Iterative design method and system for cultural and creative products based on user emotional feedback

Abstract

The present invention relates to the field of deep learning technology, and discloses a method and system for iterative design of cultural and creative products based on user emotional feedback. The method includes: fusing multiple features of a first user evaluation text to obtain a fine-grained emotional dictionary in the field of cultural and creative products; extracting enhanced word representations and constructing a user emotional feedback information extraction model; processing a second user evaluation text and calculating the design parameters to be optimized; performing grid division to generate a diversified set of optimized design solutions; performing score prediction through an initial design solution evaluation model to obtain a set of candidate design solutions; iteratively optimizing the set of candidate design solutions, updating the initial design solution evaluation model, and outputting a target design solution evaluation model and a target cultural and creative product design solution when a termination condition is met. The present invention can accurately capture user needs, and also has the ability to generate diversified innovative solutions, and can effectively evaluate and screen the solutions.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an iterative design method and system for cultural and creative products based on user emotional feedback. Background Art

[0002] Traditional cultural and creative product design methods often rely on the subjective experience of designers, making it difficult to accurately grasp user needs and emotional preferences. With the popularization of the Internet, a large amount of user evaluation texts have made it possible to deeply understand user emotions. However, how to extract valuable user emotional feedback information from massive unstructured texts and effectively transform it into a basis for design optimization is still a difficult problem that needs to be solved.

[0003] In addition, the design of cultural and creative products often involves parameters in multiple dimensions. How to quickly locate the optimal solution in a huge design space also brings challenges to the design process. Traditional manual iterative optimization methods are inefficient and difficult to cope with rapidly changing market demands. Therefore, there is an urgent need for a method that can automatically process user emotional feedback and achieve efficient design optimization. Summary of the invention

[0004] The present invention provides an iterative design method and system for cultural and creative products based on user emotional feedback, which is used to accurately capture user needs, has the ability to generate diversified innovative solutions, and can effectively evaluate and screen the solutions.

[0005] In a first aspect, the present invention provides an iterative design method for cultural and creative products based on user emotional feedback, the iterative design method for cultural and creative products based on user emotional feedback comprising:

[0006] The first user's evaluation text is subjected to multi-feature fusion to obtain a sentiment seed word set and cluster analysis is performed to obtain a fine-grained sentiment dictionary in the field of cultural and creative products.

[0007] Extracting the word context representation of the first user evaluation text, performing first-order and second-order feature fusion to obtain enhanced word representation, and constructing a user sentiment feedback information extraction model;

[0008] Processing the second user evaluation text based on the user sentiment feedback information extraction model to obtain feature importance weights and evaluation vocabulary weight values, and calculating the design parameters to be optimized;

[0009] Performing grid division in a multidimensional design space based on the design parameters to be optimized to obtain a set of design points, and generating a set of diversified optimization design solutions based on the set of design points;

[0010] Inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions;

[0011] Iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and when the termination condition is met, output the target design solution evaluation model and the target cultural and creative product design solution.

[0012] In a second aspect, the present invention provides an iterative design system for cultural and creative products based on user emotional feedback, the iterative design system for cultural and creative products based on user emotional feedback comprising:

[0013] A fusion module is used to fuse multiple features of the first user's evaluation text to obtain a sentiment seed word set and perform cluster analysis to obtain a fine-grained sentiment dictionary in the field of cultural and creative products;

[0014] A construction module is used to extract the word context representation of the first user evaluation text, perform first-order and second-order feature fusion, obtain enhanced word representation, and construct a user sentiment feedback information extraction model;

[0015] A calculation module, used to process the second user evaluation text based on the user emotion feedback information extraction model, obtain feature importance weights and evaluation vocabulary weight values, and calculate the design parameters to be optimized;

[0016] A partitioning module, used to perform grid partitioning in a multidimensional design space based on the design parameters to be optimized, obtain a set of design points, and generate a set of diversified optimization design solutions based on the set of design points;

[0017] A prediction module, used for inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions;

[0018] The output module is used to iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and output the target design solution evaluation model and the target cultural and creative product design solution when the termination condition is met.

[0019] The third aspect of the present invention provides an iterative design device for cultural and creative products based on user emotional feedback, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the iterative design device for cultural and creative products based on user emotional feedback executes the above-mentioned iterative design method for cultural and creative products based on user emotional feedback.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned iterative design method of cultural and creative products based on user emotional feedback.

[0021] The technical solution provided by the present invention improves the accuracy and comprehensiveness of user emotional feedback information extraction. A fine-grained emotional dictionary is constructed through multi-feature fusion and cluster analysis, and enhanced word representation is obtained by combining first-order and second-order feature fusion, so that the emotional information extracted from the user evaluation text is more accurate and comprehensive. Intelligent optimization of design parameters is achieved. Based on the user emotional feedback information extraction model, the feature importance weight and the evaluation vocabulary weight value are calculated, and combined with the feature combination effect analysis, the intelligent optimization of design parameters is achieved, so that the optimized design scheme is more in line with user needs. The efficiency and diversity of design scheme generation are improved. Through the grid division of multi-dimensional design space and the adaptive Gaussian mutation strategy, a diverse set of optimized design schemes can be quickly generated, which greatly improves the design efficiency and ensures the innovation of the scheme. The accuracy of design scheme evaluation is enhanced. The design scheme evaluation model with a multi-layer perceptron structure is used, combined with iterative optimization and user feedback, and the evaluation model is continuously updated and improved, so that the score prediction of the design scheme is more accurate. The automation and intelligence of the design process are realized. The entire design method covers the entire process from user sentiment analysis to solution generation, evaluation and optimization, realizing the automation and intelligence of cultural and creative product design, greatly reducing manual intervention and improving design efficiency. It also enhances the matching degree between design solutions and user needs. Through the multi-task learning framework, sentiment analysis and keyword extraction are performed simultaneously, combined with feature correlation analysis and fuzzy hierarchical analysis method, so that the final generated design solution is more in line with the user's emotional needs and preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A flowchart of an iterative design method for cultural and creative products based on user emotional feedback provided in an embodiment of the present application;

[0024] Figure 2 A schematic block diagram of the structure of an iterative design system for cultural and creative products based on user emotional feedback provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0027] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0028] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0029] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0030] See also Figure 1 , Figure 1 A flow chart of an iterative design method for cultural and creative products based on user emotional feedback provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the iterative design method for cultural and creative products based on user emotional feedback provided in the embodiment of the present application includes steps S100 to S600.

[0031] Step S100: performing multi-feature fusion on the first user evaluation text to obtain a sentiment seed word set and performing cluster analysis to obtain a fine-grained sentiment dictionary in the field of cultural and creative products;

[0032] It is understandable that the execution subject of the present invention may be a cultural and creative product iterative design system based on user emotional feedback, or may be a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0033] Specifically, the first user evaluation text is segmented and POS tagged to generate a list of words. In this process, the words are identified and their POS are tagged, so that the grammatical roles and functions of different words in the sentence can be distinguished during the analysis process. Each word in the word list is vectorized, and each word is converted into a vector representation in a high-dimensional space through natural language processing technology, so that the computer can understand the semantics and mutual relationship of the words through vector mathematical operations. Multidimensional feature extraction is performed on the word vector, including calculating word frequency statistical features, sentiment intensity features and semantic features of word vectors. These three types of features reflect the frequency of occurrence of words in the text, the strength of the emotions they carry, and the semantic connotation of the words. By integrating these different dimensions of information, the sentiment and semantic features of each word are captured more comprehensively. The multidimensional feature vectors are weighted and fused to obtain a comprehensive score for each word, so as to measure the importance of the word in the expression of emotion. According to the preset threshold, the comprehensive score is screened, and the words with higher scores are retained as the initial sentiment seed words. At the same time, the emoticons in the first user's evaluation text are identified. Emoticons often carry strong emotional meanings in modern social expressions. The identification of emoticons can further enrich the emotional information library. The words that co-occur with emoticons are extracted to obtain multiple expression-related words. The initial emotional seed words are merged with the expression-related words to obtain the emotional seed word set. The words in the emotional seed word set are hierarchically clustered. By analyzing the similarities and correlations between these words, they are divided into different emotional categories to form a structured fine-grained emotional dictionary.

