A health care product recommendation method and system based on RMFEP model

Abstract

The present application provides a health care product recommendation method and system based on the RMFEP model. Obtain a number of first research and analysis networks, each of which includes a first component cluster with a consistent composition, and the first research and analysis network also includes a first component structure outside the first component cluster, integrate the first component clusters included in each first research and analysis network to obtain a first integrated cluster, combine the first integrated cluster and the first component structure included in each first research and analysis network to obtain a first integrated network, adjust the first integrated network to obtain a first neural network, and when the first neural network meets the integration stop requirements, use the first neural network as a multi-dimensional research and analysis network. The present application integrates different networks to obtain a first integrated network, adjusts the first integrated network to obtain a multi-dimensional research and analysis network, and the accuracy and reliability of the multi-dimensional research and analysis network are improved.

Description

Technical Field

[0001] The present application relates to the fields of machine learning and text processing, and more specifically, to a method and system for recommending health care products based on the RMFEP model. Background Art

[0002] With the improvement of people's living standards and the increasing attention to health, the health care industry has developed rapidly. As the core part of the health care industry, health care products are becoming increasingly diverse, covering a variety of areas from hardware facilities to service products. In order to meet the diverse needs of consumers for health care products and help companies better develop, promote and optimize health care products, it is particularly important to conduct a comprehensive and in-depth analysis and evaluation of health care products.

[0003] The analysis of health care products involves multiple key dimensions, such as resource dimension (including the various types of resources that the product relies on), market dimension (such as market demand, competitive situation, etc.), functional dimension (specific functional characteristics of the product), experience dimension (consumer experience of using the product) and product dimension (the attributes and characteristics of the product itself). Traditional analysis methods often focus on a single dimension and lack the ability to conduct comprehensive multi-dimensional analysis of health care products.

[0004] In the field of text analysis, existing technical means usually design independent analysis networks for specific dimensions or specific tasks. For example, there may be a set of special network models for resource analysis, and another set of different models for market analysis. Although these independent models can achieve certain analysis results in their respective fields, they have many limitations. There is a lack of effective integration and coordination between independent analysis networks. When it is necessary to conduct multi-dimensional analysis of health care products, it is often necessary to run multiple different models separately, which not only increases the computing cost and time cost, but also may cause inconsistency and contradiction in the analysis results due to the differences between different models, making it difficult to form a unified and comprehensive analysis conclusion. Independent analysis networks lack flexibility and adaptability when facing new analysis tasks or data. Since each model is trained and optimized for a specific dimension, when new dimension requirements arise or data distribution changes, new models need to be redesigned and trained, which not only consumes a lot of manpower, material resources and time, but also may cause the previous model resources to be ineffectively utilized.

[0005] Existing analysis networks also have certain deficiencies in accuracy and reliability. Due to the relatively limited training data and optimization objectives of a single model, it may not be possible to fully learn the inherent associations and complex relationships between different dimensions, thus affecting the accuracy and credibility of the analysis results. In practical applications, inaccurate analysis results, such as misleading consumer choices, are also not conducive to the improvement and optimization of products by enterprises. Summary of the invention

[0006] The purpose of the present invention is to provide a health care product recommendation method and system based on the RMFEP model. This application is implemented as follows:

[0007] In a first aspect, the present application provides a health care product recommendation method based on the RMFEP model, the method comprising: obtaining a target health care product research text; using a multi-dimensional research and analysis network to perform text analysis of one or more target dimensions on the target health care product research text, and obtaining text analysis results of each target dimension.

[0008] The process of acquiring the multi-dimensional survey and analysis network includes: acquiring a plurality of first survey and analysis networks, any one of which is configured to perform a text analysis of a corresponding dimension on a target health product survey text, each of which includes a first component cluster with a consistent composition, and any one of which also includes a first component structure outside the first component cluster; wherein the dimension includes a resource dimension, a market dimension, a function dimension, an experience dimension, and a product dimension; integrating the first component clusters included in each of the first survey and analysis networks to obtain a first integrated cluster; wherein the first integrated cluster and the first component cluster have a consistent composition, and the configuration parameter of the first integrated cluster is determined by using the configuration parameter of the first component cluster included in each of the first survey and analysis networks; combining the first integrated cluster and the first component structure included in each of the first survey and analysis networks to obtain a first integrated network; adjusting the first integrated network to obtain a first neural network; when the first neural network meets the integration stop requirement, using the first neural network as a multi-dimensional survey and analysis network, the multi-dimensional survey and analysis network is configured to perform a text analysis of one or more target dimensions in the text analysis of the dimensions corresponding to each of the first survey and analysis networks on the target health product survey text.

[0009] In a second aspect, the present application provides a computer system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0010] In the embodiment of the present application, each first survey and analysis network includes a first component cluster with a consistent composition. The first component clusters included in each first survey and analysis network are integrated to obtain a first integrated cluster, and the first integrated cluster and the first component structure outside the first component cluster in each first survey and analysis network are combined to complete the integration of component clusters with a consistent composition in different networks, and the remaining network structures of different networks are combined to integrate different networks to obtain a first integrated network. The first integrated network includes the performance of each first survey and analysis network, and the application of the first integrated network is more extensive, and it can be analyzed in multiple analysis dimensions such as resources, market, function, experience and product (ie, RMFEP mode) of smart health care products. The multi-dimensional survey and analysis network is obtained by adjusting the first integrated network, which not only increases the accuracy of the multi-dimensional survey and analysis network, improves the reliability of dimensional analysis using the multi-dimensional survey and analysis network, but also can maintain the multi-dimensional survey and analysis network at a relatively low cost to achieve text analysis in multiple dimensions.

[0011] Other features will be described in part in the following description. Those skilled in the art will discover these features in part upon inspection of the following content and drawings, or may learn these features through production or use. The features of the present application may be implemented and obtained by practicing or using various aspects of the methods, tools, and combinations listed in the detailed examples described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0013] Figure 1 It is a flow chart of a health care product recommendation method based on the RMFEP model provided in an embodiment of the present application.

[0014] Figure 2 It is a flowchart of the process of acquiring a multi-dimensional research and analysis network provided in an embodiment of the present application.

[0015] Figure 3 It is a schematic diagram of the composition of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method part of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0017] In the embodiment of the present application, the executor of the health care product recommendation method based on the RMFEP model is a computer system, including but not limited to servers, personal computers, laptops, tablet computers, smart phones, etc.

[0018] like Figure 1 As shown, the health care product recommendation method based on the RMFEP model provided in the embodiment of the present application includes the following steps:

[0019] Step S1: Obtain a target health care product research text, where the target health care product research text records the content of the target health care product in at least one dimension of resource dimension, market dimension, function dimension, experience dimension or product dimension.

[0020] The target health care product research text is a record of relevant information about a specific health care product, including descriptions and related information of the health care product from different dimensions. By analyzing these texts, we can gain an in-depth understanding of the characteristics, advantages, and market demand of health care products, thereby providing a basis for accurately recommending appropriate health care products. For example, for a hot spring health care product, the research text includes information such as the distribution of hot spring resources (resource dimension), the popularity of the product in the market and the situation of competitors (market dimension), the therapeutic function of the hot spring (functional dimension), consumers' feelings and evaluations during the hot spring experience (experience dimension), and information such as the product's packaging and pricing (product dimension).

[0021] The acquisition of target health care product research texts is carried out under the permission of laws and regulations, such as based on web crawler technology, data collection interface, manual entry and collation, etc.

[0022] Step S2: Use a multi-dimensional survey analysis network to perform text analysis on the target health care product survey text in one or more target dimensions to obtain text analysis results for each target dimension.

[0023] The multi-dimensional survey and analysis network is a neural network model that has been specifically designed and trained. Its purpose is to conduct in-depth analysis of the target health product survey text and to mine valuable information in different dimensions. It can comprehensively consider factors such as resource dimension, market dimension, function dimension, experience dimension and product dimension, so as to fully and accurately understand the characteristics and market demand of health products. For example, for a new type of health tourism product, the multi-dimensional survey and analysis network can analyze the richness of its tourism resources (resource dimension), its competitiveness in the tourism market (market dimension), whether the health functions provided meet user needs (function dimension), tourists' actual experience feedback (experience dimension) and the product's own pricing, service content, etc. (product dimension), providing a basis for product optimization and recommendation.

[0024] The target dimension is one or more dimensions selected from the resource dimension, market dimension, functional dimension, experience dimension and product dimension according to specific analysis needs and application scenarios. For example, when evaluating the marketing strategy of a health food, you may pay more attention to the market dimension and product dimension to understand the positioning of the food in the market and its own characteristics; when studying the user satisfaction of a health equipment, the experience dimension and functional dimension may be the target dimensions that are focused on. The computer system can determine the target dimension according to preset rules or user input. For example, the user inputs the dimension information to be analyzed through the interface, and the computer system determines the target dimension according to the user input; or the computer system automatically determines the target dimension according to the pre-set logic based on the type of product and the purpose of analysis.

[0025] When the computer system uses a multi-dimensional survey analysis network to perform text analysis on the target health care product survey text, it can use general natural language processing technology (NLP), for example, pre-processing the target health care product survey text, including lexical analysis, syntactic analysis, part-of-speech tagging, etc. For example, for a user comment about a health care tourism product, "The natural scenery here is beautiful and the service is very thoughtful", the computer system decomposes the text into individual words through lexical analysis, such as "natural scenery", "beauty", "service", "thoughtful", etc.; then the part of speech of each word is determined through part-of-speech tagging, such as "natural scenery" as a noun, "beauty" as an adjective, etc. These pre-processing operations help the computer system better understand the semantic information of the text. Word vector representation In order for the computer system to better process text data, it is necessary to convert the words in the text into vector form, that is, word vector. Word vectors can represent words in a semantic space so that words with similar semantics are closer in the vector space. Exemplary word vector models include Word2Vec, GloVe, etc. For example, the word vectors of "hot spring" and "bath" obtained through Word2Vec model training are close in the semantic space because they have a strong semantic association. Computer systems can measure the semantic relationship between words by calculating the similarity between word vectors.

[0026] After preprocessing and word vector representation, the computer system extracts representative features from the text. These features can be word frequency, part of speech, word vector, etc. For example, when analyzing the functional dimensions of health care products, the computer system can extract keywords related to product functions as features. For example, for a smart health care bracelet, keywords such as "heart rate monitoring" and "sleep monitoring" are important functional features. Through feature extraction, text data can be converted into numerical data that can be processed by the computer system, providing a basis for subsequent analysis and modeling.

[0027] Deep learning model Multi-dimensional survey analysis network can use deep learning model for text analysis. Deep learning model has powerful automatic learning and feature representation capabilities, and can learn complex patterns and rules from large amounts of data. Exemplary deep learning models include convolutional neural network (CNN), recurrent neural network (RNN) and its variants long short-term memory network (LSTM), gated recurrent unit (GRU), etc. For example, when analyzing the user experience dimension of health care products, the computer system can use LSTM network to process the user's comment sequence, because LSTM network can effectively capture the long-term dependencies in the sequence data, so as to better understand the user's experience and feelings at different time points.

