Commodity recommendation copywriting generation method based on multi-modal artificial intelligence large model

Abstract

The invention provides a document generation method, belongs to the technical field of multi-modal artificial intelligence large models, and particularly relates to a commodity recommendation document generation method based on a multi-modal artificial intelligence large model. The problems that in the prior art, limitations exist in the aspects of data expansion, feature extraction, dimensionality reduction and multi-modal text generation, for example, the diversity of generative adversarial networks is insufficient, the training stability of neural networks is poor, the dimensionality reduction effect of an auto-encoder is poor, the modal feature fusion effect of a Transform model is poor, and model generalization and performance are affected are solved.

Description

Technical Field

[0001] The present invention provides a copywriting generation method, which belongs to the technical field of multimodal artificial intelligence big models, and in particular relates to a commodity recommendation copywriting generation method based on a multimodal artificial intelligence big model. Background Art

[0002] With the rapid development of e-commerce platforms, the types and quantities of goods have increased rapidly. Merchants need to generate a large number of high-quality recommendation copywriting to attract customers and increase sales. Traditional copywriting generation methods mainly rely on manual writing, which is not only time-consuming and labor-intensive, but also difficult to ensure the consistency and quality of the copywriting. With the development of artificial intelligence and machine learning technology, data-driven automated copywriting generation methods have gradually become a research hotspot.

[0003] The Chinese invention patent with application number CN202410293678,7 proposes a global visual guided image description generation method based on a large cross-modal model, which belongs to the field of image processing. It includes: inputting the classification label and the image to be tested into the cross-modal model to obtain the predicted label; combining the predicted label with the grammatical prompt sentence to construct the language model prompt sentence; inputting the prompt sentence into the large-scale pre-trained language model to obtain the pre-selected vocabulary, and obtaining the vocabulary candidate list based on the pre-selected vocabulary and the predicted label; obtaining the generated vocabulary based on the scores corresponding to the vocabulary in the candidate list; merging the generated vocabulary with the prompt sentence and inputting it into the language model for iteration until the number of iterations reaches the threshold or the end symbol is generated, and obtaining the image description of the image to be tested according to the scores of all generated vocabulary. This method improves the accuracy and naturalness of the image description by inputting the labels that are strongly related to the image features into the language model and incorporating them into the vocabulary candidate list, and obtaining the generated vocabulary through the weighted image-text consistency score and semantic fluency score.

[0004] The Chinese invention patent with application number CN202410822792, 4 proposes an event extraction method, device, equipment and medium based on a large model, which relates to the field of natural language processing. The method includes: performing dependency syntax analysis on the event instance to obtain the triple of the event instance; according to the event instance and the triple, obtaining the feature vector corresponding to the event instance through the extraction model, and placing the feature vector in the vector database; confirming the input of the first large model according to the event definition, generation rules, the vector database and the event to be extracted of the event instance, and obtaining the extraction result of the event to be extracted through the first large model, wherein the event definition includes the event type, and the generation rules are used to provide rules for the first large model to extract the event to be extracted; updating the generation rules according to the extraction results through the second large model. The method of the present application can achieve high-quality event capture and improve the model learning ability.

[0005] The Chinese invention patent with application number CN202311295531, 3 proposes a method and system for automatic labeling of visual sample data based on a large model, including: generating a visual word histogram according to visual features, calculating the potential topic distribution of each image in the visual sample data according to the visual word histogram; generating visual sample training features according to the potential topic distribution and the original label, and training the convolutional neural network using the visual sample training features; labeling the visual sample data to be labeled using the visual sample labeling model, and calculating the first labeling word confidence of the initial labeling word; calculating the visual word relevance between the initial labeling word and the visual sample data, and calculating the second labeling word confidence of the visual sample data to be labeled according to the visual word relevance; generating labeling data of the visual sample data according to the first labeling word confidence and the second labeling word confidence. The present invention can improve the accuracy of automatic labeling of visual sample data.

[0006] Based on the above-mentioned various existing solutions, the existing technology has the following deficiencies:

[0007] 1. In terms of data expansion, the traditional generative adversarial network is insufficient in the diversity and richness of sample generation, resulting in insufficient training data and poor generalization ability of the model.

[0008] 2. In terms of feature extraction, traditional neural networks are prone to gradient vanishing, gradient exploding, or falling into local optimal solutions during training, which affects the stability of feature extraction and model performance.