[0034] Step S200, extracting the word context representation of the first user evaluation text, performing first-order and second-order feature fusion to obtain enhanced word representation, and constructing a user sentiment feedback information extraction model;

[0035] Specifically, the first user evaluation text is encoded based on a pre-trained language model. The pre-trained language model is usually a deep learning model, such as BERT or GPT. These models can understand the contextual relationship between words through large-scale corpus training. In this process, each word is encoded into a context representation vector, which contains the semantic information of the word itself and reflects the change of its meaning in the context. It can capture the polysemy of the word and its emotional meaning in different contexts. The context representation vector of each word is linearly transformed to convert the original context representation vector into a first-order feature vector. By performing an outer product operation on the context representation vector, a second-order feature matrix is ​​generated to capture more complex correlation features between words and model the relationship between words. The first-order feature vector and the second-order feature matrix are concatenated to form an enhanced word representation. Based on the enhanced word representation, a multi-task learning framework is constructed. The multi-task learning framework includes two main tasks: sentiment analysis task and keyword extraction task. These two tasks share the underlying enhanced word representation and can complement each other to improve the overall effect of the model. In this framework, the sentiment analysis task uses a classifier based on the attention mechanism. Through the attention mechanism, the model can focus on the key sentiment words in the evaluation text and more accurately judge the sentiment tendency of the text. At the same time, the keyword extraction task uses the task target layer of sequence annotation. The model can automatically identify the important keywords in the evaluation text and annotate them. The entire multi-task learning framework is optimized through end-to-end training. The end-to-end training method can adjust the parameters of the sentiment analysis task and the keyword extraction task at the same time, so that the two tasks promote each other on the basis of sharing the underlying features. During the training process, the loss functions of the sentiment analysis task and the keyword extraction task are combined in a weighted manner to form an overall objective function. By adjusting the weight coefficient, the balance of the model when processing different tasks is ensured to avoid over-optimization or under-optimization of a certain task. After continuous training and optimization, the model can generate an efficient user sentiment feedback information extraction model, which can accurately extract sentiment information and keywords in user evaluations.

[0036] Step S300: Process the second user evaluation text based on the user sentiment feedback information extraction model to obtain feature importance weights and evaluation vocabulary weight values, and calculate the design parameters to be optimized;

[0037] Specifically, the second user evaluation text is processed, including word segmentation, stop word removal, and part-of-speech tagging. In this stage, irrelevant words in the text are filtered out, words with actual semantic information are retained, and they are classified to obtain the processed evaluation text. The processed evaluation text is analyzed based on the user emotional feedback information extraction model, and the sentiment analysis results and keyword extraction results in the text are automatically extracted. The sentiment analysis results can identify the user's emotional tendency towards the product, while the keyword extraction results help find the most representative words in the user evaluation. Based on the sentiment analysis and keyword extraction results, a two-dimensional evaluation matrix of the evaluation grid is constructed. In this matrix, the horizontal axis represents the product feature dimension, and the vertical axis represents the user demand dimension, forming a framework that can reflect the relationship between user emotion and product features. The sentiment intensity is assigned to each cell in the two-dimensional evaluation matrix. According to the strength of the user's emotion, a specific intensity value is assigned to the emotional information in each cell to generate an emotional intensity matrix. By calculating the average sentiment intensity of each feature dimension in the sentiment intensity matrix, the initial importance weight is obtained, which preliminarily reflects the importance of each product feature to the user's emotional feedback. Based on the fuzzy analytic hierarchy process, a fuzzy judgment matrix is ​​constructed, and fuzzy operations are performed on the fuzzy judgment matrix to calculate the evaluation vocabulary weight value, which reflects the importance of keywords in user evaluation. The feature importance weight of each product feature is calculated by combining the initial importance weight and the evaluation vocabulary weight value. The feature correlation analysis is performed on the sentiment intensity matrix to obtain the feature combination effect matrix, which reflects the correlation and interaction between product features. Combining the feature importance weight, the evaluation vocabulary weight value and the feature combination effect matrix, the comprehensive attractiveness score of each product feature is calculated to measure its importance in product design and its attractiveness to users. The feature attractiveness vector is constructed based on the comprehensive attractiveness score, and the threshold screening method is used to determine which features need to be optimized, and finally the design parameters to be optimized are obtained.

[0038] Step S400, gridding is performed in a multidimensional design space based on the design parameters to be optimized to obtain a set of design points, and a set of diversified optimization design solutions is generated according to the set of design points;

[0039] Specifically, a multidimensional design space is constructed according to the design parameters to be optimized. In this design space, the value range of each design parameter to be optimized is divided into multiple discrete intervals. By equally dividing these intervals, an initial grid structure is obtained, which contains different combinations of design parameters, so that the design space is systematically divided into multiple sub-regions, each of which represents a potential design solution. The Euclidean distance between adjacent grid points in the initial grid structure is calculated. The Euclidean distance reflects the distance between the design parameter changes in the design space and is an important indicator for measuring the similarity between different design solutions. According to the Euclidean distance, a grid spacing vector is generated. In order to better optimize the design solution, when the initial grid structure is adaptively adjusted, the sensitivity of different regions in the design space is paid attention to. For high-sensitivity regions, it usually means that a small change in the design parameter may have a large impact on the user experience. Therefore, the density of the grid is increased in these regions, so that the subtle changes of different design parameters can be captured more finely. In the low-sensitivity region, the grid density is appropriately reduced, thereby optimizing the allocation of computing resources during the design process. In this way, an optimized grid structure is obtained. In the optimized grid structure, each grid point is encoded to generate a set of design points. The design point set is actually a different instantiation of various design parameter combinations, and each design point represents a possible product design solution. Based on the design point set, the initial design solution is randomly selected to form the initial design solution set. In order to evaluate the pros and cons of the initial design solution, the weighted Euclidean distance between each solution in the initial design solution set and the user preference interval is calculated based on the feature attractiveness vector to generate a solution distance matrix. This matrix quantifies the degree of match between each design solution and the user preference. The smaller the distance, the more it meets the user's needs. Based on the solution distance matrix, the initial design solution set is sorted to screen out the optimal solution and the suboptimal solution. The adaptive Gaussian mutation strategy is applied to mutate the initial design solution to generate a diversified set of optimized design solutions. The adaptive Gaussian mutation strategy is a method for dynamically adjusting the mutation intensity. As the number of iterations increases, the mutation intensity will gradually weaken. The initial mutation is strong, allowing a large adjustment of the design parameters. As the number of iterations increases, the mutation gradually becomes more refined to ensure that the final design solution can fully meet the user's needs and does not deviate from the target design direction. Through the above steps, the final generated diversified set of optimized design solutions can better meet the user's emotional feedback and actual needs.