[0028] In the embodiment of the present application, the focus is on the acquisition of a multi-dimensional research and analysis network, that is, how to determine a low-cost multi-dimensional research and analysis network that can integrate the functions of multiple analysis dimensions. For details, please refer to Figure 2 , the present application embodiment adopts the following steps to obtain:

[0029] Step S10: Obtain a plurality of first research and analysis networks, wherein any first research and analysis network is configured to perform a text analysis of a corresponding dimension on the research text of the target health care product, each first research and analysis network includes a first component cluster with a consistent composition, and any first research and analysis network also includes a first component structure outside the first component cluster; wherein the dimensions include resource dimension, market dimension, function dimension, experience dimension and product dimension.

[0030] The first research and analysis network is a neural network model that performs text analysis on specific dimensions in health care product research texts. Its function is to deeply explore and analyze the relevant information of the target health care product in a specific dimension. For example, the first research and analysis network for the resource dimension has the main task of analyzing the various resources that the target health care product relies on, such as land resources, water resources, biological resources, etc.; the first research and analysis network for the market dimension focuses on analyzing the performance of the product in the market, including market share, competitive situation, consumer demand, etc. Each first research and analysis network focuses on the analysis of a specific dimension to ensure accurate understanding and grasp of the information in that dimension.

[0031] The first component cluster is a consistent part of each first survey and analysis network, which contains a series of basic components for text analysis. These components work together to realize functions such as processing and feature extraction of text data. For example, the first component cluster may contain word vector generation components, text encoding components, etc. The word vector generation component is responsible for converting words in the text into vector form so that the computer system can perform numerical calculations and analysis. The text encoding component further encodes the word vectors to extract more representative features. To illustrate with a simple example, for a description text about a health tourism product "There are rich hot spring resources here and the surrounding environment is beautiful", the word vector generation component will convert words such as "hot spring resources", "surrounding environment" and "beautiful" into corresponding word vectors, and the text encoding component will encode these word vectors according to certain rules to extract features related to the resource dimension.

[0032] The first component structure is the part of the first survey and analysis network other than the first component cluster. It is customized for the analysis needs of specific dimensions to achieve more in-depth and professional text analysis. For example, for the first survey and analysis network in the resource dimension, its first component structure may include a resource classification module and a resource evaluation module. The resource classification module is used to classify different types of resources, such as natural resources into water resources, land resources, etc.; the resource evaluation module evaluates the quality and quantity of various types of resources according to certain standards. Taking a health farm as an example, the resource classification module will classify the land, water sources, crops and other resources in the farm, and the resource evaluation module will evaluate these resources according to indicators such as soil fertility, water quality, and crop yields, thereby obtaining a comprehensive analysis result of the health farm in the resource dimension.

[0033] When the computer system obtains the first survey and analysis network, it can collect a large amount of text data related to health care products, which covers various dimensions such as resources, markets, functions, experiences and products. For example, crawl text data such as user comments, industry reports, news information and other text data of various health care products from the Internet. Then, preprocess these data, including data cleaning, word segmentation, part-of-speech tagging and other operations. Data cleaning is to remove noise and invalid information in the data, such as removing duplicate data, special characters, etc.; word segmentation is to divide the text into words for subsequent analysis; part-of-speech tagging is to mark the part of speech for each word, such as noun, verb, adjective, etc. For example, for a user comment "This health care product has powerful functions and a good user experience", after word segmentation processing, "this", "health care product", "function", "very", "powerful", "user experience", "also", "very good" are obtained. After part-of-speech tagging, "health care product" is marked as a noun, and "powerful" is marked as an adjective, etc.

[0034] Model architecture design According to the analysis requirements of different dimensions, the computer system designs the corresponding model architecture of the first survey analysis network. The design of the model architecture needs to take into account the characteristics of text data and the complexity of the analysis task. For example, for the analysis of the functional dimension, since it is necessary to focus on the various specific functions and characteristics of the product, a convolutional neural network (CNN) architecture with strong feature extraction capabilities is used; for the analysis of the experience dimension, since user experience often has time series, a recurrent neural network (RNN) or its variant long short-term memory network (LSTM) architecture is used. The design of the model architecture needs to be adjusted and optimized according to the specific situation to improve the performance and accuracy of the model.

[0035] Model training and optimization After determining the model architecture, the computer system uses the collected preprocessed data to train the first survey analysis network. During the training process, the model learns the patterns and regularities in the text data and adjusts its own parameters to minimize the difference between the predicted results and the true labels. Commonly used training algorithms include stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, Adam, etc. During the training process, the model also needs to be optimized to avoid overfitting and improve the generalization ability of the model. Exemplary optimization methods include adding regularization terms, data enhancement, early stopping, etc. For example, adding an L2 regularization term can add a square sum of parameters to the loss function to penalize excessive parameter values, thereby preventing the model from overfitting.

[0036] Model evaluation and verification After training is completed, the computer system evaluates and verifies the first survey analysis network to ensure the performance and accuracy of the model. Common evaluation indicators include accuracy, recall, F1 value, etc. By analyzing the evaluation results of the model on the validation set, if the performance of the model does not meet the requirements, it is necessary to adjust the model parameters or redesign the model architecture until the model achieves satisfactory performance.

[0037] Step S20: Integrate the first component clusters included in each first survey and analysis network to obtain a first integrated cluster; wherein the first integrated cluster and the first component cluster have the same composition, and the configuration parameters of the first integrated cluster are determined by using the configuration parameters of the first component clusters included in each first survey and analysis network.

[0038] The first component cluster contains basic components for functions such as text processing and feature extraction. However, since each first research and analysis network analyzes different dimensions, there may be problems of information isolation and resource waste when used alone. By integrating these first component clusters, the computer system can build a more unified and efficient first integration cluster to achieve information sharing and collaborative processing, and improve the comprehensive analysis capabilities of the target health product research text. For example, when analyzing a health tourism product, the first component cluster in the first research and analysis network of the resource dimension may focus on extracting features related to natural resources, and the first component cluster of the market dimension may pay more attention to features such as market demand and competitive situation. After integrating these component clusters, the first integration cluster can comprehensively consider the characteristics of multiple dimensions and more comprehensively understand the characteristics and market value of health products.

[0039] The first integrated cluster is consistent with the first component cluster, that is, the component types and structures contained in the first integrated cluster are the same as those of the first component clusters in each first survey and analysis network. For example, if the first component cluster contains a word vector generation component, a text encoding component, and a feature extraction component, then the first integrated cluster also has these three components. This consistency ensures the compatibility and synergy between the components during the integration process, so that the integrated cluster can smoothly inherit and integrate the functions of each first component cluster.

[0040] The process of determining the first integrated cluster configuration parameter by the computer system involves comprehensive processing of the first component cluster configuration parameter in each first survey and analysis network. Exemplary technical means include weighted average, parameter fusion and other methods.

[0041] When the computer system performs the first component cluster integration, it can be performed according to the following steps:

[0042] First, the computer system identifies the first component clusters in each first survey and analysis network and determines the corresponding relationship between them. For example, by analyzing the network structure and the functional definition of the components, components with the same functions in different networks, such as word vector generation components, text encoding components, etc., are found and matched.

[0043] Configuration parameter extraction For each corresponding component, the computer system extracts its configuration parameters. These configuration parameters may include weights, biases, learning rates, etc. For example, for a convolution layer component, its configuration parameters may include the weights and biases of the convolution kernel.

[0044] Parameter Processing and Integration Based on selected technical means, such as weighted averaging or parameter fusion, the computer system processes and integrates the extracted configuration parameters and calculates the configuration parameters of the corresponding components in the first integrated cluster.

[0045] Constructing the first integrated cluster Finally, the computer system constructs the first integrated cluster based on the integrated configuration parameters. Components with the same functions are merged and assigned the integrated configuration parameters, thereby forming a unified first integrated cluster that can comprehensively process multi-dimensional information.

[0046] Through the above steps, the computer system completes the integration of the first component clusters in each first research and analysis network and obtains the first integrated cluster. This first integrated cluster will serve as an important basis for building a multi-dimensional research and analysis network and provide stronger support for subsequent text analysis.

[0047] Step S30: combining the first integrated cluster and the first component structures included in each first research and analysis network to obtain a first integrated network.

[0048] In the process of determining the multi-dimensional research and analysis network, step S30 is to combine the first integrated cluster and the first component structures included in each first research and analysis network to obtain the first integrated network. This step is crucial for building a comprehensive network architecture that can comprehensively analyze multi-dimensional information.

[0049] The first integration cluster integrates the corresponding component clusters in each first research and analysis network, and has the ability to jointly process basic information from multiple dimensions. The first component structure is the customized part designed by each first research and analysis network for specific dimensional analysis needs, including specific functional modules such as resource classification module for resource dimension analysis and market trend prediction module for market dimension analysis. These modules provide targeted support for in-depth analysis of information in each dimension.

[0050] When the computer system performs the combination operation, it will first perform a structural adaptation check on the first component structure and the first integrated cluster of each first survey and analysis network. Ensure that they are compatible with each other in terms of data input and output format, data dimension, etc. For example, if the dimension of the feature vector output by the first integrated cluster is

[100] , and the dimension of the feature vector expected to be input by a first component structure is also

[100] , then they are compatible in this regard. If they are not compatible, the computer system uses data transformation technology, such as linear transformation, to adjust the data dimension to an appropriate range. Assuming that the feature vector output by the first integrated cluster is x, with a dimension of m, and the required input dimension of a first component structure is n, the linear transformation matrix A can be used to make y=Ax, where the dimension of y is n, and A is an n×m matrix, whose elements are calculated and determined according to actual needs and data characteristics.

[0051] Next, the computer system connects the first integration cluster with each first component structure according to a specific logic. To illustrate with a simplified example, assume that there are three first research and analysis networks, corresponding to the resource dimension, market dimension, and functional dimension. The first integration cluster has completed the integration of the basic components of each network. For the first component structure of the resource dimension, it may have a resource evaluation module. The computer system accurately inputs the resource-related information that has been preliminarily processed in the first integration cluster into the resource evaluation module, allowing the module to conduct a more in-depth resource evaluation analysis based on this information. Similarly, for the first component structure of the market dimension, the market trend prediction module therein will receive the relevant market information output by the first integration cluster and conduct a market trend prediction analysis; the functional feature extraction module in the first component structure of the functional dimension will further process the function-related information output by the first integration cluster to extract the core functional features of the product.

[0052] In the process of combination, the computer system also needs to optimize the connection relationship to ensure the efficiency and accuracy of information transmission. For example, the transmission strength of different information can be controlled by adjusting the connection weight. The idea similar to the back propagation algorithm in the neural network can be adopted to continuously adjust the connection weight according to the feedback of the final analysis results. Suppose the connection weight is , represents the connection weight from the i-th node of the first integrated cluster to the j-th node of a first component structure. After each analysis, according to the error E, according to the formula Update, where is the learning rate, which is used to control the step size of weight update.

