptability.

[0101] Optionally, a structured outline can be used. The knowledge transfer accuracy score and fusion prediction value in the fusion adaptation degree Combined with interest matching score to predict value Input the multimodal creation model, fill in visual elements (such as suggestions for opera costumes) and language style (such as age-appropriate speech rate annotations) to generate details, and output the final age-appropriate multimedia data. Among them, the knowledge transfer accuracy score is a fusion prediction value. Combined with interest matching score to predict value Used to guide the generation process, such as At a lower level, enhance the expression of knowledge. To make it more appealing, add fun elements and age-friendly multimedia data, including videos.

[0102] Furthermore, after generating age-friendly multimedia data, a pre-trained scoring model can be used to score the age-friendly multimedia data to obtain a quality score. If the quality score is less than a preset threshold (such as 0.7), it indicates that the quality of the age-friendly multimedia data is low and it needs to be regenerated.

[0103] In addition, after generating age-friendly multimedia data, input-output consistency verification can be performed to verify whether the manual annotation and the fit assessment model are completely consistent at the input and output levels. For example, input consistency includes that both use "health knowledge + cross-border content" as input and the feature extraction methods are completely corresponding; output consistency includes that the A, I, N, C predicted by the model are completely consistent with the definitions of the indicators in the manual annotation, and the comprehensive fit score S is calculated in the same way.

[0104] In summary, the technical solution provided in this application has at least the following advantages: 1. Pioneering Adaptability Quantitative Evaluation System: Transforms subjective creative problems into calculable multimodal adaptation problems, achieving quantitative control that balances knowledge accuracy and fun, and solving the limitations of existing technologies that rely on experience-based creation.

[0105] 2. Deep semantic fusion of multimodal features: Through cross-attention bidirectional interaction and HybridTransformer-GRU architecture, the medical entity features of health knowledge and multimodal features from cross-domain fields (such as opera audio and historical dynasties) are mapped to a shared semantic space, effectively solving the "modal gap" problem.

[0106] 3. Structured modeling of age-friendly features: The cultural background and cognitive characteristics of elderly users are transformed into computable feature vectors, realizing a technological leap from "superficial age-friendly" to "deep age-friendly", effectively improving the age-friendly optimization effect.

[0107] 4. Data-driven closed-loop generation mechanism: The fitness evaluation model is trained using manually labeled sample datasets to replace traditional experience-based judgment; combined with A / B testing framework and online learning, the generation strategy is dynamically optimized, and the stability of content quality is significantly improved.

[0108] 5. Synergistic optimization of knowledge accuracy and dissemination effectiveness: Accurate delivery of medical knowledge is ensured through knowledge entity attention calculation; at the same time, the acceptance rate of elderly users is improved through interest matching quantification. Experiments show that the completion rate is significantly higher than that of traditional popular science content.

[0109] Based on the above technical solution, the following experimental verification was conducted: 1. Experimental Design (1) Dataset: 1200 labeled samples (including health knowledge-cross-border content sample pairs and corresponding fit labeling information, sourced from the health knowledge base of tertiary hospitals and the elderly culture database), divided into a training set of 840 labeled samples, a validation set of 180 labeled samples, and a test set of 180 labeled samples.

[0110] (2) Comparison Model Baseline 1: Traditional template matching method (without adaptability evaluation); Baseline 2: Single-modal fusion model (text features only, no Cross-Attention); Baseline 3: Ordinary Transformer model (time series modeling without bidirectional gated loops).

[0111] (3) Evaluation indicators Main indicator: Mean Absolute Error (The lower the better); Auxiliary metrics: Pearson correlation coefficient ρ (the correlation between model predictions and human annotations, the higher the better), Top-3 domain accuracy (the proportion of correctly selected high-fit domains).

[0112] 2. Experimental Results The experimental results are shown in Table 1: Table 1 Experimental results for each model

[0113] 3. Verification of the age-appropriate effect A controlled experiment was conducted with 100 elderly users (aged 60-75) who watched traditional science videos and compared them with the cross-disciplinary videos generated by this application. The results are as follows: Completion rate: 32% for traditional videos, 87% for videos submitted in this application; Knowledge mastery rate (questionnaire test): 45% for traditional videos, 78% for videos in this application; Subjective satisfaction (out of 5): Traditional video 2.3 points, this application video 4.6 points.

[0114] 4. Conclusion The fit assessment model provided in this application is significantly superior to existing technologies in terms of fit assessment accuracy and the effectiveness of age-friendly dissemination, verifying the feasibility and superiority of the solution.

[0115] On the other hand, this application also provides a computer system, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute any of the above-mentioned methods for generating age-friendly multimedia data based on medical and cross-disciplinary integration.

[0116] For example, program code can be divided into one or more modules / units, which are stored in memory and executed by a processor to complete the present invention. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in a computer system.

[0117] The computer system can be a desktop computer, laptop, handheld computer, or cloud server, and may optionally include, but is not limited to, processors and memory. Those skilled in the art will understand that it may also include input / output devices, network access devices, buses, etc.

[0118] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0119] Memory can be an internal storage unit of a terminal device, such as a hard drive or RAM. Memory can also be an external storage device of a terminal device, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal and external storage units of a computer system. Memory is used to store program code and other programs and data required by the computer system. Memory can also be used to temporarily store data that has been output or will be output.