[0040] Step S500, inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions;

[0041] Specifically, the diversified optimized design scheme set is input into the initial design scheme evaluation model, which adopts a multi-layer perceptron structure. The multi-layer perceptron model consists of an input layer, three hidden layers and an output layer, and its purpose is to predict the score of each design scheme. The input layer receives the feature vector of the design scheme, whose dimension is N, that is, the number of design parameters. Each feature vector contains the values ​​of all design parameters in the design scheme. The data of the input layer will be passed to the three hidden layers in sequence, with the number of neurons being 128, 64 and 32 respectively, so that the model can extract the features of the design scheme layer by layer, and gradually compress the feature representation to extract the core information related to the design evaluation. In each hidden layer, the ReLU (linear rectification) activation function is used to ensure that the model can capture more complex design features during the nonlinear transformation process. At the same time, a Dropout layer is added after each hidden layer, and the dropout rate is set to 0.3 to prevent the model from overfitting. Dropout makes the model more generalized by randomly "discarding" a part of the neurons, and can maintain good prediction accuracy when dealing with new design schemes. The output layer is composed of a single neuron, using the Sigmoid activation function, which maps the output of the model to a range between 0 and 1, indicating the predicted score of each design solution. The closer the score is to 1, the more the solution meets the user's emotional feedback and design goals. Feature extraction is performed on each design solution in the diversified optimization design solution set to obtain an N-dimensional design solution feature vector. All feature vectors are combined to form a design solution feature matrix. The dimension of the matrix is ​​M×N, where M represents the number of design solutions and N represents the feature dimension of each solution. The design solution feature matrix will be used as the input of the input layer. The feature matrix is ​​calculated by the first hidden layer to obtain a 128-dimensional intermediate feature representation, and the preliminary features of the solution are extracted through the connection of neurons; the 128-dimensional intermediate feature representation is input into the second hidden layer, and a 64-dimensional high-level feature representation is calculated and extracted, which contains important patterns and information in the solution design; the 64-dimensional high-level feature representation is processed by the third hidden layer and finally compressed into a 32-dimensional final feature representation, which contains the most critical summary information for the design solution score prediction. The final 32-dimensional feature representation is input to the output layer, and a prediction score between 0 and 1 is generated for each design solution through the Sigmoid activation function. These scores constitute an M-dimensional prediction score vector, which represents the scoring results of all design solutions. Based on this score vector, the diversified optimization design solution set is sorted in descending order, and the top K design solutions with the highest scores are selected to form a high-scoring solution set. K-means clustering analysis is performed on the high-scoring solution set to cluster similar design solutions together to help identify the diverse features in the design solution set. In each cluster center, the design solution closest to the cluster center and with the highest score is selected. These solutions can represent the optimal solutions in each design feature direction. The cluster center solutions with high scores are selected to form a set of candidate design solutions.

[0042] Step S600: iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and when the termination condition is met, output the target design solution evaluation model and the target cultural and creative product design solution.

[0043] Specifically, parameter encoding is performed on each scheme in the candidate design scheme set, and it is converted into a quantifiable scheme parameter matrix. Based on the scheme parameter matrix, the Euclidean distance between different design schemes is calculated to construct a scheme distance matrix. The Euclidean distance reflects the degree of difference between the design schemes and can help identify the similarity between the schemes. The candidate design scheme set is clustered using the scheme distance matrix, and similar design schemes are clustered together to obtain the scheme clustering results. Representative design schemes are selected from the scheme clustering results, and these schemes can represent the core characteristics of each cluster. Crossover and mutation operations are performed on the representative schemes to generate a new generation of design scheme sets. The crossover operation realizes the exchange of characteristics between schemes and enhances the diversity of schemes by combining the parameters of different design schemes; while the mutation operation introduces innovative design possibilities by randomly perturbing some design parameters, so that the design schemes have greater optimization space in subsequent iterations. The new generation of design scheme sets are input into the initial design scheme evaluation model for score prediction. The model scores each new scheme based on the design parameters and previous training results. Based on the predicted scores of the schemes, the new generation of design scheme sets are sorted, and the highest-scoring schemes are selected as elite individuals to form an elite scheme set. The elite solution set represents the design solutions with the best performance and the most potential in the current stage, with higher user satisfaction and design quality. User feedback is collected on the elite solution set, and users score each elite solution through actual experience or evaluation to generate user rating data. Based on the user's rating information, the parameters of the initial design solution evaluation model are updated so that the model can more accurately predict the rating of future design solutions, and obtain an updated design solution evaluation model that is more accurate and adaptable to user needs. During the model update, it is determined whether the current iteration process meets the termination condition. Generally, the termination condition can be that the number of iterations has reached a predetermined upper limit, or the rating of the design solution has stabilized after several iterations and has not been significantly improved. If the termination condition is not met, a new iteration is continued, and the above-mentioned cycle of crossover, mutation, evaluation and user feedback is repeated until the termination condition is met. When the termination condition is met, the solution with the highest score and that meets all design constraints is selected from the elite solution set as the final target cultural and creative product design solution, and the updated design solution evaluation model is determined as the target design solution evaluation model.

[0044] In the embodiment of the present invention, the accuracy and comprehensiveness of the extraction of user emotional feedback information are improved. A fine-grained emotional dictionary is constructed by multi-feature fusion and cluster analysis, and enhanced word representation is obtained by combining the fusion of first-order and second-order features, so that the emotional information extracted from the user evaluation text is more accurate and comprehensive. Intelligent optimization of design parameters is achieved. Based on the user emotional feedback information extraction model, the feature importance weight and the evaluation vocabulary weight value are calculated, and combined with the feature combination effect analysis, the intelligent optimization of design parameters is achieved, so that the optimized design scheme is more in line with user needs. The efficiency and diversity of design scheme generation are improved. Through the grid division of the multi-dimensional design space and the adaptive Gaussian mutation strategy, a diverse set of optimized design schemes can be quickly generated, which greatly improves the design efficiency and ensures the innovation of the scheme. The accuracy of design scheme evaluation is enhanced. The design scheme evaluation model with a multi-layer perceptron structure is used, combined with iterative optimization and user feedback, and the evaluation model is continuously updated and improved, so that the score prediction of the design scheme is more accurate. The automation and intelligence of the design process are realized. The entire design method covers the entire process from user sentiment analysis to solution generation, evaluation and optimization, realizing the automation and intelligence of cultural and creative product design, greatly reducing manual intervention and improving design efficiency. It also enhances the matching degree between design solutions and user needs. Through the multi-task learning framework, sentiment analysis and keyword extraction are performed simultaneously, combined with feature correlation analysis and fuzzy hierarchical analysis method, so that the final generated design solution is more in line with the user's emotional needs and preferences.

[0045] In a specific embodiment, the process of executing step S100 may specifically include the following steps:

[0046] Perform word segmentation and part-of-speech tagging on the first user evaluation text to obtain a word list, and perform vectorization processing on each word in the word list to obtain a word vector;

[0047] Calculate the word frequency statistical features, sentiment intensity features and word vector semantic features for each word in the word list to obtain a multi-dimensional feature vector;

[0048] Perform weighted fusion on multi-dimensional feature vectors to obtain a comprehensive score, and filter the comprehensive score according to a preset threshold to obtain multiple initial sentiment seed words;

[0049] Recognize the emoticons in the first user's evaluation text, and extract words that co-occur with the emoticons to obtain a plurality of emoticon-related words;

[0050] Multiple initial sentiment seed words are merged with multiple expression-related words to obtain a sentiment seed word set, and the words in the sentiment seed word set are hierarchically clustered to obtain a fine-grained sentiment dictionary in the field of cultural and creative products.