[0053] Finally, through the above series of operations, the computer system combines the first integrated cluster and the first component structures of each first research and analysis network to form a first integrated network. This first integrated network combines the multi-dimensional basic processing capabilities of the first integrated cluster and the targeted analysis capabilities of each first component structure, laying a solid network architecture foundation for a more comprehensive and in-depth analysis of the target health product research text in the future. For example, it can comprehensively analyze a health product from multiple dimensions such as resources, market and functions at the same time, and obtain more valuable analysis results.

[0054] Step S40: calibrate the first integrated network to obtain a first neural network.

[0055] The computer system obtains the health care product survey training texts of each first survey and analysis network, which are the data basis for adjusting the corresponding first survey and analysis network. For example, for the first survey and analysis network in the resource dimension, its health care product survey training text may contain a large amount of resource information descriptions about different health care products, such as detailed data on the land area, water resources, and vegetation coverage of various health care resorts. These data, as training samples, can help the network learn the relationship between different features and health care products under the resource dimension.

[0056] For any health care product research training text of the first research and analysis network, the computer system uses the first integrated network to perform text analysis of the corresponding dimension to obtain the first analysis result corresponding to the first research and analysis network. The specific process is as follows:

[0057] The computer system uses the first integrated cluster to perform implicit text representation on any health care product research training text of the first research and analysis network, that is, the process of feature extraction, to obtain the first training semantic representation vector. For example, for a training text describing a health care hot spring product, "The hot spring has moderate water temperature, rich mineral content, and beautiful surrounding environment", the text implicit representation component in the first integrated cluster will convert the key information in the text into a vector form, such as mapping the features such as "moderate water temperature", "rich mineral content", and "beautiful surrounding environment" to specific positions in the vector space, and finally form a first training semantic representation vector that can represent the semantic information of the text.

[0058] Next, the computer system adopts the first component structure included in any one of the first survey and analysis networks, and determines the corresponding first analysis result based on the first training semantic representation vector. For example, for the first survey and analysis network of the market dimension, its first component structure may include a market demand prediction module. This module will predict the demand for the health care product in the market through specific calculation and analysis logic based on the market-related feature information in the first training semantic representation vector, such as the popularity and market share of the product, thereby obtaining the first analysis result of the market dimension.

[0059] After obtaining the first analysis results corresponding to each first survey and analysis network, the computer system will adjust the first integrated network based on these results to optimize the performance and accuracy of the network. The adjustment process can use an optimization algorithm to adjust the parameters of the network so that the difference between the output result of the network and the expected result is minimized. Exemplary optimization algorithms include stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, Adam, etc.

[0060] During the calibration process, the computer system calculates a loss function based on the difference between the first analysis result and the expected result. The loss function is used to measure the gap between the network output and the true label. Exemplary loss functions include mean square error (MSE), cross entropy loss, etc.

[0061] By continuously iterating and updating the network parameters, the computer system gradually reduces the value of the loss function, so that the performance of the first integrated network is optimized, and finally the first neural network is obtained. After adjustment, this first neural network can more accurately perform multi-dimensional text analysis on the target health product research text, providing a more reliable basis for subsequent health product recommendations. For example, when analyzing a new health food, the first neural network can more accurately evaluate its performance in multiple dimensions such as resources, market, function, experience and product, so as to provide consumers with more appropriate recommendations.

[0062] Step S50: When the first neural network meets the integration stop requirement, the first neural network is used as a multi-dimensional survey and analysis network, and the multi-dimensional survey and analysis network is configured to perform text analysis of one or more target dimensions among the text analysis of the dimensions corresponding to each first survey and analysis network on the target health care product survey text.

[0063] The integration stop requirement is a comprehensive criterion that aims to ensure that the first neural network has sufficient accuracy, stability, and generalization ability in text analysis tasks in multiple dimensions, and can be effectively applied to actual health care product research text analysis scenarios. This requirement usually involves the consideration of multiple performance indicators, such as accuracy, recall, F1 value, mean square error, etc. These indicators reflect the performance of the neural network when processing text data from different angles.

[0064] The computer system compares the actual performance measurement results of the first neural network in each dimension with the preset reference performance measurement results one by one. If the performance measurement results of the first neural network in all dimensions meet or exceed the corresponding reference performance measurement results, then the computer system determines that the first neural network meets the integration stop requirements, and at this time, the first neural network can be put into use as a multi-dimensional research and analysis network.

[0065] On the contrary, if the performance measurement result of any dimension does not reach the corresponding reference performance measurement result, the computer system will determine that the first neural network does not meet the integration stopping requirements and needs to be further adjusted and optimized. This may involve readjusting network parameters, adding training data, improving network structure, etc., and then repeating the above evaluation process until the first neural network meets the integration stopping requirements.

[0066] As an implementation mode, step S40, adjusting the first integrated network to obtain the first neural network, includes:

[0067] Step S41: obtaining health care product survey training texts of each first survey and analysis network, and the health care product survey training texts of any first survey and analysis network are used to adjust and obtain any first survey and analysis network;

[0068] Step S42: for any health care product research training text of the first research and analysis network, use the first integrated network to perform text analysis of the corresponding dimension on the health care product research training text of any first research and analysis network to obtain a first analysis result corresponding to any first research and analysis network;

[0069] Step S43: Based on the first analysis results corresponding to each first survey and analysis network, the first integrated network is adjusted to obtain a first neural network.

[0070] In step S41, the computer system obtains the health care product survey training texts of each first survey and analysis network, which are the data basis for training and adjusting the corresponding first survey and analysis network. The health care product survey training text contains rich information, covering multiple dimensions such as resources, market, function, experience and product. For example, for a survey training text set for forest health care products, it may contain multiple samples, each of which is a detailed description of a specific forest health care project. These descriptions may include the area of ​​the forest, the type of vegetation (resource dimension), the popularity of the project in the local tourism market, the flow of tourists (market dimension), the health care service items provided, the facilities (functional dimension), the evaluation of tourists, the satisfaction feedback (experience dimension) and the price of the product, the package content (product dimension) and other information. The computer system can obtain these text data in a variety of ways, such as crawling user comments and product introduction information from health care industry-related websites, forums, and social media platforms on the Internet, or purchasing sorted survey data from professional health care industry research institutions.

[0071] Step S42 uses the first integrated network to analyze the training text of the specific first research and analysis network to obtain corresponding analysis results. Taking forest health products as an example, assuming that the current focus is on the first research and analysis network of the market dimension, the computer system inputs the health product research training text corresponding to the network into the first integrated network for text analysis of the market dimension.

[0072] Text preprocessing Before analysis, the computer system can preprocess the input text. This includes but is not limited to data cleaning, removing noise data (such as special characters, repeated information, etc.) in the text; word segmentation operation, dividing the text according to words, such as dividing "forest health care projects are very popular in the local market" into "forest", "health care projects", "in", "local", "market", "very", "popular", etc.; part-of-speech tagging, determining the part of speech of each word, such as "forest" as a noun, "popular" as an adjective, etc. These preprocessing operations help the computer system better understand the semantic information of the text and provide a basis for subsequent analysis. After preprocessing, the first integration network will extract features from the text. For example, for text in the market dimension, for example, market-related features such as "market share", "competitive situation", "consumer demand", etc. are extracted. The computer system uses natural language processing technology and machine learning algorithms to identify and extract these features. For example, words are converted into vector representations through a word vector model so that the computer system can perform numerical calculations and analysis. Then, the first integration network will analyze based on these features, such as using different machine learning methods such as classification algorithms and regression algorithms to predict relevant indicators of the market dimension based on the extracted features. For example, predict the market share change trend of the forest health product in the future and obtain a specific prediction value as part of the first analysis result.

[0073] In addition to the analysis of a single feature, the first integration network will also conduct a comprehensive analysis of multiple features to obtain a more comprehensive first analysis result. For example, by combining multiple features such as "market share", "competitive situation" and "consumer demand", the overall competitiveness and development potential of the forest health care product in the market are analyzed. By performing weighted calculations, logical reasoning and other operations on these features, a comprehensive first analysis result is generated.

[0074] In step S43, after obtaining the first analysis results corresponding to each first survey and analysis network, the computer system adjusts the first integrated network according to these results to optimize the performance of the network to make it more in line with actual needs, thereby obtaining the first neural network.

[0075] For any first survey and analysis network, the computer system determines the target training error corresponding to the first survey and analysis network based on one or all of the first training error and the second training error corresponding to the first survey and analysis network. The first training error is determined based on the first analysis result and the second analysis result corresponding to the first survey and analysis network. The second analysis result is obtained by using the first survey and analysis network to perform dimensional text analysis on the health care product survey training text. For example, for the market dimension first survey and analysis network of the above-mentioned forest health care product, assuming that the first analysis result predicts that the future market share of the product is 30%, and the second analysis result obtained through actual market research and statistical data shows that its actual market share is 25%, then the first training error can be obtained by calculating the difference between the predicted value and the actual value, such as using the mean square error (MSE) formula. The second training error is determined based on the first analysis result and the training label corresponding to the first survey and analysis network, and the training label is used to indicate the analysis result obtained by performing the corresponding dimensional text analysis on the health care product survey training text of the first survey and analysis network. For example, the training label indicates that the forest health care product is positioned in the market as the mid-to-high-end market, while the first analysis result positions it as the mass market, then the second training error can be calculated based on this difference.

[0076] The computer system adjusts the first integrated network based on the target training errors corresponding to each first survey and analysis network. This involves adjusting the parameters in the network to minimize the target training error. An exemplary calibration method is to use an optimization algorithm, such as stochastic gradient descent (SGD) and its variants Adagrad, Adadelta, Adam, etc. By continuously iteratively updating the network parameters, the computer system gradually reduces the target training error, so that the performance of the first integrated network is optimized, and finally the first neural network is obtained. For example, in each iteration, the gradient is calculated according to the current target training error, and then the parameters such as weights and biases in the network are updated according to the gradient optimization until the target training error reaches a smaller value and meets the preset stop condition.

[0077] As an implementation mode, step S42, using the first integrated network to perform text analysis of the corresponding dimension on the health care product research training text of any first research and analysis network, to obtain the first analysis result corresponding to any first research and analysis network, includes:

[0078] Step S421: using the first integrated cluster to perform implicit text representation on any health care product research training text of the first research and analysis network to obtain a first training semantic representation vector;

[0079] Step S422: using the first component structure included in any first survey and analysis network, and determining a first analysis result corresponding to any first survey and analysis network based on the first training semantic representation vector.

[0080] In step S421, the implicit representation of the text is the process of converting text data into a vector form that can be understood and processed by the computer system, that is, the process of feature extraction. Through this process, the first training semantic representation vector that can represent the semantic information of the text can be obtained. When the computer system performs this step, it will use a variety of technical means to achieve effective implicit representation of the text. First, the computer system can convert the words in the text into word vectors. Word vectors are a representation method that maps words to a low-dimensional vector space, so that words with similar semantics are closer in the vector space. Exemplary word vector models include Word2Vec, GloVe, etc. For example, for words such as "hot springs" and "bathhouses" in the health care product research training text, the word vectors obtained by training the Word2Vec model will be closer in the semantic space because they have a strong semantic correlation.