[0120] The computer system described above is created based on the above-mentioned method for generating age-friendly multimedia data based on medical and cross-disciplinary integration. Its technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0121] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the protection scope of this invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for generating age-friendly multimedia data based on the integration of medicine and cross-disciplinary fields, characterized in that, The method includes: S1. Construct and train an adaptation evaluation model that takes health knowledge features and cross-border content features as inputs and the degree of adaptation between the two as output. S2, acquire target health knowledge data and candidate cross-border content data set, and determine the predicted fit value of each candidate cross-border content data in the candidate cross-border content data set based on the fit evaluation model. S3. Based on the predicted fit values ​​of each candidate cross-border content data, determine the target cross-border content data, and generate age-friendly multimedia data based on the target health knowledge data, the target cross-border content data, and the fit evaluation model.

2. The method according to claim 1, characterized in that, Step S2 includes: S21, Based on the target health knowledge data, extract the target health knowledge features; S22, Based on the candidate cross-border content data set, extract the candidate cross-border content features of each candidate cross-border content data; S23, the target health knowledge features and each of the candidate cross-border content features are input into the fit evaluation model for processing, and the fit prediction value corresponding to each of the candidate cross-border content data is determined.

3. The method according to claim 2, characterized in that, Step S21 includes: S211, extract the medical entity features, knowledge point complexity features and knowledge attribute features of the target health knowledge data. The medical entity features are used to indicate the medical entity vectors in the health knowledge data. The knowledge point complexity features are used to indicate the rarity of medical entities in the health knowledge data. The knowledge attribute features are used to indicate the knowledge level and applicable scenarios of the health knowledge data. S212, the medical entity features, the knowledge point complexity features, and the knowledge attribute features are concatenated and projected onto the shared space to obtain the target health knowledge features.

4. The method according to claim 2, characterized in that, Step S22 includes: S221, determine the cross-domain to which each of the candidate cross-domain content data belongs, including any one or more of history, opera and cooking; S222, based on the feature extraction method corresponding to the cross-border domain to which each candidate cross-border content data belongs, extract the sub-features corresponding to each candidate cross-border content data, and after feature splicing each sub-feature, project it to the shared space to obtain the candidate cross-border content features corresponding to each candidate cross-border content data. Among them, the sub-features in the field of history include any one or more of the following: dynastic features, features of personal relationships, and features of cultural symbols; the sub-features in the field of opera include any one or more of the following: genre classification features, vocal style and emotional features, and repertoire features; and the sub-features in the field of cooking include any one or more of the following: ingredient attribute features, cooking scene features, and solar term association features.

5. The method according to claim 2, characterized in that, In step S23, the processing procedure of the fitness evaluation model includes: S231, Input the target health knowledge feature and the candidate cross-border content feature to be tested, wherein the candidate cross-border content feature to be tested is any of the candidate cross-border content features; S232, Based on the cross-attention mechanism, bidirectional feature interaction is performed on the target health knowledge features and the cross-border content features to be tested to obtain fused attention-enhanced features; S233, Perform bidirectional temporal feature fusion on the fused attention enhancement features to extract narrative temporal features; S234, Based on the fused attention-enhanced features, extract global features; S235, Dynamically fuse the narrative temporal features and the global features to determine the final fused features; S236, Based on the final fusion features, a prediction is made to obtain the predicted fit between the target health knowledge features and the cross-border content features to be tested.

6. The method according to claim 5, characterized in that, Step S233 includes: S2331, Perform the first bidirectional temporal feature fusion on the fused attention enhancement features to obtain the first positive hidden state sequence and the first negative hidden state sequence, and splice them together to obtain the shallow temporal feature sequence; S2332, perform a second bidirectional temporal feature fusion on the shallow temporal feature sequence to obtain a second forward hidden state sequence and a second reverse hidden state sequence, and concatenate the last hidden state in the second forward hidden state sequence and the first hidden state in the second reverse hidden state sequence to obtain the narrative temporal feature.

7. The method according to claim 1, characterized in that, Step S3 includes: S31, based on the adaptation prediction value of each candidate cross-border content data, determine the candidate cross-border content data with the highest preset number of adaptation prediction values ​​as the target cross-border content data; S32, Input the target health knowledge data and all the target cross-border content data into the fit evaluation model for processing to obtain the fusion fit; S33, Based on the target health knowledge data, the target cross-border content data, and the fusion adaptation degree, generate the age-friendly multimedia data.

8. The method according to claim 7, characterized in that, Step S33 includes: S331. Based on the target health knowledge data, the cross-border content data of each target, and the preset age-appropriate parameters, determine the structured outline. The preset age-appropriate parameters include any one or more of the preset speech rate, preset font size, and preset color contrast. S332, Based on the structured outline and the fusion adaptability, generate the age-friendly multimedia data.

9. The method according to any one of claims 1 to 8, characterized in that, Step S1 includes: S11, The fitness evaluation model is constructed based on a hybrid Transformer and a bidirectional gated loop unit; S12, acquire several health knowledge sample data and cross-border content sample data, and construct several health knowledge-cross-border content sample pairs; S13, perform fit labeling on each of the sample pairs to construct a training dataset with health knowledge sample data, cross-border content sample data and fit as the core dimensions. The fit includes a total score and four components. The total score is the comprehensive fit score. The components are the knowledge transfer accuracy score, interest matching score, narrative integration naturalness score and dissemination age-appropriateness score. S14. Based on the training dataset, the fitness evaluation model is trained using a loss function. The loss function includes a weighted sum of component losses and a regularization term. The component losses are used to characterize the mean squared error loss between the predicted value and the corresponding labeled value of each of the four components.

10. A computer system, characterized in that, It includes a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to perform the method of any one of claims 1 to 9.

Images