[0051] Specifically, the first user evaluation text is segmented and POS tagged. Word segmentation is to divide the continuous text stream into independent word units, while POS tagging is to assign each word its grammatical role in the sentence, such as noun, verb, adjective, etc. With the help of natural language processing tools, such as Jieba word segmentation library or NLTK, the text is converted into a word list. Each word in the word list is vectorized to map the words in the text into a multidimensional space so that it can be used in mathematical calculations. Word vectorization methods usually include TF-IDF methods based on co-occurrence matrices or word embedding methods based on neural networks such as Word2Vec or GloVe. For example, using Word2Vec, each word in the text is mapped to an N-dimensional vector space, where the vector representation of each word captures the semantic information of the word. For word vectors , assuming its dimension is N, each word vector can be represented as an N-dimensional vector , where each Represents the value of a word in the jth dimension. Three types of features are calculated for each word, namely, word frequency statistical features, sentiment intensity features, and semantic features of word vectors. Word frequency statistical features are based on the frequency of occurrence of words in the text. Suppose word The number of times it appears in the text is , the total number of words is , then the word frequency statistical features can be expressed as:

[0052] ;

[0053] The word frequency feature can reflect the importance of words in the evaluation text. The sentiment intensity feature is a sentiment score assigned to each word based on existing sentiment dictionaries (such as NRC or SentiWordNet). Each word may have different sentiment dimensions, such as positive sentiment, negative sentiment, anger, happiness, etc. The word vector semantic feature captures the semantic relevance information of words through its word vector representation. These three types of features are combined to generate a multidimensional feature vector for each word. The weighted fusion method is used to assign weights to different features according to their importance, and a comprehensive score is calculated. The weighted fusion method is used to assign weights to different features according to their importance, and a comprehensive score is calculated. Assume that the weights of each type of feature are , then the comprehensive score of each word can be expressed as:

[0054]

[0055] in, The modulus of the word vector, reflecting the weight of the semantic information of the word. By adjusting different weights, the importance of each word in sentiment analysis is measured. The comprehensive score is screened based on the preset threshold, and those words with a comprehensive score higher than the threshold are retained as initial sentiment seed words. Identify emoticons in the text. Common emoticons include smiley faces, anger, etc. Each emoticon conveys a certain emotional meaning. After identifying the emoticons in the text through sentiment analysis tools or predefined expression and emotion mapping libraries, extract the words that co-occur with these emoticons. After obtaining the initial sentiment seed words and expression-related words, merge these two types of words to form a more comprehensive set of sentiment seed words. Perform hierarchical clustering analysis on the words, gradually aggregate similar words, form a multi-level hierarchical structure of sentiment words, and obtain a fine-grained sentiment dictionary in the field of cultural and creative products.

[0056] In a specific embodiment, the process of executing step S200 may specifically include the following steps:

[0057] Encode the first user evaluation text based on the pre-trained language model to obtain a context representation vector for each word;

[0058] Perform a linear transformation on the context representation vector to obtain a first-order feature vector, and perform an outer product operation on the context representation vector to obtain a second-order feature matrix;

[0059] Concatenate the first-order feature vector and the second-order feature matrix to obtain the enhanced word representation;

[0060] Construct a multi-task learning framework, which includes sentiment analysis tasks and keyword extraction tasks. The sentiment analysis tasks and keyword extraction tasks share the underlying features of enhanced word representation. The sentiment analysis task uses a classifier based on the attention mechanism, and the keyword extraction task uses a task target layer with sequence annotation.

[0061] The parameters of the multi-task learning framework are optimized through end-to-end training to obtain a user sentiment feedback information extraction model. The end-to-end training includes combining the loss functions of the sentiment analysis task and the keyword extraction task in a weighted manner to form an overall objective function.

[0062] Specifically, the first user evaluation text is encoded based on the pre-trained language model, and a corresponding context representation vector is generated for each word. The pre-trained language model (such as BERT, GPT, etc.) has been pre-trained on a large amount of corpus and can well capture the semantic information in the context. The context representation vector of each word not only contains the semantic information of the word, but also includes the dependency relationship with other words in the context. The context representation vector can be used Indicates that For words The context in the text is represented by a vector. Assume that the dimension of each vector is ,but Represents the deep semantic information of the word in the context. Feature extraction is performed on the context representation vector, which is specifically divided into first-order features and second-order features. The context representation vector is converted into a first-order feature vector using linear transformation. Each vector is linearly transformed, namely:

[0063] ;

[0064] in, is the weight matrix of the linear transformation, is the bias term, It is a first-order feature vector, representing the word The basic feature extraction of It still contains the semantic information of the words, but retains the core features in the dimension reduction while reducing the model complexity. At the same time, the second-order feature matrix is ​​extracted by performing an outer product operation on the context representation vector. The outer product operation is used to capture the bilinear relationship between words and can describe the interactive information between words and other words in the semantic space. For each word's context representation vector , its second-order characteristic matrix can be calculated as follows:

[0065] ;

[0066] in, represents the outer product operation, is a second-order feature matrix. The second-order feature matrix not only contains the semantic information of the word itself, but also reflects the interactive relationship between the word and the context. The first-order feature vector is concatenated with the second-order feature matrix to generate an enhanced word representation. The enhanced word representation combines basic word features with more complex interactive features, enabling the model to more accurately extract sentiment and keyword information in subsequent tasks. The enhanced word representation can be expressed as:

[0067] ;

[0068] Among them, flatten(·) means flattening the second-order feature matrix into a one-dimensional vector. The final enhanced word representation contains all the features of the word. A multi-task learning framework is constructed. The framework handles sentiment analysis tasks and keyword extraction tasks at the same time, and both share the underlying enhanced word representation. The goal of the sentiment analysis task is to judge the user's emotional inclination towards cultural and creative products by classifying the overall semantics of the text. The sentiment analysis task uses a classifier based on the attention mechanism. The attention mechanism can focus on the most critical words in the text and improve the accuracy of sentiment classification. The output of the attention mechanism can be expressed as:

[0069] ;

[0070] in, is the attention weight matrix, is the attention weight of the word, indicating the degree of attention the model pays to each word. The keyword extraction task uses sequence labeling to identify key terms in the text and mark out important words in the evaluation text. The keyword extraction task is completed by a sequence labeling algorithm (such as CRF or Bi-LSTM-CRF). The sequence labeling task relies on enhancing the underlying features of word representation, combining contextual information and word features to determine whether each word is a keyword. The entire multi-task learning framework is optimized through end-to-end training. End-to-end training means that the parameters of the sentiment analysis task and the keyword extraction task are updated in the same process. During the training process, an overall objective function is constructed, which combines the loss functions of the two tasks. Suppose the loss function of the sentiment analysis task is , the loss function of the keyword extraction task is , then the overall objective function can be expressed as:

[0071] ;

[0072] in, and is the weight coefficient of the two tasks, which is used to balance the contribution of different tasks to model optimization. By weighting, the loss of the two tasks is ensured to be reasonably optimized during the model training process. During the model training process, sentiment analysis and keyword extraction tasks are carried out simultaneously. By sharing the underlying features of enhanced word representation, the model can improve the accuracy of sentiment analysis while improving the performance of keyword extraction. After multiple iterations of training, the model gradually learns how to accurately extract sentiment feedback and key terms from different user reviews.

[0073] In a specific embodiment, the process of executing step S300 may specifically include the following steps:

[0074] Performing word segmentation, removing stop words and part-of-speech tagging on the second user's evaluation text to obtain a processed evaluation text;

[0075] The processed evaluation text is processed based on the user sentiment feedback information extraction model to obtain sentiment analysis results and keyword extraction results;

[0076] A two-dimensional evaluation matrix of the evaluation grid is constructed based on the sentiment analysis results and keyword extraction results, where the horizontal axis of the two-dimensional evaluation matrix represents the product feature dimension and the vertical axis represents the user demand dimension;

[0077] Assign sentiment intensity to each cell in the two-dimensional evaluation matrix to obtain a sentiment intensity matrix, and calculate the average sentiment intensity of each feature dimension in the sentiment intensity matrix to obtain the initial importance weight;

[0078] Based on the fuzzy analytic hierarchy process, a fuzzy judgment matrix is ​​constructed, and fuzzy operations are performed on the fuzzy judgment matrix to obtain the evaluation vocabulary weight value, and based on the initial importance weight and the evaluation vocabulary weight value, the feature importance weight is calculated;

[0079] Perform feature correlation analysis on the sentiment intensity matrix to obtain the feature combination effect matrix, and combine the feature importance weight, evaluation word weight value and feature combination effect matrix to calculate the comprehensive attractiveness score of each product feature;

[0080] A characteristic attraction vector is constructed based on the comprehensive attraction score, and a threshold screening method is used to determine the design parameters to be optimized according to the characteristic attraction vector.