[0081] After obtaining the word vector, the computer system further extracts the features of the text and encodes it. This can be achieved in many ways, such as using deep learning models such as convolutional neural networks (CNN) or recurrent neural networks (RNN). Taking CNN as an example, it can perform convolution operations on word vectors through convolution layers to extract local features. Assume that the input text is represented by word vectors to form a matrix X with a dimension of n×m, where n represents the number of words and m represents the dimension of the word vector. The size of the convolution kernel is k×m, where k represents the length of the convolution kernel in the direction of the word sequence. Through the convolution operation, a feature map can be obtained, and the calculation formula is: ,in represents the element in the i-th row and j-th column of the feature map, represents the element at the corresponding position in the input matrix X, represents the weight of the lth row and jth column in the convolution kernel, and b represents the bias term.

[0082] Pooling operation To further reduce the data dimension and extract more representative features, the computer system performs a pooling operation on the feature map. Exemplary pooling methods include maximum pooling and average pooling. Taking maximum pooling as an example, it selects the maximum value in each local area of ​​the feature map as the representative value of the area. Assuming that the dimension of the feature map is p×q and the size of the pooling window is r×s, after the maximum pooling operation, the dimension of the pooling result P is ,in , Represents the pooling window area with (i, j) as the upper left corner.

[0083] After the above series of operations, the computer system concatenates or otherwise processes the finally obtained feature representations to form a first training semantic representation vector. For example, the feature vectors after the pooling operation are concatenated to obtain a vector with a higher dimension as the first training semantic representation vector.

[0084] In step S422, after obtaining the first training semantic representation vector, the computer system further processes the vector using the first component structure in the first survey analysis network to determine the corresponding first analysis result. The first component structure is a customized design part for specific dimensional analysis requirements, which can achieve more in-depth and professional text analysis.

[0085] For example, the first training semantic representation vector is further processed. For example, the vector is linearly transformed and nonlinearly activated through a fully connected layer to enhance the expressive power of the feature. Assume that the first training semantic representation vector The dimension of is d, the weight matrix of the fully connected layer is W, the dimension is h×d, and the bias vector is , the dimension is h, then the output after the fully connected layer is The calculation formula is: ,in Represents a nonlinear activation function, such as the ReLU function, (x)=max(0, x).

[0086] According to the dimension corresponding to the first survey and analysis network, the first component structure will perform specific analysis operations. For example, for the first survey and analysis network of the resource dimension, its first component structure, for example, classifies and evaluates the resources that the health care products rely on based on the relevant features in the first training semantic representation vector. Assuming that certain features in the first training semantic representation vector represent information such as the mineral content and water temperature of the hot spring bath, the first component structure will judge the quality level of the hot spring resources based on these features, such as evaluating hot spring resources with high mineral content and suitable water temperature as high-quality resources, and obtain corresponding analysis results.

[0087] After dimension-specific analysis, the first component structure generates the first analysis result corresponding to the first survey and analysis network. This result may be a specific value, a category label, or a descriptive text. For example, for the first survey and analysis network of the market dimension, the first analysis result may be the predicted market share of the health care product in the future, such as 30%; for the first survey and analysis network of the experience dimension, the first analysis result may be the user's satisfaction evaluation label for the health care product, such as "satisfied" or "unsatisfied".

[0088] Taking forest health products as an example, assume that the first training semantic representation vector contains resource-related feature information such as forest area, vegetation type, and air quality. For the first survey and analysis network in the resource dimension, its first component structure will evaluate and classify forest resources based on these features, and finally generate the first analysis result, such as "the resource level of this forest health product is high-quality, with rich vegetation resources and good air quality."

[0089] As an implementation mode, the first integrated cluster includes one or more text implicit representation components; step S421, using the first integrated cluster to perform text implicit representation on any health care product survey training text of the first survey analysis network to obtain a first training semantic representation vector, including:

[0090] Step S4211: for the first text implicit representation component, use the first text implicit representation component to perform text implicit representation on any health care product research training text of the first research and analysis network, and obtain the execution result of the first text implicit representation component;

[0091] Step S4212: For text implicit representation components other than the first one, use the text implicit representation components other than the first one to perform text implicit representation on the execution result of the previous text implicit representation component of the text implicit representation component other than the first one, obtain the execution result of the text implicit representation components other than the first one, and use the execution result of the last text implicit representation component as the first training semantic representation vector.

[0092] In step S4211, the first text implicit representation component plays an important role in the entire text implicit representation process. It is responsible for the preliminary feature extraction and representation of the original health care product research training text. This component usually contains one or more branch components, each of which includes a weight focusing module (for example, using an attention mechanism) and a multi-layer perceptron.

[0093] The main function of the weighted focus module is to assign different weights to different parts of the text to highlight the information that is more important to the analysis task. Taking the attention mechanism as an example, it determines the importance of each word or element in the text by calculating its relevance to other parts. Specifically, for a given health product research training text ,in represents the i-th word or element in the text, and the attention mechanism calculates each x i The weight a i A common way to calculate the weights is by computing the sum of the query vector q and each x i The key vector k i The similarity between them, for example, using the dot product attention mechanism, is calculated as: , where · represents the dot product operation. In this way, the weighted focusing module can assign different importance weights to different parts of the text, so that subsequent analysis can focus more on important information.

[0094] After the weight focusing module assigns weights to the elements in the text, the multi-layer perceptron (MLP) further processes and extracts features from the text based on these weights. The multi-layer perceptron is a neural network structure that contains multiple neurons and can perform nonlinear transformations on input data. and its weight , the multilayer perceptron first transforms the weighted elements As input, features are extracted through a series of linear transformations and nonlinear activation functions. Assume that the multilayer perceptron contains L layers, and the weight matrix of the lth layer is , the bias vector is , the activation function is , then the output h of the lth layer l It can be calculated by the following formula: ,in Through the processing of multi-layer perceptron, the information in the text can be mapped into a more representative feature space.

[0095] For example, for a health product research training text "This forest health base has rich negative oxygen ion resources, fresh air, and complete supporting facilities", the weight focus module assigns higher weights to key information such as "rich negative oxygen ion resources" and "complete supporting facilities", and assigns lower weights to some connecting words and auxiliary words. Then, the multi-layer perceptron processes the text according to these weights, extracts features related to the resources and facilities of the forest health base, and obtains the execution result of the first text implicit representation component, which may be a feature vector, such as [0.8, 0.6, 0.9, ...], where each element represents the feature value of different aspects.

[0096] In step S4212, after the first text implicit representation component completes the preliminary processing of the original text, the subsequent text implicit representation components will further process and extract features of the execution results of the previous component in turn to gradually build a richer and more abstract text semantic representation. The processing process of the subsequent text implicit representation components is similar to that of the first text implicit representation component, and each subsequent text implicit representation component may also include weight focusing modules and multi-layer perceptrons. The difference is that their input is the execution result of the previous component, not the original text data. For example, the second text implicit representation component will take the execution result of the first text implicit representation component (such as the feature vector mentioned above) as input. The weight focusing module will again assign weights to the input features to highlight the more important parts. In this way, the subsequent text implicit representation components process the execution results of the previous component in turn until the last text implicit representation component completes the processing. The execution result of the last text implicit representation component is the first training semantic representation vector, which integrates the processing results of each text implicit representation component and contains deeper and more abstract semantic information about the health care product research training text. For example, after being processed by multiple text implicit representation components, the final first training semantic representation vector may be a feature vector with higher dimension and more representativeness, in which each element contains the semantic features of the text at different levels and aspects.

[0097] Through steps S4211 and S4212, the computer system gradually and deeply performs text implicit representation of the health care product research training text, and through the synergy of multiple text implicit representation components, obtains a more representative and rich first training semantic representation vector, laying a solid foundation for subsequent analysis and processing based on the vector.

[0098] As an implementation mode, the first text implicit representation component includes one or more branch components, and any branch component includes a weight focusing module and a multi-layer perceptron; in step S4211, the first text implicit representation component is used to perform text implicit representation on any health care product survey training text of the first survey analysis network, and the execution result of the first text implicit representation component is obtained, including:

[0099] Step S42111: for the first branch component, use the weight focusing module in the first branch component to perform weight focusing on any health care product research training text of the first research and analysis network to obtain a first focusing result, and use the multilayer perceptron in the first branch component to determine the execution result of the first branch component according to the first focusing result;

[0100] Step S42112: For the first branch component outside, use the weight focusing module in the first branch component outside to perform weight focusing on the execution result of the previous branch component outside the first branch component to obtain a second focusing result, use the multilayer perceptron in the first branch component outside to determine the execution result of the first branch component outside based on the second focusing result, and use the execution result of the last branch component as the execution result of the first text implicit representation component.

[0101] In step S42111, the weight focusing module and the multi-layer perceptron work together to achieve preliminary feature extraction and representation of the health care product research training text.

[0102] The core function of the weighted focus module is to assign different weights to different parts of the input text, thereby highlighting the information that is more important to the analysis task. In practical applications, the commonly used technical means is the attention mechanism. The attention mechanism determines the importance of each element (such as words, characters, etc.) by calculating the correlation between each element in the text (such as words, characters, etc.) and other elements, and assigns weights accordingly. Taking a specific health product survey training text as an example, suppose the text is "The hot spring water temperature of the hot spring health resort is moderate, the mineral content is rich, and the surrounding environment is very beautiful." The computer system will first perform word segmentation on the text to obtain a word sequence [the, hot spring, health, resort, the, hot spring, water temperature, moderate, mineral, content, rich, surrounding, environment, very, beautiful]. For the attention mechanism, it will calculate a weight for each word. The specific calculation process is as follows: First, generate a key vector k for each word i and the query vector q i (It can be obtained through word vector model or other encoding methods). Then, by calculating the query vector q i Key vector k with other words j The weight is determined by the similarity between them. Exemplary similarity calculation methods include dot product, cosine similarity, etc. For example, for the word "hot spring", after calculating its similarity with other words, its weight is 0.15; for the part of "rich in mineral content", since it is more important for describing the characteristics of hot spring health resorts, it is assigned a higher weight, such as 0.25. Through such calculations, the computer system obtains the weight of each word, that is, the first focus result.

[0103] After obtaining the first focusing result, the multi-layer perceptron (MLP) will further process and extract features from the text according to these weights. The multi-layer perceptron is a neural network structure containing multiple neurons, which can perform nonlinear transformations on input data. i and its weight ai , the multilayer perceptron first transforms the weighted element x i '=x i a i As input. Assume that the multilayer perceptron contains L layers, and the weight matrix of the lth layer is W l , the bias vector is b l , the activation function is σ l (Example activation functions include ReLU, Sigmoid, etc.) For the weighted elements of the input, the output h1 of the first layer can be calculated by the following formula: h1=σ l (W1x i '+b1). Then, the input of the second layer is the output h1 of the first layer, and its output h2 is calculated as: h2=σ2(W2h1+b2). ​​And so on, until the last layer L, the final output h L , this output is the execution result of the branch component.

[0104] In step S42112, after the first branch component completes processing, the subsequent branch components will further process the execution results of the previous branch component in turn to gradually build a richer and more abstract text semantic representation.