[0009] 3. In terms of feature dimensionality reduction, conventional autoencoders cannot fully capture and retain the intrinsic correlation and structural characteristics of the data during the dimensionality reduction process, resulting in poor dimensionality reduction effects.

[0010] 4. In terms of existing technologies for copywriting generation, the traditional Transformer model has insufficient optimization of the attention mechanism when processing multimodal data, resulting in poor inter-modal feature fusion effect and low text generation quality. Summary of the invention

[0011] In order to make up for the shortcomings of the prior art, the embodiments of the present application provide a method for generating product recommendation texts based on a multimodal artificial intelligence large model, thereby solving the limitations of the prior art in data expansion, feature extraction, dimensionality reduction and multimodal text generation, such as insufficient diversity of generative adversarial networks, poor stability of neural network training, poor dimensionality reduction effect of autoencoders, and poor modal feature fusion effect of Transformer models, which affect model generalization and performance.

[0012] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for generating product recommendation copy based on a multimodal artificial intelligence large model, comprising the following steps:

[0013] S1. Data collection and labeling: Collect multimodal information of products from multiple e-commerce platforms, including product images, parameter descriptions, etc., and manually label them;

[0014] S2. Data expansion: Generate samples using a generative adversarial network based on the quantum tunneling effect to achieve data expansion;

[0015] S3, feature extraction: a 3-layer fully connected neural network is used to extract features based on the adaptive biological impact optimization algorithm;

[0016] S4, feature dimensionality reduction: feature dimensionality reduction is performed using an autoencoder neural network based on feature adaptive refinement;

[0017] S5. Generative AI model training: training is performed using the Transformer model based on high-order tensor accumulation;

[0018] S6. Product recommendation copy generation: Use the trained model to generate recommendation copy for new products.

[0019] Preferably, the generation of an adversarial network based on the quantum tunneling effect in step S2 includes:

[0020] S201, initializing the weight parameters of the generator and the discriminator;

[0021] S202, converting input data into a high-dimensional quantum state through multi-state encoding of quantum bit superposition;

[0022] S203, the generator simulates the data generation process using a quantum tunneling algorithm to generate new data;

[0023] S204, the discriminator adopts a quantum state estimation strategy to evaluate the statistical fit between the generated samples and the real samples;

[0024] S205, iterative training until the preset conditions are met.

[0025] Preferably: the generator and the weight of the generator in step S201 are subject to a specific distribution, and the specific distribution has 0 as the mean, is a normal distribution with a variance; in the step S202, the input data is converted into a high-dimensional quantum state through multi-state encoding of quantum bit superposition, so that each data point is represented as a quantum state probability distribution; in the step S203, the generator simulates the data generation process through a quantum tunneling algorithm, combines quantum error correction coding to improve robustness, and uses quantum logic gate serialization to represent; in the steps S204 and 205, during the adversarial training stage, the generator deceives the discriminator by minimizing the loss function, and the discriminator evaluates the similarity between the generated samples and the real samples; the loss function includes the logarithmic loss of the generated samples and the real samples, and a regularization term; the output of the discriminator is based on the measurement probability of the quantum state, and the weights of the generator and the discriminator are updated by a quantum derivative optimization method; the step S2 iterates the training until the maximum number of iterations is met.

[0026] Preferably: the parameters of the neural network in step S3 are optimized by simulating the survival competition and resource adaptation process of organisms;

[0027] Preferably: during the training process of the neural network, normal distribution is used to initialize weights and biases; after each iteration, the network parameter performance is evaluated and the loss and environmental fitness are calculated; the weights and biases are updated according to resource competition, and the momentum term is used to smooth the learning process; the learning rate and sensitivity parameters are adjusted to adapt to environmental changes; the resource allocation sensitivity hyperparameters are dynamically adjusted to avoid local optimality; the step S3 is iterated until a preset maximum number of times or a stop condition is reached.

[0028] Preferably, in step S4, the autoencoder neural network realizes low-dimensional conversion and high-dimensional reconstruction of features through an encoder and a decoder, and simultaneously uses a feature coupling mechanism to capture the intrinsic correlation and structural features of the data.

[0029] Preferably: in the training process of the autoencoder, the network parameters are initialized using a normal distribution; the input data is encoded into a low-dimensional representation and reconstructed by forward propagation; dynamic weight adjustment is applied in the hidden layer to achieve adaptive feature refinement; the loss function (including mean square error and L1 regularization) is calculated; the parameters are updated using the gradient descent method; the learning rate is dynamically adjusted to adapt to different training stages; the step S4 is iterated until a preset maximum number of times or a stop condition is reached.