[0081] Specifically, the second user evaluation text is preprocessed, including word segmentation, stop word removal and part-of-speech tagging, to obtain the processed evaluation text. Word segmentation is to split the continuous text stream into independent word units; stop words are those words that have little effect on the overall meaning of the text. After removing the stop words, the key information of the text will be more prominent. Part-of-speech tagging is to assign grammatical roles to each word, such as nouns, verbs, adjectives, etc., which is helpful for subsequent keyword extraction and sentiment analysis. The processed evaluation text is processed based on the user sentiment feedback information extraction model, and the model will perform sentiment analysis and keyword extraction on the text. The sentiment analysis results will indicate the sentiment tendency of the user evaluation. Based on the sentiment analysis results and keyword extraction results, a two-dimensional evaluation matrix of the evaluation grid is constructed. The horizontal axis of the two-dimensional evaluation matrix represents the product feature dimension, and the vertical axis represents the user demand dimension. In the constructed two-dimensional evaluation matrix, the sentiment intensity is assigned to each cell to form a sentiment intensity matrix. The sentiment intensity assignment is based on the sentiment analysis results, and the sentiment information in each cell is quantified into a specific value. The average sentiment intensity of each feature dimension in the sentiment intensity matrix is ​​calculated to obtain the initial importance weight. In order to improve the accuracy of feature weights, the fuzzy analytic hierarchy process is used to construct a fuzzy judgment matrix. The fuzzy analytic hierarchy process is a multi-criteria decision-making method that combines fuzzy logic with the analytic hierarchy process to handle uncertainty and fuzzy information. The fuzzy judgment matrix is ​​used to compare the importance of different features. The weight values ​​of the evaluation words are obtained by performing fuzzy operations on the fuzzy judgment matrix. The elements of the fuzzy judgment matrix are: Representation characteristics Relative to the feature The importance of each keyword is calculated through fuzzy operation. The feature importance weight is calculated based on the initial importance weight and the evaluation vocabulary weight value. The formula is as follows:

[0082] ;

[0083] in, Representation characteristics The final importance weight of is the initial importance weight of the feature, is the evaluation vocabulary weight value, and are the balance coefficients of the two weight items respectively. Through this formula, the sentiment intensity in sentiment analysis is combined with the evaluation weight in keyword extraction to obtain a more accurate feature importance assessment. Feature correlation analysis is performed on the sentiment intensity matrix to obtain the feature combination effect matrix. The feature combination effect matrix is ​​used to measure the correlation between different features. For example, "sound quality" and "comfort" may have a strong correlation with the user experience of headphones, so the combination effect between them is higher. By analyzing the correlation between features, we can better understand how to optimize multiple feature combinations to improve user satisfaction. Combining the feature importance weights, evaluation vocabulary weight values ​​and feature combination effect matrix, the comprehensive attractiveness score of each product feature is calculated. The comprehensive attractiveness score reflects the overall influence of each feature on the user experience. The formula is as follows:

[0084] ;

[0085] in, It is a feature The overall attractiveness score of is the feature importance weight, It is a feature and Features The combined effect weight between It is a feature The formula not only considers the importance of a single feature, but also integrates the synergy between features. Based on the comprehensive attractiveness score, a feature attractiveness vector is constructed, and the design parameters to be optimized are determined by the threshold screening method. The threshold screening method sets a minimum attractiveness threshold value, and only features that exceed this threshold are considered design elements that need to be optimized. For example, if the comprehensive attractiveness score of a feature is higher than 0.7, it is considered that the feature has great optimization potential, otherwise it will not be optimized for the time being. Through this process, emotional feedback information is extracted from user reviews and converted into quantifiable feature attractiveness, providing a basis for the optimization of cultural and creative product design.

[0086] In a specific embodiment, the process of executing step S400 may specifically include the following steps:

[0087] A multidimensional design space is constructed according to the design parameters to be optimized, and the value range of each design parameter is equally divided to obtain an initial grid structure;

[0088] The Euclidean distance between adjacent grid points in the initial grid structure is calculated to obtain a grid spacing vector, and the initial grid structure is adaptively adjusted based on the grid spacing vector, the grid density is increased in the high-sensitivity area, and the optimized grid structure is obtained;

[0089] Encode each grid point in the optimized grid structure to obtain a set of design points;

[0090] An initial design scheme is randomly selected based on the design point set to obtain an initial design scheme set, and a weighted Euclidean distance between each scheme in the initial design scheme set and the user preference interval is calculated based on the characteristic attractiveness vector to obtain a scheme distance matrix;

[0091] The initial design scheme set is sorted based on the scheme distance matrix, and an adaptive Gaussian mutation strategy is applied to generate a diversified optimization design scheme set, where the mutation intensity of the adaptive Gaussian mutation strategy is dynamically adjusted with the number of iterations.

[0092] Specifically, a multidimensional design space is constructed based on the design parameters to be optimized. Each design parameter has a range of values, and the range of values ​​of each parameter is defined according to market demand or user feedback. The range of values ​​is divided equally, and the value of each parameter is divided into multiple discrete points to form a grid structure in the design space, where each grid point represents a combination of design solutions. The Euclidean distance between adjacent grid points in the initial grid structure is calculated to measure the similarity of different design parameter combinations. The Euclidean distance is a standard method for measuring the straight-line distance between two points in a space. For two design solutions A and B, assume that their design parameters are and ,in is the number of design parameters, then the Euclidean distance between the two solutions can be expressed as:

[0093] ;

[0094] in, Represents the distance between scheme A and scheme B in the design space, and Respectively represent the scheme A and scheme B in The grid spacing vector is obtained by calculating the Euclidean distance between all adjacent grid points, which is used to represent the distance between different design points in each dimension. The initial grid structure is adaptively adjusted based on the grid spacing vector. In the design space, some areas may be more sensitive to design optimization, that is, in these areas, slight changes in design parameters will have a greater impact on user experience. Such highly sensitive areas can be judged based on historical user feedback or feature attractiveness vectors. To this end, the density of the grid is increased in the highly sensitive areas, that is, the design parameter values ​​in these areas are subdivided to capture the slight differences in design details. Correspondingly, for low-sensitivity areas, the density of the grid can be reduced and the coarser divisions can be retained. Adaptive adjustment optimizes the distribution of the design space so that computing resources are concentrated in important design areas. After obtaining the optimized grid structure, each grid point is encoded and each design solution is converted into a unique vector form. The encoding of each grid point can be composed of its values ​​in each parameter dimension. By encoding all grid points, a set of design points is generated, and each element in the set is a possible design solution. Based on the set of design points, some initial design solutions are randomly selected to obtain the initial design solution set. The initial design solution set is a subset randomly selected from the entire set of design points, representing different combinations in the design space. In order to evaluate the pros and cons of these solutions, the weighted Euclidean distance between each solution and the user preference interval is calculated based on the feature attractiveness vector. The feature attractiveness vector reflects the importance of each design feature to the user, and the user preference interval represents the user's ideal value for certain parameters. The weighted Euclidean distance calculation formula is as follows:

[0095] ;

[0096] in, is the weighted Euclidean distance, is the weight in the feature attraction vector, indicating the The importance of design parameters and are design solution A and user preference interval in The weighted distance can better evaluate the degree of match between each design solution and user expectations. The initial set of design solutions is sorted based on the solution distance matrix, and arranged from high to low according to the degree of match with user preferences. After sorting, the adaptive Gaussian mutation strategy is applied to generate a diversified set of optimized design solutions. Gaussian mutation is a method of adjusting design parameters by introducing random noise. New design solutions are generated by adding random changes that follow a Gaussian distribution to each parameter of the design solution. The adaptive Gaussian mutation strategy means that the mutation intensity is dynamically adjusted with the number of iterations. Initially, the mutation intensity is large, allowing the design parameters to be adjusted to a large extent to explore the diversity of the design space; as the number of iterations increases, the mutation intensity gradually decreases, and fine optimization is concentrated in a local range. Assume that the design solution parameters The initial variation of , then The mutation at the iteration can be expressed as:

[0097] ;

[0098] in, For the The mutation intensity of the iteration, is the initial mutation intensity, is the attenuation coefficient, is the current iteration number. In this way, the mutation strategy can conduct extensive exploration in the early stage and perform local fine-tuning in the later stage to ensure the diversity and optimization effect of the scheme. Through the above steps, the final generated diversified optimization design scheme set can cover a wide range of design spaces and gradually approach the user preference range in each iteration process.