[0105] Operation of the weight focusing module For branch components other than the first one, the input of the weight focusing module is the execution result of the previous branch component. Taking the aforementioned health care product research training text as an example, assuming that the execution result of the first branch component for the word "hot spring" is the feature vector [0.3, 0.5, 0.2,…], the weight focusing module of the second branch component will weight this feature vector. The attention mechanism is also used to calculate the correlation between each element in the feature vector and other elements, so as to determine the weight of each element. Assume that the weight of the first element 0.3 of the feature vector is 0.2 after calculation, and the weight of the second element 0.5 is 0.3, etc. Through such calculation, a weight vector [0.2, 0.3,…] is obtained, and then the weight vector is multiplied by the corresponding element of the feature vector to obtain the second focusing result, such as [0.3×0.2, 0.5×0.3,…]=[0.06, 0.15,…].

[0106] Processing of multi-layer perceptron and determination of final execution result After obtaining the second focusing result, the multi-layer perceptron in the branch component is calculated in a similar manner to the multi-layer perceptron in the first branch component. The second focusing result is used as input, and after multiple layers of linear transformation and nonlinear activation function processing, the execution result of the branch component is obtained. For example, after calculation, the multi-layer perceptron of the second branch component converts the second focusing result [0.06, 0.15,…] into a new feature vector [0.4, 0.6, 0.1,…]. The subsequent branch components repeat the above process in sequence, and perform weighted focusing and multi-layer perceptron processing on the execution result of the previous branch component until the last branch component completes the processing. The execution result of the last branch component is the execution result of the first text implicit representation component. For example, after being processed by multiple branch components, the execution result of the first text implicit representation component may be a more representative and abstract feature vector [0.5, 0.7, 0.2,…]. This feature vector integrates the processing information of each branch component on the text and more comprehensively reflects the semantic characteristics of the health care product research training text.

[0107] As an implementation manner, step S422, using the first component structure included in any first survey and analysis network, and determining the first analysis result corresponding to any first survey and analysis network based on the first training semantic representation vector, includes:

[0108] Step S4221: using any first component structure included in the first survey analysis network to perform text implicit representation on the first training semantic representation vector to obtain a second training semantic representation vector;

[0109] Step S4222: using the first component structure included in any first survey and analysis network, mapping the second training semantic representation vector to obtain a first analysis result corresponding to any first survey and analysis network.

[0110] In step S4221, the computer system uses the first component structure to further perform text implicit representation on the first training semantic representation vector, with the purpose of extracting deeper and more targeted feature information, thereby obtaining a second training semantic representation vector.

[0111] The first component structure is a part of the first research and analysis network that is customized for specific dimension analysis requirements. It contains a series of modules and algorithms specifically used to process and analyze information of specific dimensions. For example, for the first research and analysis network of the market dimension, its first component structure may include a market trend prediction module, a consumer demand analysis module, etc.; for the first research and analysis network of the function dimension, its first component structure may include a function feature extraction module, a function effect evaluation module, etc. These modules work together to perform more professional and in-depth processing on the first training semantic representation vector.

[0112] When the computer system uses the first component structure to perform text implicit representation on the first training semantic representation vector, it will use a variety of technical means. An exemplary technical means is to use neural network structures in deep learning, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) and their variants. Taking the convolutional neural network as an example, assuming that the dimension of the first training semantic representation vector is n, the computer system regards it as a sequence of length n. First, the convolutional layer defines a set of convolutional kernels, and the size of each convolutional kernel is k (k < n). The convolutional kernels perform sliding convolutional operations on the first training semantic representation vector, and extract local features by calculating the dot product of the convolutional kernels and the local regions of the vector. The specific calculation formula is as follows:

[0113] For the first training semantic representation vector , , at position , the convolutional result c i is:

[0114] ;

[0115] Through the sliding convolutional operation, the computer system can obtain a series of convolutional results, and these results form a feature map. Then, for example, use the pooling layer to perform dimensionality reduction and feature selection on the feature map. Exemplary pooling operations include max pooling and average pooling. Max pooling will select the maximum value in each local region of the feature map as the representative value of that region, and average pooling will calculate the average value of the local region as the representative value. For example, for a feature map of size m×p, using a pooling window of size q×q for max pooling operation, the size of the resulting pooling result is . After the convolutional and pooling operations, the resulting feature vector is the second training semantic representation vector .

[0116] In step S4222, after obtaining the second training semantic representation vector, the computer system maps it to a specific analysis result to obtain the first analysis result corresponding to the first survey analysis network. The mapping process is implemented, for example, by a fully connected layer. Each neuron in the fully connected layer is connected to all neurons in the previous layer. Its function is to perform linear transformation and nonlinear activation on the input feature vector, thereby mapping the feature to the target space. Assume that the second training semantic representation vector The dimension of is d, the weight matrix of the fully connected layer is W, the dimension is h×d, and the bias vector is , with a dimension of h. The output of the fully connected layer It can be calculated by the following formula:

[0117] ;

[0118] in, Represents a nonlinear activation function. Common activation functions include ReLU, Sigmoid, Tanh, etc. The role of a nonlinear activation function is to introduce nonlinear factors and enhance the network's expressive power. For example, the definition of the ReLU function is: (x)=max(0, x), which truncates the negative part to 0 and retains the positive part.

[0119] According to the different dimensions corresponding to the first survey and analysis network, the first analysis result obtained by mapping has different meanings and forms. For example, for the first survey and analysis network of the resource dimension, the first analysis result may be the evaluation level of the resource quality of forest health care products, such as "high quality", "good", "average", etc. The computer system determines the evaluation level of resource quality based on the result of mapping the second training semantic representation vector through the fully connected layer, and through the set threshold or classification rule. Assume that the result after mapping If the value of is in the interval [0.8, 1], the resource quality is judged to be "high quality"; in the interval [0.6, 0.8) it is "good"; in the interval [0, 0.6) it is "average". For the first research and analysis network in the market dimension, the first analysis result may be the market potential forecast value of the health care product, such as the expected growth percentage of market share in the next year. The computer system directly outputs the specific value as the market potential forecast value based on the mapping result.

[0120] Through steps S4221 and S4222, the computer system can use the first component structure to further process and map the first training semantic representation vector to obtain a first analysis result corresponding to a specific dimension, which provides an important basis for subsequent adjustment of the first integration network based on the analysis result.

[0121] As an implementation mode, step S43, based on the first analysis results corresponding to each first survey and analysis network, calibrating the first integrated network to obtain the first neural network includes:

[0122] Step S431: for any first investigation and analysis network, based on one or both of the first training error and the second training error corresponding to any first investigation and analysis network, determining a target training error corresponding to any first investigation and analysis network;

[0123] Step S432: Based on the target training errors corresponding to the first survey and analysis networks, the first integrated network is calibrated to obtain a first neural network.

[0124] Among them, the first training error corresponding to any first survey and analysis network is determined based on the first analysis result and the second analysis result corresponding to any first survey and analysis network, and the second analysis result corresponding to any first survey and analysis network is obtained by using any first survey and analysis network to perform dimensional text analysis on the health care product survey training text of any first survey and analysis network; the second training error corresponding to any first survey and analysis network is determined based on the first analysis result and training label corresponding to any first survey and analysis network, and the training label corresponding to any first survey and analysis network is used to indicate the analysis result obtained by performing corresponding dimensional text analysis on the health care product survey training text of any first survey and analysis network.

[0125] In step S431, the computer system calculates and comprehensively considers the first training error and the second training error to determine a target training error for network calibration.

[0126] The first training error is determined based on the first analysis result and the second analysis result corresponding to the first survey and analysis network. The first analysis result is obtained by analyzing the health care product survey training text through the first integrated network, and the second analysis result is obtained by using the first survey and analysis network to perform dimensional text analysis on the health care product survey training text. The purpose of calculating the first training error is to measure the difference in analysis results between the first integrated network and the original first survey and analysis network, reflecting the accuracy and consistency of the first integrated network in analyzing texts of specific dimensions. An exemplary method for calculating the first training error is to use Mean Squared Error (MSE).

[0127] The second training error is determined based on the first analysis result and training label corresponding to the first survey and analysis network. The training label is used to indicate the analysis result obtained by performing text analysis of the corresponding dimension of the health care product survey training text of the first survey and analysis network. The second training error mainly measures the difference between the first analysis result and the true label, reflecting the degree of deviation between the analysis result of the first integrated network and the expected result. Commonly used methods for calculating the second training error include cross entropy loss, etc., which are suitable for classification problems; for regression problems, mean square error, etc. can also be used. Taking the classification problem as an example, assuming that the first analysis result is the predicted category probability distribution , where k is the number of categories and the training label is the true category y (y is an integer between 1 and k, representing the true category), then the formula for calculating the second training error E2 by the cross entropy loss is:

[0128] ;

[0129] in, is an indicator function, which takes the value of 1 when y=j, and takes the value of 0 otherwise. For example, for the first survey analysis network of the experience dimension, assuming that the user experience of health care products is divided into three categories: "satisfied", "average", and "unsatisfied" (i.e. k=3), the first analysis result predicts that the probability of a certain health care product user experience being "satisfied" is 0.6, the probability of being "average" is 0.3, and the probability of being "unsatisfied" is 0.1 (i.e. =[0.6, 0.3, 0.1]), while the training label indicates that the real user experience of the product is "satisfactory" (i.e. y=1), then the second training error .

[0130] The computer system determines the target training error based on one or both of the first training error and the second training error according to specific application requirements and network calibration strategies. For example, the first training error and the second training error can be simply added together as the target training error, that is, E 目标 =E1+E2; different weights can also be assigned to the two according to their importance.

[0131] In step S432, after determining the target training errors corresponding to each first survey and analysis network, the computer system adjusts the first integrated network according to the error information to optimize the performance of the network to make it more in line with actual needs, and finally obtains the first neural network.

[0132] A feasible tuning technique is to use an optimization algorithm to update the parameters of the network to minimize the target training error. Exemplary optimization algorithms include Stochastic Gradient Descent (SGD) and its variants Adagrad, Adadelta, Adam, etc. These algorithms calculate the gradient of the target training error to the network parameters, and then update the parameters according to the gradient direction and a certain learning rate. In each iteration, the target training error corresponding to each first survey analysis network is calculated based on the current first integrated network parameters. Then, the gradient of the target training error to each parameter in the network is calculated by the back propagation algorithm. The back propagation algorithm is an efficient method for calculating gradients. It starts from the output layer, calculates the error contribution of each neuron forward in sequence, and calculates the gradient according to the chain rule.

[0133] As an implementation mode, step S432, based on the target training errors corresponding to the first survey and analysis networks, calibrating the first integrated network to obtain the first neural network includes:

[0134] Step S4321: based on the target training errors corresponding to each first survey and analysis network, adjusting the first integrated cluster to obtain a first target cluster;

[0135] Step S4322: Combine the first target cluster and the first component structure included in each first survey and analysis network to obtain a first neural network.