[0030] Preferably, the training process of the Transformer model based on high-order tensor accumulation in step S5 includes:

[0031] S501, initializing the parameters of the Transformer model;

[0032] S502, performing a self-attention operation on the input features;

[0033] S503, using high-order tensor operations to combine the features of each mode;

[0034] S504, calculating the error based on the loss function and updating the model parameters through the back propagation algorithm;

[0035] S505: Iterate training until preset conditions are met.

[0036] Preferably: in the steps S502-S505, the attention score matrix is ​​decomposed into multiple smaller matrices, each matrix focuses on the relationship between different modalities, thereby improving the ability of text generation; in the step S501, the parameters of the Transformer model are randomly initialized using an approximate normal distribution; in the step S502, feature vectors are defined and processed, including feature extraction and dimensionality reduction, and the weights between features are calculated through a self-attention mechanism; in the step S503, the interaction between features is optimized through a dual decomposition technique, and a tensor product operation is used to achieve feature fusion; in the step S504, the cross entropy loss and the regularization term constitute a loss function, the cross entropy loss calculates the error between the model output and the true label, the regularization term is used to control the model complexity and prevent overfitting, and the back propagation algorithm is used to calculate the gradient of the loss function to the model parameters, and the parameters are updated through a gradient descent method.

[0037] Preferably, the Transformer model generates a recommendation text based on the processed feature values ​​of the input.

[0038] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0039] 1. A generative adversarial network based on the quantum tunneling effect is used to generate samples with higher diversity and richness through multi-state coding of quantum bit superposition and quantum error correction coding function, thereby effectively expanding the training data.

[0040] 2. A three-layer fully connected neural network is used, based on an adaptive biological impact optimization algorithm, to simulate the survival competition and resource adaptation process of organisms, optimize the parameters of the neural network, avoid gradient vanishing, gradient explosion and local optimal solution problems, and improve feature extraction capabilities.

[0041] 3. Adopt an autoencoding neural network based on feature adaptive refinement, convert high-dimensional input features into low-dimensional hidden representations through the encoder, and then reconstruct them back to high-dimensional space through the decoder, so as to capture the intrinsic correlation and structural characteristics of the input data and optimize the information flow of the network.

[0042] 4. Adopting the Transformer model based on high-order tensor accumulation, by optimizing the self-attention mechanism of Transformer, the attention score matrix is ​​decomposed into multiple smaller matrices, focusing on the specific relationship between different modalities and improving the text generation ability.

[0043] Based on the above innovations, the following technical effects are achieved:

[0044] 1. The diversity and richness of samples generated by data augmentation technology are significantly improved, which enhances the generalization ability of the model.

[0045] 2. Feature extraction technology enhances the stability of the neural network through an adaptive biological influence optimization algorithm, avoids gradient-related problems, and improves the accuracy and efficiency of feature extraction.

[0046] 3. Feature dimensionality reduction technology uses an autoencoding neural network with adaptive feature refinement to effectively reduce the data dimension while retaining the inherent correlation of the data and improving the effect of feature dimensionality reduction.

[0047] 4. The copywriting generation technology is based on the Transformer model with high-order tensor accumulation, which significantly improves the fusion capability of different modal features and the quality of text generation.

[0048] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The present invention provides a method for generating product recommendation texts based on a multimodal artificial intelligence large model, and a training process of a neural network algorithm based on adaptive biological influence optimization;

[0050] Figure 2 The present invention discloses a training process of an autoencoding neural network algorithm based on feature adaptive refinement for a method for generating product recommendation text based on a multimodal artificial intelligence large model. DETAILED DESCRIPTION

[0051] 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 only 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.

[0052] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used in this document are only for illustrative purposes and do not represent the only implementation method.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more related listed items.

[0054] like Figure 1 and Figure 2 As shown, the present invention proposes a method for generating product recommendation copy based on a multimodal artificial intelligence large model, and the main steps are as follows:

[0055] S1. Data collection and annotation

[0056] The data of the present invention is collected from multiple e-commerce platforms. The data includes multimodal information such as product patterns and parameter introductions. The collected data is stored in JSON format and the data is annotated manually, that is, a professional annotation team generates corresponding recommendation texts for each product data.