[0099] In a specific embodiment, the process of executing step S500 may specifically include the following steps:

[0100] The diversified optimized design scheme set is input into the initial design scheme evaluation model, which adopts a multi-layer perceptron structure, including an input layer, three hidden layers and an output layer; the input layer receives the feature vector of the design scheme, with a dimension of N, where N is the number of design parameters; the number of neurons in the three hidden layers is 128, 64 and 32 respectively; the output layer is a single neuron, which outputs the predicted score; the hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function to map the output to between 0 and 1; a Dropout layer is added after each hidden layer, with a dropout rate of 0.3;

[0101] Extract features from each design solution in the diversified optimization design solution set to obtain an N-dimensional design solution feature vector; combine all feature vectors into a design solution feature matrix with a dimension of M×N, where M is the number of design solutions;

[0102] The design scheme feature matrix is ​​input into the input layer of the initial design scheme evaluation model, and a 128-dimensional intermediate feature representation is obtained through calculation in the first hidden layer; the 128-dimensional intermediate feature representation is input into the second hidden layer, and a 64-dimensional high-level feature representation is obtained through calculation; the 64-dimensional high-level feature representation is input into the third hidden layer, and a 32-dimensional final feature representation is obtained through calculation;

[0103] The 32-dimensional final feature representation is input to the output layer, and the predicted score is obtained through the Sigmoid activation function to form an M-dimensional predicted score vector; based on the predicted score vector, the diversified optimization design scheme set is sorted in descending order, and the top K design schemes are selected to obtain a high-scoring scheme set;

[0104] Perform K-means clustering analysis on the high-scoring solution set, and select the solution with the highest score closest to each cluster center to form a set of candidate design solutions.

[0105] Specifically, the multilayer perceptron model is a feedforward neural network, whose structure includes an input layer, multiple hidden layers, and an output layer. In order to evaluate the design scheme, the input layer receives the feature vector of each design scheme. The feature vector of the design scheme can be expressed as:

[0106] ;

[0107] in, is the eigenvector of the ith design, represents the value of the i-th design solution at the j-th design parameter, and N is the total number of design parameters. The feature vectors of all design solutions are combined into a design solution feature matrix. Assuming there are M design solutions, the dimension of the design solution feature matrix is , which can be expressed as:

[0108] ;

[0109] The design feature matrix is ​​input to the input layer of the initial design evaluation model. The dimension of the input layer is N, which is the length of the feature vector of each design. The data flows through three hidden layers, with 128, 64, and 32 neurons respectively. The calculation results of each hidden layer are linearly transformed based on the output of the previous layer and activated by ReLU (linear rectification function). The formula of ReLU activation function is:

[0110] ;

[0111] in, It is the output of the linear transformation of the previous layer. The role of ReLU is to introduce nonlinear characteristics so that the model can learn more complex features. After passing through the first hidden layer, the features of each design solution are converted into a 128-dimensional intermediate feature representation. The 128-dimensional features are input into the second hidden layer to extract 64-dimensional high-level features. The 64-dimensional features are compressed into a 32-dimensional final feature representation through the third hidden layer. A Dropout layer is added after each hidden layer. The role of Dropout is to prevent the model from overfitting. The Dropout layer reduces the model's dependence on specific paths by randomly "discarding" a part of the neurons during the training process. It is usually set to a dropout rate of 30%, that is, 30% of the neurons are randomly set to 0 after each hidden layer calculation. In the final output layer, the model uses the Sigmoid activation function to map the output value to between 0 and 1, representing the predicted score of the design solution. The formula of the Sigmoid function is:

[0112] ;

[0113] in, is the linear output of the previous layer. The Sigmoid function compresses the input value between 0 and 1 as the score of the design solution, indicating the relative merits of each design solution. The model generates an M-dimensional predicted score vector for M design solutions:

[0114] ;

[0115] in, The predicted score of the i-th design solution is as follows: the closer the score is to 1, the more the solution meets the design goal or user needs. Based on the predicted score vector, the diversified optimization design solution set is sorted in descending order, and the top K design solutions with the highest scores are selected to form a high-scoring solution set. For example, if there are 100 design solutions, the top 10 solutions with the highest scores are selected and put into the high-scoring solution set. For the high-scoring solution set, K-means clustering analysis is used to identify representative solutions with different characteristics, and the data set is divided into K clusters, each represented by a cluster center. The algorithm optimizes the clustering results by minimizing the distance from each data point to the cluster center to which it belongs. For each design solution feature vector, the K-means algorithm performs clustering through the following steps: randomly initialize K cluster centers; calculate the Euclidean distance between each design solution and each cluster center; assign the design solution to the cluster center with the closest distance; recalculate the center of each cluster and update the cluster center; repeat the above steps until the cluster center converges. For each cluster, the design solution with the closest distance to the cluster center and the highest score is selected as the representative solution. The final set of candidate design solutions not only includes the solutions with the highest scores, but also covers different design feature areas.

[0116] In a specific embodiment, the process of executing step S600 may specifically include the following steps:

[0117] Parameter encoding is performed on each scheme in the candidate design scheme set to obtain a scheme parameter matrix, and the Euclidean distance between schemes is calculated based on the scheme parameter matrix to construct a scheme distance matrix;

[0118] Based on the scheme distance matrix, cluster analysis is performed on the candidate design scheme set to obtain the scheme clustering results, and representative schemes are selected from the scheme clustering results. Crossover and mutation operations are performed on the representative schemes to obtain a new generation of design scheme sets;

[0119] Input the new generation design solution set into the initial design solution evaluation model to obtain the solution prediction score, and sort the new generation design solution set based on the solution prediction score, select the solutions with the highest scores as elite individuals, and obtain the elite solution set;

[0120] Collect user feedback on the elite solution set to obtain user rating data, and update the parameters of the initial design solution evaluation model based on the user rating data to obtain an updated design solution evaluation model;

[0121] Determine whether the termination conditions are met. If not, continue to iterate. If so, select the solution with the highest score and that meets all constraints from the elite solution set as the target cultural and creative product design solution, and use the updated design solution evaluation model as the target design solution evaluation model.

[0122] Specifically, each design scheme in the candidate design scheme set is parameter-encoded. Each design scheme consists of several design parameters, which can be represented as a vector. For example, if the design parameters are the "size", "weight" and "color" of the product, the parameters of each design scheme can be encoded as a three-dimensional vector, represented as:

[0123] ;

[0124] in, is the parameter vector of the ith design, Indicates the value of the scheme at the jth design parameter. The parameter vectors of all candidate design schemes are combined to form a scheme parameter matrix. Assuming there are M candidate design schemes, each with N design parameters, the scheme parameter matrix can be expressed as:

[0125] ;

[0126] The Euclidean distance between each two design solutions is calculated based on the solution parameter matrix. The Euclidean distance is an important indicator used to measure the similarity between two solutions in the design space. The calculation formula is:

[0127] ;