[0136] In step S4321, the first integrated cluster is an important component of the first integrated network, and its performance directly affects the analysis effect of the entire network. The computer system performs targeted adjustments on the first integrated cluster according to the target training errors corresponding to each first survey and analysis network to optimize its processing capabilities for information of different dimensions.

[0137] Impact on the First Integration Cluster The target training errors of each first survey and analysis network reflect the degree of deviation of the first integration network in different dimensional analysis tasks. These error information will be used by the computer system to evaluate the deficiencies of the first integration cluster in feature extraction and information processing. For example, if the target training error of the first survey and analysis network in the market dimension is large, it means that the first integration cluster may not accurately extract key market-related features, or there are problems in the way it processes market information. This may be due to unreasonable parameter settings of some components in the first integration cluster (such as the text implicit representation component), resulting in inaccurate feature extraction of market dimension text.

[0138] After analyzing the target training errors corresponding to each dimension and adjusting the parameters, the computer system fully adjusts the first integrated cluster and finally obtains the first target cluster. This first target cluster has been optimized in terms of information processing capabilities in each dimension and can more accurately extract and process multi-dimensional feature information related to health care products.

[0139] In step S4322, after obtaining the first target cluster, the computer system recombines it with the first component structures included in each first survey and analysis network to construct a complete first neural network. The purpose of this step is to combine the optimization capabilities of the adjusted first target cluster with the professional analysis capabilities of the first component structure for different dimensions to form an overall network with better performance.

[0140] The computer system connects and integrates the first target cluster with each first component structure according to a specific network architecture design. In terms of the connection method, it is ensured that information can be smoothly transmitted between the first target cluster and the first component structure. For example, the optimized feature information output by the first target cluster can be accurately input into the corresponding first component structure for further analysis and processing.

[0141] The first target cluster is responsible for basic, multi-dimensional feature extraction and preliminary processing of the input health care product research text, while the first component structure conducts in-depth analysis and transformation of these features according to the needs of different dimensions. For example, for the first component structure of the functional dimension, it will receive the feature information related to the product function extracted by the first target cluster, and then further analyze the specific functions and effects of the product through its internal functional characteristic extraction module, functional effect evaluation module, etc. This combination method gives full play to the advantages of both, enabling the first neural network to understand and analyze the information of health care products in various dimensions more comprehensively and deeply. Adaptation and adjustment during the combination process During the combination process, the computer system ensures the compatibility and adaptability between the first target cluster and the first component structure. This may involve adjustments to data formats, dimensions, etc. For example, if the dimension of the feature vector output by the first target cluster is inconsistent with the dimension of the expected input of the first component structure, the computer system, for example, uses data transformation technology, such as linear transformation or nonlinear transformation, to adjust the dimension so that the two can be smoothly connected.

[0142] Assume that the dimension of the resource dimension feature vector output by the first target cluster is m, and the expected input dimension of the first component structure of the functional dimension is The computer system can transform the resource dimension feature vector output by the first target cluster through a linear transformation matrix A (dimension is n×m) Transform and obtain the new feature vector , so that its dimension meets the requirements of the first component structure of the functional dimension.

[0143] At the same time, the computer system also fine-tunes the connection weights, for example, to optimize the efficiency and accuracy of information transmission. Through continuous adjustment and optimization, it is ensured that the first target cluster and the first component structure can work together to achieve the best performance.

[0144] By combining the first target cluster and each first component structure, the computer system successfully obtains the first neural network. This first neural network combines the advantages of both, and can provide more accurate and comprehensive analysis results when conducting multi-dimensional analysis of health care product research texts, providing solid and reliable support for subsequent health care product recommendations and decisions.

[0145] As an embodiment, the method further includes:

[0146] Step S60: obtaining a first performance measurement result of the first neural network on the text analysis of the dimension corresponding to each first research and analysis network;

[0147] Step S70: when each first performance measurement result is greater than the corresponding reference performance measurement result, determining that the first neural network does not meet the integration stop requirement;

[0148] Step S80: When a first performance measurement result is not greater than a corresponding reference performance measurement result, it is determined that the first neural network meets the integration stop requirement; wherein the reference performance measurement result corresponding to any first performance measurement result is determined based on the performance measurement result of the first survey and analysis network corresponding to any first performance measurement result.

[0149] In step S60, after completing the construction of the first neural network, the computer system evaluates its performance in the text analysis tasks of each dimension. The performance measurement results can reflect the accuracy, reliability and effectiveness of the first neural network in analyzing the health care product survey text. Exemplary performance measurement indicators include F1 score, precision, recall rate, accuracy, etc. Different indicators measure the performance of the network from different angles.

[0150] The F1 score is the harmonic mean of precision and recall, which takes into account the precision and recall of the model and can comprehensively evaluate the performance of the model. For example, when analyzing the market dimension of health care products, assume that the computer system divides 100 health care products into two categories: popular products and unpopular products. Among them, the actual situation is that there are 60 popular products and 40 unpopular products. The first neural network predicts 50 popular products, of which 40 are correct. Then, precision = 0.8, recall = 0.67, according to the F1 score calculation formula: F1 = 0.73. Precision refers to the proportion of samples predicted as positive examples that are actually positive examples, which is used to measure the accuracy of the model's prediction results. Recall refers to the proportion of samples predicted as positive examples that are actually positive examples, which is used to measure the model's coverage of positive examples. Accuracy refers to the proportion of samples predicted correctly to the total number of samples, reflecting the overall prediction accuracy of the model. The computer system tests a large number of health care product research texts on the text analysis tasks corresponding to the dimensions of each first research analysis network, and calculates the corresponding first performance measurement results. For example, for the resource dimension, market dimension, function dimension, experience dimension and product dimension, the corresponding F1 score, precision rate, recall rate and accuracy rate are calculated respectively as the first performance measurement results.

[0151] In step S70, the reference performance measurement result is determined based on the performance measurement results of each first survey and analysis network, and it represents the minimum performance standard expected to be achieved in this dimension. The computer system compares the first performance measurement result of the first neural network in each dimension with the corresponding reference performance measurement result.

[0152] If the first performance measurement results of each dimension are greater than the corresponding reference performance measurement results, it means that the first neural network performs well in the text analysis tasks of each dimension, and there is room for further improvement and optimization. Therefore, it is determined that the first neural network does not meet the integration stop requirements. For example, suppose that in the market dimension, the reference performance measurement results are set as F1 score greater than 0.6, precision greater than 0.7, recall greater than 0.65, and accuracy greater than 0.75. The first performance measurement results of the first neural network in the market dimension are F1 score 0.8, precision 0.85, recall 0.75, and accuracy 0.82. Since each indicator is greater than the corresponding reference performance measurement results, the computer system determines that the first neural network does not meet the integration stop requirements in the market dimension and needs to continue to be optimized and adjusted.

[0153] In step S80, when the first performance measurement result of at least one dimension is not greater than the corresponding reference performance measurement result, it means that the performance of the first neural network in this dimension has not yet reached the expected standard, and there may be problems such as inaccurate analysis and instability. At this time, the computer system determines that the first neural network meets the integration stop requirements, that is, no further optimization and integration will be performed on it.

[0154] For example, in the experience dimension, the reference performance measurement results are set as F1 score greater than 0.7, precision greater than 0.75, recall greater than 0.7, and accuracy greater than 0.8. The first performance measurement results of the first neural network in the experience dimension are F1 score 0.65, precision 0.7, recall 0.68, and accuracy 0.78, where the F1 score and accuracy are not greater than the corresponding reference performance measurement results, so the computer system determines that the first neural network meets the integration stop requirements. At this time, the performance of the first neural network has reached the expectations of the current stage, and can be used as a multi-dimensional research and analysis network for subsequent health care product recommendations and analysis tasks.

[0155] In summary, steps S60 to S80 determine whether the first neural network meets the integration stop requirement by obtaining the first performance measurement results of the first neural network in each dimension and comparing them with the corresponding reference performance measurement results. This evaluation and judgment mechanism can ensure that the final multi-dimensional research and analysis network has good performance in each dimension, thereby providing reliable support for the recommendation and analysis of health care products.

[0156] As an implementation, the first neural network includes a first target cluster, each first component structure includes a second component cluster with a consistent composition, and the first component structure also includes a second component structure outside the second component cluster; the method further includes:

[0157] Step S90: when the first neural network does not meet the integration stop requirement, integrating the second component clusters included in each of the first component structures to obtain a second integrated cluster;

[0158] Step S100: combining the first target cluster, the second integrated cluster, and the second component structures included in each first component structure to obtain a second integrated network;

[0159] Step S110: calibrating the second integrated network to obtain a second neural network;

[0160] Step S120: When the second neural network meets the integration stop requirement, the second neural network is used as a multi-dimensional research and analysis network.

[0161] In step S90, when the computer system determines that the first neural network does not meet the integration stop requirement, it means that the network has room for improvement in text analysis performance in certain dimensions. At this time, the computer system focuses on the second component clusters contained in each first component structure and integrates them. The second component clusters are the parts with consistent composition in each first component structure. Integrating them helps to further extract and integrate common features in different dimensions and improve the comprehensive analysis ability of the network.

[0162] For example, suppose there are three first research and analysis networks, corresponding to the resource dimension, market dimension, and functional dimension respectively. Each first component structure contains a second component cluster, which is used to process information related to the basic attributes of health care products, such as the basic description of the product, the applicable population, etc. In the first component structure of the resource dimension, the second component cluster may extract basic resource-related features by analyzing information in the text about the geographical location and area of ​​the product; in the first component structure of the market dimension, the second component cluster may extract basic features from aspects such as the product's market positioning and target customer groups; in the first component structure of the functional dimension, the second component cluster may focus on basic information such as the product's main functional characteristics and technical parameters.

[0163] When integrating these second component clusters, the computer system will first identify the commonalities and differences between them. The common parts can be merged directly. For example, the processing methods for the basic description of the product may be similar in various dimensions. The computer system can unify these similar processing logics and parameters. For the parts with differences, the computer system uses methods such as weighted averaging and feature selection to process them. For example, if the description methods of the applicable population are different in different dimensions, the computer system can use weighted averaging to assign weights to different description methods according to the importance of each dimension, and then synthesize a unified processing method.

[0164] In step S100, after obtaining the second integrated cluster, the computer system combines it with the first target cluster and the second component structures in each first component structure except the second component cluster to construct a second integrated network. The first target cluster is the part that performs better in multi-dimensional feature extraction after adjustment, the second integrated cluster focuses on processing the common basic features under each dimension, and the second component structure provides support for in-depth analysis of specific dimensions.

[0165] Taking the market dimension of health care products as an example, the first target cluster may be able to extract macro characteristics related to the market, such as market size, development trend, etc.; the second integration cluster can further process the basic attribute information of the product, such as brand awareness, product price range, etc.; the second component structure focuses on specific analysis of the market dimension, such as competitor analysis, market segmentation, etc. When combining these parts, the computer system needs to ensure the smooth transmission and effective integration of information between different parts.

[0166] In step S110, in order to make the second integrated network have better performance in text analysis tasks in various dimensions, the computer system adjusts it to optimize the parameters and structure of the network to make it more in line with actual needs, thereby obtaining a second neural network.