[0057] Furthermore, the collected data is converted into a vector form to adapt the machine learning model. The present invention uses advanced text embedding technology and image feature extraction technology to perform vectorization conversion of multimodal data. Specifically, text information is converted into text vectors through a pre-trained natural language processing model, and images are extracted through convolutional neural networks and converted into image vectors. Furthermore, the text vector and the image vector are cascaded to obtain the original data vector feature.

[0058] S2 Data Augmentation

[0059] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a generative adversarial network based on the quantum tunneling effect to generate samples, thereby achieving data expansion. The generative adversarial network based on the quantum tunneling effect includes a generator and a discriminator. The generator uses a multi-state encoding method of quantum bit superposition to simulate the super-position state in quantum computing to increase the diversity and richness of the generated samples; the discriminator uses a quantum state estimation strategy to evaluate the statistical fit between the generated samples and the real samples to improve the accuracy of the discrimination.

[0060] Specifically, the training process of the generative adversarial network algorithm based on the quantum tunneling effect is as follows:

[0061] S201, initializing and setting weight parameters of the generator and the discriminator. In one embodiment, the initialization method is expressed as:

[0062]

[0063] Where W cG and W cD are the weights of the generator and the discriminator respectively; ~ means obeying a specific distribution; It means that 0 is the mean value. is a normal distribution with a variance. Preferably, Set to 0, 01.

[0064] S202, converting the input data into a high-dimensional quantum state through multi-state encoding of quantum bit superposition, each data point is represented as a quantum state probability distribution, and the calculation method is expressed as:

[0065]

[0066] Where, X c To generate the input data of the adversarial network, Ψ c For data X c The corresponding quantum state vector; α ci represents the complex amplitude of the superposition of the i-th quantum state; |φ ci > represents the ground state of the i-th qubit; NL is the total number of qubits.

[0067] In one embodiment, the ground state of each qubit is defined by a quantum Fourier transform, which is calculated as:

[0068]

[0069] Where |k> represents the k-state of the quantum bit; i * Is an imaginary unit.

[0070] S203, when the generator receives the encoded quantum state input, it uses the quantum tunneling algorithm to simulate the data generation process, and the way to generate new data is expressed as:

[0071]

[0072] Where ×° is the element-by-element point multiplication operation; E() is the quantum error correction coding function; is the data generated by the generator; G() is the generator function; Θ c are the parameters of the generator.

[0073] In one embodiment, the quantum error correction coding function encodes the original quantum state Ψ c To a higher-dimensional error correction space to achieve, so as to improve the robustness of data in the quantum tunneling process, the calculation method is expressed as:

[0074]

[0075] In the formula, β cij represents the coding coefficient of the i-th dimension; Jr is the dimension of the coding space; φ cj is the quantum state of the jth dimension; |> represents the right vector of the quantum state. Preferably, the dimension Jr of the encoding space is twice the dimension NL of the original quantum bit.

[0076] Furthermore, in this embodiment, the generator is enhanced by a serialized representation of quantum logic gate operations to simulate the physical process of quantum tunneling, which is represented as:

[0077]

[0078] In the formula, U(Θ c ) depends on the parameter Θ c The quantum logic gates is the parameter Θ c Preferably, the quantum logic gate adopts a Hadamard gate.

[0079] S204. In the adversarial training stage, the discriminator evaluates the similarity between the generated samples and the real samples, and the generator tries to deceive the discriminator so that it cannot distinguish between real and fake samples. The training process is constrained by minimizing the loss function, which is expressed as:

[0080]

[0081] In the formula, λ c and γ c is the regularization coefficient; || || 2 represents the L2 norm; is the loss function of the generative adversarial network; D() is the discriminator function; X ci is the i-th real data sample; is the i-th generated data sample; mLa is the total number of generated samples. Preferably, λ c and γ c Set to 0,3 and 0,2 respectively.

[0082] In one embodiment, the output of the discriminator is the measurement probability based on the quantum state, expressed as:

[0083] D(X ci )=Tr(ρ ci M ci )

[0084] In the formula, ρ ci is the input data X ci The corresponding density matrix, M ci is with X ci The associated quantum measurement operator, Tr() is the trace operation.

[0085] S205. The weights of the generator and the discriminator are updated using an optimization method based on quantum derivatives. The parameter update method is expressed as:

[0086]

[0087] In the formula, ← is the parameter update operation, Θ c and W cD are the parameters of the generator and the discriminator respectively; η c is the learning rate; and They are the loss functions About Θ c and W cD Preferably, η c Set to 0, 01.