[0128] in, Represents the distance between the i-th solution and the j-th solution in the design parameter space. By calculating the pairwise distances of all the design solutions in the candidate design solution set, the solution distance matrix is ​​obtained, and its dimension is , each element in the matrix represents the distance between two design solutions. After obtaining the solution distance matrix, cluster analysis is performed based on the matrix, and the design solutions are divided into multiple classes according to similarity to find solutions that represent different design trends. K-means clustering algorithm is a commonly used clustering method. Through iterative optimization, each solution is assigned to the nearest cluster center. The basic steps of K-means include: randomly initializing K cluster centers, calculating the distance between each solution and these centers, assigning the solution to the nearest cluster, and updating the location of the cluster center until the clustering results converge. Select representative solutions from the clustering results. Representative solutions are usually the solutions closest to each cluster center, which reflect different design trends and directions. Crossover and mutation operations are performed on representative solutions. The crossover operation simulates the reproduction process in the genetic algorithm, and generates a new solution by exchanging some design parameters of the two solutions. For example, suppose there are two solutions A and B, with parameters and respectively. After the crossover operation, a new solution may be generated, that is, the new solution inherits some features of A and B. At the same time, the mutation operation introduces diversity by randomly changing the values ​​of certain design parameters. This random change is usually carried out within a certain range, such as slightly adjusting the parameters to new values ​​to generate new design solutions. After crossover and mutation operations, a new generation of design solution sets is formed. The new generation of design solution sets are input into the initial design solution evaluation model to obtain the predicted score of each solution. The initial design solution evaluation model scores the input design parameters through the previously trained multi-layer perceptron (MLP) structure. The output of the model is the predicted score of each design solution, which can be used to measure the pros and cons of the solutions. Based on the predicted scores, the new generation of design solution sets are sorted, and several solutions with the highest scores are selected as elite individuals to form the elite solution set. The elite solution set represents the best design solutions in the current iteration, which perform well in terms of the combination of design parameters and user satisfaction. On this basis, user feedback is collected for the elite solution set. These solutions are displayed to users, and user rating data is collected. The rating data fed back by users can reflect the real user preferences and experience, and provide information with more practical reference value than model predictions. Based on user feedback data, the parameters of the initial design solution evaluation model are updated. The evaluation model is retrained using user rating data to update the weight parameters in the model so that it can better predict the ratings of future design solutions. In this way, the evaluation model gradually becomes more accurate and better adapted to actual user needs. In each iteration, it is necessary to determine whether the termination condition is met. The termination condition can be set based on the number of iterations, the convergence of the model score, or the change in user feedback. If the termination condition is not met, the next round of iteration is continued, that is, the new set of design solutions is input into the evaluation model, a new set of elite solutions is generated, and the model parameters are updated.If the termination conditions are met, the solution with the highest score and meeting all design constraints will be selected from the elite solution set as the final target cultural and creative product design solution.

[0129] See also Figure 2 , Figure 2 A schematic block diagram of the structure of a cultural and creative product iterative design system 200 based on user emotional feedback provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the iterative design system 200 for cultural and creative products based on user emotional feedback includes:

[0130] A fusion module 210 is used to perform multi-feature fusion on the first user evaluation text to obtain a sentiment seed word set and perform cluster analysis to obtain a fine-grained sentiment dictionary in the field of cultural and creative products;

[0131] A construction module 220 is used to extract the word context representation of the first user evaluation text, perform first-order and second-order feature fusion, obtain enhanced word representation, and construct a user sentiment feedback information extraction model;

[0132] A calculation module 230 is used to process the second user evaluation text based on the user sentiment feedback information extraction model, obtain feature importance weights and evaluation vocabulary weight values, and calculate the design parameters to be optimized;

[0133] A partitioning module 240 is used to perform grid partitioning in the multidimensional design space based on the design parameters to be optimized, obtain a set of design points, and generate a set of diversified optimization design solutions based on the set of design points;

[0134] Prediction module 250, used for inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions;

[0135] The output module 260 is used to iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and output the target design solution evaluation model and the target cultural and creative product design solution when the termination condition is met.

[0136] Through the collaboration of the above components, the accuracy and comprehensiveness of user emotional feedback information extraction are improved. A fine-grained emotional dictionary is constructed through multi-feature fusion and cluster analysis, and enhanced word representation is obtained by combining the fusion of first-order and second-order features, making the emotional information extracted from the user evaluation text more accurate and comprehensive. Intelligent optimization of design parameters is achieved. Based on the user emotional feedback information extraction model, the feature importance weights and evaluation vocabulary weight values ​​are calculated, and combined with the feature combination effect analysis, the intelligent optimization of design parameters is achieved, so that the optimized design scheme is more in line with user needs. The efficiency and diversity of design scheme generation are improved. Through the grid division of multi-dimensional design space and the adaptive Gaussian mutation strategy, a diverse set of optimized design schemes can be quickly generated, which greatly improves the design efficiency and ensures the innovation of the scheme. The accuracy of design scheme evaluation is enhanced. The design scheme evaluation model with a multi-layer perceptron structure is combined with iterative optimization and user feedback to continuously update and improve the evaluation model, making the score prediction of the design scheme more accurate. The automation and intelligence of the design process are achieved. The entire design method covers the entire process from user sentiment analysis to solution generation, evaluation and optimization, realizing the automation and intelligence of cultural and creative product design, greatly reducing manual intervention and improving design efficiency. It also enhances the matching degree between design solutions and user needs. Through the multi-task learning framework, sentiment analysis and keyword extraction are performed simultaneously, combined with feature correlation analysis and fuzzy hierarchical analysis method, so that the final generated design solution is more in line with the user's emotional needs and preferences.

[0137] The present application also provides an iterative design device for cultural and creative products based on user emotional feedback. The iterative design device for cultural and creative products based on user emotional feedback includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the iterative design method for cultural and creative products based on user emotional feedback in the above-mentioned embodiments.

[0138] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the iterative design method for cultural and creative products based on user emotional feedback.

[0139] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0140] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0141] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for iterative design of cultural and creative products based on user emotional feedback, characterized in that: include: The first user evaluation text is subjected to multi-feature fusion to obtain a sentiment seed word set and cluster analysis to obtain a fine-grained sentiment dictionary in the field of cultural and creative products; specifically, the method includes: performing word segmentation and part-of-speech tagging on the first user evaluation text to obtain a word list, and performing vectorization processing on each word in the word list to obtain a word vector; calculating the word frequency statistical features, sentiment intensity features and word vector semantic features of each word in the word list to obtain a multi-dimensional feature vector; performing weighted fusion on the multi-dimensional feature vector to obtain a comprehensive score, and filtering the comprehensive score according to a preset threshold to obtain a plurality of initial sentiment seed words; identifying emoticons in the first user evaluation text, and extracting words co-occurring with emoticons to obtain a plurality of expression-related words; merging the plurality of initial sentiment seed words with the plurality of expression-related words to obtain a sentiment seed word set, and performing hierarchical clustering on the words in the sentiment seed word set to obtain a fine-grained sentiment dictionary in the field of cultural and creative products; Extracting the word context representation of the first user evaluation text, performing first-order and second-order feature fusion to obtain enhanced word representation, and constructing a user sentiment feedback information extraction model; Processing the second user evaluation text based on the user sentiment feedback information extraction model to obtain feature importance weights and evaluation vocabulary weight values, and calculating the design parameters to be optimized; Based on the design parameters to be optimized, grid division is performed in a multidimensional design space to obtain a set of design points, and a diversified set of optimized design solutions is generated according to the set of design points; specifically, the method comprises: constructing a multidimensional design space according to the design parameters to be optimized, dividing the value range of each design parameter into equal parts, and obtaining an initial grid structure; calculating the Euclidean distance between adjacent grid points in the initial grid structure to obtain a grid spacing vector, and adaptively adjusting the initial grid structure based on the grid spacing vector, increasing the grid density in a high-sensitivity area, and obtaining an optimized grid structure; encoding each grid point in the optimized grid structure to obtain a set of design points; randomly selecting an initial design solution based on the set of design points to obtain an initial design solution set, and calculating the weighted Euclidean distance between each solution in the initial design solution set and the user preference interval based on the feature attractiveness vector to obtain a solution distance matrix; sorting the initial design solution set based on the solution distance matrix, and applying an adaptive Gaussian mutation strategy to generate a diversified set of optimized design solutions, wherein the mutation intensity of the adaptive Gaussian mutation strategy is dynamically adjusted with the number of iterations; Inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions; Iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and when the termination condition is met, output the target design solution evaluation model and the target cultural and creative product design solution.