[0167] The tuning process mainly involves training the network using training data and adjusting the network parameters based on the training results. Exemplary training methods include optimization algorithms based on gradient descent, such as stochastic gradient descent (SGD), Adagrad, Adadelta, Adam, etc. These algorithms calculate the gradient of the loss function to the network parameters, and then update the parameters according to the gradient direction and a certain learning rate to minimize the loss function.

[0168] In step S120, after obtaining the second neural network, the computer system determines whether it meets the integration stop requirement. The integration stop requirement is a set of performance indicators set according to actual application scenarios and requirements, and is used to measure whether the performance of the neural network in the text analysis tasks of various dimensions meets expectations.

[0169] For example, the set integration stop requirements may include that the accuracy, recall, F1 score and other indicators in each dimension have reached a certain threshold. The computer system calculates the performance measurement results of the second neural network in each dimension and compares them with the set threshold. If the performance measurement results of the second neural network in all dimensions meet the integration stop requirements, it means that the network has good performance in multi-dimensional text analysis tasks and can accurately and comprehensively analyze the information of health care products in multiple dimensions such as resources, market, function, experience and product. At this time, the computer system uses the second neural network as the final multi-dimensional research and analysis network.

[0170] For example, suppose that in the resource dimension, the accuracy threshold is set to 90%, the recall threshold is set to 85%, and the F1 score threshold is set to 88%; in the market dimension, the accuracy threshold is set to 88%, the recall threshold is set to 82%, and the F1 score threshold is set to 85%, etc. After evaluating the performance of the second neural network, the accuracy in the resource dimension is 92%, the recall rate is set to 87%, and the F1 score is set to 90%; the accuracy in the market dimension is set to 89%, the recall rate is set to 84%, and the F1 score is set to 86%, etc. Since the performance measurement results in each dimension meet the set thresholds, the computer system determines that the second neural network meets the integration stop requirements and uses it as a multi-dimensional research and analysis network for actual health care product recommendation and analysis tasks.

[0171] Through the operations of steps S90 to S120, when the first neural network does not meet the integration stop requirements, the computer system integrates the second component cluster, recombines and adjusts the network, and finally obtains a multi-dimensional research and analysis network that meets the requirements, further improving the analysis ability and accuracy of the health care product research text, and providing stronger support for the recommendation and decision-making of health care products.

[0172] As an implementation method, step S2 uses a multi-dimensional research and analysis network to perform text analysis of one or more target dimensions on the target health care product research text to obtain text analysis results of each target dimension, including:

[0173] Step S21: presenting text analysis of several candidate dimensions corresponding to the multi-dimensional research and analysis network;

[0174] Step S22: In response to the user's confirmation of the text analysis of one or more target dimensions in the text analysis of several candidate dimensions, a multi-dimensional survey analysis network is used to perform text analysis of each target dimension on the target health care product survey text to obtain text analysis results of each target dimension.

[0175] Before the computer system conducts text analysis on the target dimension, it needs to first present to the user text analysis options for several candidate dimensions corresponding to the multi-dimensional research and analysis network. These candidate dimensions usually include resource dimensions, market dimensions, functional dimensions, experience dimensions, and product dimensions, and each dimension represents an analytical perspective on different aspects of health care products.

[0176] The display method can be presented to the user through a user interface (UI). For example, in a software interface for health care product analysis, the computer system displays the text analysis options of each candidate dimension in the form of a list or menu. Each option is accompanied by a brief description to help users understand the specific content and significance of the analysis of this dimension. For example, for the text analysis option of the resource dimension, the description may be "analyze the various resource conditions that health care products rely on, such as land resources, water resources, biological resources, etc."; for the text analysis option of the market dimension, the description may be "evaluate the performance of health care products in the market, including market share, competitive situation, consumer demand, etc."

[0177] Computer systems can use graphical interface design techniques to achieve this presentation. For example, front-end technologies such as HTML, CSS, and JavaScript can be used to build an intuitive and user-friendly user interface. On the interface, the text analysis options for each candidate dimension can be presented in the form of buttons or links, and users can select the target dimension of interest by clicking the corresponding button or link.

[0178] In addition, in order to help users better understand the meaning and role of each candidate dimension, the computer system can also provide some examples or explanatory information. For example, next to the text analysis option of each candidate dimension, some examples of the analysis results of health care products that have been analyzed on this dimension are displayed, or some relevant industry knowledge and research report links are provided for user reference.

[0179] When the user selects and confirms one or more target dimensions from the text analysis of several candidate dimensions presented, the computer system uses a multi-dimensional research and analysis network to perform text analysis of the target health care product research text on the corresponding target dimensions based on the user's selection.

[0180] The computer system monitors the user's operations on the user interface in real time. When the user clicks the confirmation button or selects a specific target dimension, the system will obtain the user's selection information. For example, if the user selects the resource dimension and the market dimension as the target dimension on the interface, the computer system will record this selection.

[0181] Before conducting text analysis, the computer system will first preprocess the target health care product research text. The preprocessing process includes but is not limited to data cleaning, word segmentation, part-of-speech tagging and other operations. Data cleaning is to remove noise data in the text, such as special characters, repeated information, etc.; word segmentation is to divide the text into words for subsequent analysis; part-of-speech tagging is to determine the part of speech of each word, such as noun, verb, adjective, etc. For example, for a health care product research text "This health care product is located in a beautiful mountainous area, with rich natural resources and broad market prospects", after word segmentation, it is obtained that "This health care product is located in a beautiful mountainous area, with rich natural resources and broad market prospects", and after part-of-speech tagging, "health care product" is marked as a noun, "beautiful" is marked as an adjective, etc.

[0182] After preprocessing, the computer system uses a multi-dimensional research and analysis network to extract features and analyze the target health care product research text. The multi-dimensional research and analysis network consists of a first target cluster and a first component structure corresponding to the text analysis of each candidate dimension. For different target dimensions, the computer system uses different first component structures for analysis.

[0183] Taking the resource dimension as an example, the first component structure focuses on words and information related to resources in the text, such as "natural resources", "land resources", "water resources", etc. The computer system performs a preliminary implicit text representation of the text through the first target cluster to obtain the target text semantic representation vector. For example, the words in the text are converted into vector representations through a word vector model, and then these word vectors are processed through deep learning models such as convolutional neural networks (CNN) or recurrent neural networks (RNN) to extract more representative feature vectors as the target text semantic representation vector. Then, this target text semantic representation vector is input into the first component structure corresponding to the resource dimension, which may contain some specific analysis modules, such as resource classification module and resource evaluation module. The resource classification module will classify the resources that health care products rely on according to the feature information in the target text semantic representation vector, such as dividing resources into natural resources and social resources; the resource evaluation module will evaluate the quality and quantity of resources based on the classification results and other relevant information to obtain the text analysis results of the resource dimension.

[0184] For the market dimension, the first component structure, for example, focuses on market-related information in the text, such as "market prospects", "market share", "competitive situation", etc. Similarly, the computer system first obtains the semantic representation vector of the target text through the first target cluster, and then inputs it into the first component structure corresponding to the market dimension. This structure may include a market trend prediction module, a consumer demand analysis module, etc. The market trend prediction module will predict the market development trend of health care products based on the feature information in the semantic representation vector of the target text, combined with market historical data and related models; the consumer demand analysis module will analyze the consumer demand characteristics and preferences for health care products reflected in the text, thereby obtaining the text analysis results of the market dimension.

[0185] After completing the feature extraction and analysis of each target dimension, the computer system will organize and present the analysis results. The presentation method can be to generate a detailed analysis report, which contains the analysis results and related explanations of each target dimension. For example, for the analysis results of the resource dimension, the report lists the various types of resources owned by health care products and their evaluation levels; for the analysis results of the market dimension, the report includes charts showing market development trends and analysis conclusions of consumer demand.

[0186] As an implementation mode, the multi-dimensional survey and analysis network includes a first target cluster and a first component structure corresponding to the text analysis of each candidate dimension; step S22, in response to the user's confirmation of the text analysis of one or more target dimensions in the text analysis of several candidate dimensions, the multi-dimensional survey and analysis network is used to perform text analysis of each target dimension on the target health care product survey text, and obtain the text analysis results of each target dimension, including:

[0187] For the text analysis of any target dimension, in response to the user's confirmation of the text analysis of any target dimension in the text analysis of several candidate dimensions, the first target cluster is used to perform implicit text representation on the target health care product research text to obtain the target text semantic representation vector; the first component structure corresponding to the text analysis of any target dimension is used to determine the text analysis result of the target dimension corresponding to the text analysis of any target dimension based on the target text semantic representation vector.

[0188] The computer system will first identify the text analysis of any target dimension confirmed by the user in the text analysis of several candidate dimensions. Assume that the candidate dimensions include resource dimension, market dimension, function dimension, experience dimension and product dimension, and the user confirms the text analysis of the function dimension. At this time, the computer system calls the first target cluster in the multi-dimensional survey analysis network. The first target cluster contains one or more text implicit representation components, which are used to extract features from the input text, that is, to convert the text into a semantic vector form that the computer can understand and process. For the first text implicit representation component, the computer system uses it to perform text implicit representation operations on the target health product survey text. Taking a simple health product survey text fragment "This health product has a variety of massage functions, adjustable strength and mode" as an example, the first branch component in the first text implicit representation component starts working. The weight focus module in the branch component uses the attention mechanism to weight focus on this text fragment. The role of the attention mechanism is to allow the computer system to focus on the important parts of the text, and to highlight key information by calculating the weights of different words or parts of the text.

[0189] Next, the multi-layer perceptron in the branch component determines the execution result of the branch component based on the first focusing result. The multi-layer perceptron is a neural network structure containing multiple neuron layers, which can perform nonlinear transformation and feature extraction on the input data. For example, the multi-layer perceptron may transform the key information "massage function" and "adjustment strength and mode" after focusing into a feature vector, which is the execution result of the branch component.

[0190] For other branch components other than the first one, the computer system processes them in the same way. That is, the weight focusing module in each branch component performs weight focusing on the execution result of the previous branch component, and then the multi-layer perceptron determines the execution result of the current branch component based on the new focusing result. Until the last branch component, its execution result will be used as the execution result of the first text implicit representation component.

[0191] Then, the computer system follows the same process, allowing other text implicit representation components other than the first text implicit representation component to perform text implicit representation on the execution result of the previous component in turn, and finally uses the execution result of the last text implicit representation component as the target text semantic representation vector. For example, the second text implicit representation component will further extract and transform the execution result of the first text implicit representation component, so that the final target text semantic representation vector can more accurately represent the semantic information of the target health care product research text in the functional dimension.

[0192] Step S222 illustrates how the computer system uses the first component structure corresponding to the text analysis of any target dimension to determine the text analysis result of the target dimension based on the target text semantic representation vector. The computer system uses a part of the first component structure corresponding to the text analysis of the target dimension to perform text implicit representation on the target text semantic representation vector to obtain a second training semantic representation vector. This step further processes the extracted target text semantic representation vector to mine deeper semantic information. For example, assuming that the target dimension is a functional dimension, this part of the corresponding first component structure may be a neural network layer specifically used to analyze function-related semantics, which will perform nonlinear transformation and feature extraction on the target text semantic representation vector to obtain a more representative second training semantic representation vector. Then, the computer system uses another part of the first component structure corresponding to the text analysis of the target dimension to map the second training semantic representation vector to obtain the text analysis result of the target dimension. For example, this other part may be a fully connected layer, which maps the second training semantic representation vector to a specific output space, and the result of this output space is the text analysis result on the functional dimension. Assuming that the output space is a vector space representing information such as function type and function strength, through the mapping of the fully connected layer, the computer system can determine the specific analysis results of the health care product in the functional dimension, such as determining that the product has multiple massage functions and the function strength is adjustable.

[0193] In the actual implementation process, the computer system can use a variety of technical means to implement the above steps. For the attention mechanism in the weight focusing module, different variants can be used, such as self-attention mechanism, multi-head attention mechanism, etc., to better capture the semantic relationship in the text. For multi-layer perceptrons and other neural network layers, suitable activation functions such as ReLU function (ReLU(x)=max(0, x)) can be used to increase the nonlinear expression ability of the model. During the training process, stochastic gradient descent (SGD) and its variants, such as Adagrad, Adadelta, Adam and other optimization algorithms can be used to adjust the parameters of the model to minimize the loss function and improve the performance and accuracy of the model.

[0194] Through the above detailed steps and implementation methods, the computer system can accurately respond to the user's confirmation of the target dimension text analysis, use the multi-dimensional survey analysis network to analyze the target health care product survey text, obtain the text analysis results of each target dimension, and provide strong support for the subsequent health care product recommendation. This series of operations can not only effectively extract key information from the text, but also improve the accuracy and reliability of the analysis results through continuous training and optimization, so as to better meet the user's needs in the selection and evaluation of health care products.

[0195] The present application embodiment provides a computer system, such as Figure 3 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.

[0196] Processor 101 may be a CPU, a general-purpose processor, a GPU, a DSP, an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 101 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0197] The bus 102 may include a path to transmit information between the above components. The bus 102 may be a PCI bus or an EISA bus, etc. The bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0198] The memory 103 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, a CD-ROM or other optical disk storage, an optical disk storage (including a compressed optical disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0199] The memory 103 is used to store application code for executing the solution of the present application, and the execution is controlled by the processor 101. The processor 101 is used to execute the application code stored in the memory 103 to implement the content shown in any of the above method embodiments.

[0200] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processor, the above-mentioned health care product recommendation method based on the RMFEP model is implemented.

[0201] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program runs on a processor, the processor can execute the corresponding content in the aforementioned method embodiment.

[0202] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A health care product recommendation method based on the RMFEP model is characterized by: The method comprises: Obtain research text on target health and wellness products; Use a multi-dimensional research and analysis network to perform text analysis on the target health care product research text in one or more target dimensions to obtain text analysis results for each target dimension. The process of obtaining the multi-dimensional research and analysis network includes: Acquire a plurality of first research and analysis networks, wherein any one of the first research and analysis networks is configured to perform text analysis on a corresponding dimension of a target health and wellness product research text, each of the first research and analysis networks including a first component cluster having a consistent composition, and any one of the first research and analysis networks also including a first component structure outside of the first component cluster; wherein the dimensions include a resource dimension, a market dimension, a functional dimension, an experience dimension, and a product dimension; Integrating the first component clusters included in each of the first investigation and analysis networks to obtain a first integrated cluster; wherein the first integrated cluster and the first component cluster have the same composition, and configuration parameters of the first integrated cluster are determined using the configuration parameters of the first component clusters included in each of the first investigation and analysis networks; combining the first integrated cluster and the first component structures included in each of the first research and analysis networks to obtain a first integrated network; Adjusting the first integrated network to obtain a first neural network; When the first neural network meets the integration stop requirement, the first neural network is used as a multi-dimensional research and analysis network, and the multi-dimensional research and analysis network is configured to perform text analysis on the target health care product research text in one or more target dimensions among the text analysis of the dimensions corresponding to the various first research and analysis networks; The step of adjusting the first integrated network to obtain a first neural network includes: Obtaining health care product research training texts of each first research and analysis network, wherein the health care product research training texts of any first research and analysis network are used to adjust and obtain the any first research and analysis network; For the health care product research training text of any one of the first research and analysis networks, using the first integrated network to perform text analysis of corresponding dimensions on the health care product research training text of any one of the first research and analysis networks to obtain a first analysis result corresponding to the any one of the first research and analysis networks; Based on the first analysis results corresponding to each first survey and analysis network, the first integrated network is adjusted to obtain a first neural network.

2. The method according to claim 1, characterized in that The step of using the first integrated network to perform text analysis of corresponding dimensions on the health care product research training text of any one of the first research and analysis networks to obtain a first analysis result corresponding to any one of the first research and analysis networks includes: Using the first integrated cluster to perform text implicit representation on the health care product research training text of any one of the first research and analysis networks to obtain a first training semantic representation vector; The first component structure included in any one of the first survey and analysis networks is adopted, and a first analysis result corresponding to the any one of the first survey and analysis networks is determined based on the first training semantic representation vector.

3. The method according to claim 2, characterized in that The first integrated cluster includes one or more text implicit representation components; the first integrated cluster is used to perform text implicit representation on the health care product survey training text of any first survey and analysis network to obtain a first training semantic representation vector, including: For a first text implicit representation component, use the first text implicit representation component to perform text implicit representation on the health care product research training text of any one of the first research and analysis networks, and obtain an execution result of the first text implicit representation component; For text implicit representation components other than the first one, the text implicit representation components other than the first one are used to perform text implicit representation on the execution result of the previous text implicit representation component of the text implicit representation component other than the first one, and the execution result of the text implicit representation component other than the first one is obtained, and the execution result of the last text implicit representation component is used as the first training semantic representation vector.

4. The method according to claim 3, characterized in that The first implicit text representation component includes one or more branch components, any one of which includes a weighted focusing module and a multi-layer perceptron; the first implicit text representation component is used to perform implicit text representation on the health care product survey training text of any first survey analysis network, and the execution result of the first implicit text representation component is obtained, including: For the first branch component, a weight focusing module in the first branch component is used to perform weight focusing on the health care product research training text of any one of the first research and analysis networks to obtain a first focusing result, and a multilayer perceptron in the first branch component is used to determine an execution result of the first branch component based on the first focusing result; For the first external branch component, using the weight focusing module in the first external branch component to perform weight focusing on the execution result of the previous branch component of the first external branch component to obtain a second focusing result, using the multilayer perceptron in the first external branch component to determine the execution result of the first external branch component based on the second focusing result, and using the execution result of the last branch component as the execution result of the first text implicit representation component; The step of using the first component structure of any one of the first survey and analysis networks and determining a first analysis result corresponding to the any one of the first survey and analysis networks based on the first training semantic representation vector includes: Using the first component structure included in any one of the first survey analysis networks, performing text implicit representation on the first training semantic representation vector to obtain a second training semantic representation vector; The first component structure included in any one of the first survey and analysis networks is used to map the second training semantic representation vector to obtain a first analysis result corresponding to the any one of the first survey and analysis networks.

5. The method according to claim 1, wherein The first integrated network is adjusted based on the first analysis results corresponding to the first survey and analysis networks to obtain the first neural network, including: For any first investigation and analysis network, determining a target training error corresponding to the any first investigation and analysis network based on one or both of a first training error and a second training error corresponding to the any first investigation and analysis network; Adjusting the first integrated network based on target training errors corresponding to each first survey and analysis network to obtain a first neural network; The first training error corresponding to any one of the first survey and analysis networks is determined based on the first analysis result and the second analysis result corresponding to any one of the first survey and analysis networks, and the second analysis result corresponding to any one of the first survey and analysis networks is obtained by performing dimensional text analysis on the health care product survey training text of any one of the first survey and analysis networks using the any one of the first survey and analysis networks; The second training error corresponding to any one of the first survey and analysis networks is determined based on the first analysis result and training label corresponding to the any one of the first survey and analysis networks, and the training label corresponding to the any one of the first survey and analysis networks is used to indicate the analysis result obtained by performing a text analysis of the corresponding dimension on the health care product survey training text of the any one of the first survey and analysis networks; The step of adjusting the first integrated network based on the target training errors corresponding to the respective first survey and analysis networks to obtain the first neural network includes: Adjusting the first integrated cluster based on target training errors corresponding to each first survey and analysis network to obtain a first target cluster; The first target cluster and the first component structures included in each of the first survey and analysis networks are combined to obtain a first neural network.

6. The method according to claim 1, wherein The method further comprises: Obtaining first performance measurement results of the first neural network on text analysis of the dimensions corresponding to each of the first research and analysis networks; When each first performance metric result is greater than the corresponding reference performance metric result, determining that the first neural network does not meet the integration stop requirement; When the first performance metric result is not greater than the corresponding reference performance metric result, determining that the first neural network meets the integration stopping requirement; The reference performance measurement result corresponding to any first performance measurement result is determined based on the performance measurement result of the first survey and analysis network corresponding to any first performance measurement result.

7. The method according to any one of claims 1 to 6, characterized in that The first neural network includes a first target cluster, each first component structure includes a second component cluster with a consistent composition, and the first component structure also includes a second component structure outside the second component cluster; the method further includes: When the first neural network does not meet the integration stop requirement, integrating the second component clusters included in each of the first component structures to obtain a second integrated cluster; combining the first target cluster, the second integrated cluster, and the second component structures included in each of the first component structures to obtain a second integrated network; Adjusting the second integrated network to obtain a second neural network; When the second neural network meets the integration stop requirement, the second neural network is used as the multi-dimensional research and analysis network.

8. The method according to claim 1, characterized in that The multi-dimensional research and analysis network is used to perform text analysis on the target health care product research text in one or more target dimensions to obtain text analysis results for each target dimension, including: Presenting text analysis of several candidate dimensions corresponding to the multi-dimensional research and analysis network; In response to the user confirming the text analysis of one or more target dimensions in the text analysis of the plurality of candidate dimensions, a multi-dimensional survey analysis network is used to perform text analysis of the target health care product survey text in each target dimension to obtain text analysis results for each target dimension; The multi-dimensional research and analysis network includes a first target cluster and a first component structure corresponding to the text analysis of each candidate dimension; in response to the user's confirmation of the text analysis of one or more target dimensions in the text analysis of the plurality of candidate dimensions, the multi-dimensional research and analysis network is used to perform text analysis of each target dimension on the target health care product research text, and obtain text analysis results for each target dimension, including: For text analysis of any target dimension, in response to a user's confirmation of the text analysis of any target dimension in the text analysis of the plurality of candidate dimensions, the first target cluster is used to perform implicit text representation on the target health care product research text to obtain a semantic representation vector of the target text; The first component structure corresponding to the text analysis of any target dimension is adopted, and the text analysis result of the target dimension corresponding to the text analysis of any target dimension is determined based on the target text semantic representation vector.

9. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 8 is implemented.

Images