2. The iterative design method for cultural and creative products based on user emotional feedback according to claim 1, characterized in that: The extracting of the word context representation of the first user evaluation text performs first-order and second-order feature fusion to obtain enhanced word representation, and constructs a user sentiment feedback information extraction model, including: Encoding the first user evaluation text based on the pre-trained language model to obtain a context representation vector for each word; Performing a linear transformation on the context representation vector to obtain a first-order feature vector, and performing an outer product operation on the context representation vector to obtain a second-order feature matrix; Concatenating the first-order feature vector with the second-order feature matrix to obtain an enhanced word representation; Constructing a multi-task learning framework, the multi-task learning framework includes a sentiment analysis task and a keyword extraction task, wherein the sentiment analysis task and the keyword extraction task share the underlying features of the enhanced word representation; the sentiment analysis task adopts a classifier based on an attention mechanism, and the keyword extraction task adopts a task target layer of sequence annotation; The parameters of the multi-task learning framework are optimized through end-to-end training to obtain a user emotion feedback information extraction model, wherein the end-to-end training includes combining the loss functions of the emotion analysis task and the keyword extraction task in a weighted manner to form an overall objective function.

3. The iterative design method for cultural and creative products based on user emotional feedback according to claim 1, characterized in that: The method of processing the second user evaluation text based on the user emotion feedback information extraction model to obtain feature importance weights and evaluation vocabulary weight values, and calculating the design parameters to be optimized includes: Performing word segmentation, stop word removal and part-of-speech tagging on the second user evaluation text to obtain a processed evaluation text; Processing the processed evaluation text based on the user sentiment feedback information extraction model to obtain sentiment analysis results and keyword extraction results; Constructing a two-dimensional evaluation matrix of an evaluation grid based on the sentiment analysis result and the keyword extraction result, wherein the horizontal axis of the two-dimensional evaluation matrix represents the product feature dimension, and the vertical axis represents the user demand dimension; Assigning sentiment intensity to each cell in the two-dimensional evaluation matrix to obtain a sentiment intensity matrix, and calculating the average sentiment intensity of each feature dimension in the sentiment intensity matrix to obtain an initial importance weight; Constructing a fuzzy judgment matrix based on the fuzzy analytic hierarchy process, performing fuzzy operation on the fuzzy judgment matrix to obtain evaluation vocabulary weight values, and calculating feature importance weights based on the initial importance weights and the evaluation vocabulary weight values; Performing feature correlation analysis on the sentiment intensity matrix to obtain a feature combination effect matrix, and calculating a comprehensive attractiveness score of each product feature by combining the feature importance weights, the evaluation vocabulary weight values ​​and the feature combination effect matrix; A characteristic attraction vector is constructed based on the comprehensive attraction score, and a threshold screening method is used to determine the design parameters to be optimized according to the characteristic attraction vector.

4. The iterative design method for cultural and creative products based on user emotional feedback according to claim 1, characterized in that: The diversified optimization design solution set is input into the initial design solution evaluation model for scoring prediction to obtain a set of candidate design solutions, including: The diversified optimization design solution set is input into the initial design solution evaluation model, and the initial design solution evaluation model adopts a multi-layer perceptron structure, including an input layer, three hidden layers and an output layer; the input layer receives the feature vector of the design solution, with a dimension of N, where N is the number of design parameters; the number of neurons in the three hidden layers is 128, 64 and 32 respectively; the output layer is a single neuron, and outputs a prediction score; the hidden layer uses a ReLU activation function, and the output layer uses a Sigmoid activation function to map the output to between 0 and 1; a Dropout layer is added after each hidden layer, and the dropout rate is 0.3; Extracting features from each design solution in the diversified optimization design solution set to obtain an N-dimensional design solution feature vector; combining all feature vectors into a design solution feature matrix with a dimension of M×N, where M is the number of design solutions; Input the design scheme feature matrix into the input layer of the initial design scheme evaluation model, and obtain a 128-dimensional intermediate feature representation through calculation by the first hidden layer; input the 128-dimensional intermediate feature representation into the second hidden layer, and obtain a 64-dimensional high-level feature representation through calculation; input the 64-dimensional high-level feature representation into the third hidden layer, and obtain a 32-dimensional final feature representation through calculation; The 32-dimensional final feature representation is input to the output layer, and a prediction score is obtained through a Sigmoid activation function to form an M-dimensional prediction score vector; based on the prediction score vector, the diversified optimization design solution set is sorted in descending order, and the top K design solutions are selected to obtain a high-scoring solution set; A K-means clustering analysis is performed on the high-scoring solution set, and the solution with the nearest and highest score to each cluster center is selected to form a set of candidate design solutions.

5. The iterative design method for cultural and creative products based on user emotional feedback according to claim 1, characterized in that: The iterative optimization of the candidate design solution set, updating the initial design solution evaluation model, and outputting the target design solution evaluation model and the target cultural and creative product design solution when the termination condition is met, comprises: Performing parameter encoding on each solution in the candidate design solution set to obtain a solution parameter matrix, and calculating the Euclidean distance between solutions based on the solution parameter matrix to construct a solution distance matrix; Performing cluster analysis on the candidate design solution set based on the solution distance matrix to obtain a solution clustering result, selecting representative solutions from the solution clustering result, performing crossover and mutation operations on the representative solutions to obtain a new generation of design solution set; Inputting the new generation design solution set into the initial design solution evaluation model to obtain a solution prediction score, and sorting the new generation design solution set based on the solution prediction score, selecting several solutions with the highest scores as elite individuals, and obtaining an elite solution set; Collecting user feedback on the elite solution set to obtain user rating data, and updating parameters of the initial design solution evaluation model based on the user rating data to obtain an updated design solution evaluation model; Determine whether the termination condition is met. If not, continue iterating. If met, select the solution with the highest score and satisfying all constraints from the elite solution set as the target cultural and creative product design solution, and use the updated design solution evaluation model as the target design solution evaluation model.

6. A cultural and creative product iterative design system based on user emotional feedback, characterized in that: Used to execute the iterative design method of cultural and creative products based on user emotional feedback as described in any one of claims 1 to 5, the iterative design system of cultural and creative products based on user emotional feedback comprises: A fusion module is used to fuse multiple features of the first user's evaluation text to obtain a sentiment seed word set and perform cluster analysis to obtain a fine-grained sentiment dictionary in the field of cultural and creative products; A construction module is used to extract the word context representation of the first user evaluation text, perform first-order and second-order feature fusion, obtain enhanced word representation, and construct a user sentiment feedback information extraction model; A calculation module, used to process the second user evaluation text based on the user emotion feedback information extraction model, obtain feature importance weights and evaluation vocabulary weight values, and calculate the design parameters to be optimized; A partitioning module, used to perform grid partitioning in a multidimensional design space based on the design parameters to be optimized, obtain a set of design points, and generate a set of diversified optimization design solutions based on the set of design points; A prediction module, used for inputting the diversified optimization design solution set into the initial design solution evaluation model for score prediction to obtain a set of candidate design solutions; The output module is used to iteratively optimize the candidate design solution set, update the initial design solution evaluation model, and output the target design solution evaluation model and the target cultural and creative product design solution when the termination condition is met.

7. A device for iterative design of cultural and creative products based on user emotional feedback, characterized in that: The device for iterative design of cultural and creative products based on user emotional feedback comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the cultural and creative product iterative design device based on user emotional feedback executes the cultural and creative product iterative design method based on user emotional feedback as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the iterative design method for cultural and creative products based on user emotional feedback as described in any one of claims 1-5 is implemented.

Images