[0088] S206, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0089] S3. Feature extraction model training

[0090] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention uses a 3-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion or falling into the local optimal solution may be encountered, affecting the stability of training and the performance of the model. The present invention uses a neural network algorithm based on adaptive biological impact optimization as a feature extraction model, inspired by the biological adaptive behavior in nature, especially the ability of organisms to adjust strategies according to resource changes in the environment, and optimizes the parameters of the neural network by simulating the survival competition and resource adaptation process of organisms, thereby avoiding the problems of gradient vanishing, gradient explosion and falling into the local optimal solution, and improving the feature extraction capability.

[0091] S301, initialize the parameters of the neural network, define the weights and biases of the neural network, in one embodiment, the initial weights and initial bias The initialization method is expressed as:

[0092]

[0093] In the formula, ~ means obeying a specific distribution, W NN and b NN are the weights and biases of the neural network, σ sn is the initialization standard deviation, and are the initial weights and initial biases of the neural network, respectively. Preferably, σ sn Set to 0, 01, is a normal distribution.

[0094] S302: At the beginning of each iteration, evaluate the performance of the current network parameters under a specific training environment. In one embodiment, calculate the loss L under the current parameter configuration. NN (W NN , b NN ), and estimate the environmental fitness as the basis for subsequent adjustments. The calculation method is expressed as:

[0095] E NN =exp(-β NN ·(W NN , b NN ))

[0096] In the formula, E NN is the environmental fitness, β NN is the sensitivity parameter, L NN () is the loss function of the neural network.

[0097] And, in this embodiment, the loss L NN (W NN , b NN ) is calculated using the cross entropy loss method and is expressed as:

[0098]

[0099] In the formula, C represents the number of categories, y c is the one-hot encoding of the true label, is the probability predicted by the model.

[0100] Furthermore, the probability predicted by the model is calculated by the softmax function and is expressed as:

[0101]

[0102] In the formula, z c is the logistic regression probability output by the network.

[0103] In the resource competition behavior in nature, each parameter is regarded as an individual, competing for resources in the data set. The resource allocation is determined by the contribution of the parameter. The resource allocation of each parameter is calculated as follows:

[0104]

[0105] In the formula, R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indices, γ NN is a hyperparameter that controls the sensitivity of resource allocation, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of .

[0106] S303: Update weight W according to the result of resource competition NN and bias b NN , the update method is expressed as:

[0107]

[0108]

[0109] In the formula, R i The resources allocated for the parameters of the i-th neural network layer, is the bias of the neurons in the i-th layer, is the weight parameter update of the lth neuron in the kth neural network layer, α NN is the learning rate, is the update amount of the bias parameter of the i-th layer of neurons. The present invention adopts the momentum term m ij To smooth the learning process of weight parameters, it can be expressed as:

[0110]

[0111] In the formula, is the momentum term of the jth neuron in the ith neural network layer at the t-1th iteration, is the momentum term of the jth neuron in the jth neural network layer at the tth iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-1th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-th iteration, β xm is the momentum decay parameter. Preferably, β xm Set to 0,95.

[0112] S304: Adjust the sensitivity parameter and the learning rate parameter according to the result of the environmental perception evaluation. The adjustment method is expressed as follows:

[0113]

[0114] In the formula, ΔENN is the change in environmental fitness, that is, the difference between the environmental fitness of this iteration and the previous iteration; η NN and κ NN is the adjustment factor, is the updated learning rate parameter, is the updated sensitivity parameter. Preferably, η NN and κ NN Both are set to 0 and 1.

[0115] S305, dynamically adjusting the hyperparameters controlling the sensitivity of resource allocation between the global search and the local search to avoid falling into the local optimal solution, the adjustment method is expressed as:

[0116]

[0117] In the formula, || is the absolute value symbol, is the hyperparameter of the adjusted control resource allocation sensitivity, and ρ is a balanced hyperparameter. Preferably, ρ NN Set to 0, 1.

[0118] S306, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0119] S4. Feature dimensionality reduction model training

[0120] The features after feature extraction are input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention adopts an autoencoder neural network algorithm based on feature adaptive refinement to perform feature dimensionality reduction. The autoencoder neural network algorithm includes two parts: an encoder and a decoder. The encoder converts the features extracted from the high-dimensional input into a low-dimensional hidden representation, and the decoder reconstructs this hidden representation back to the original high-dimensional space to form a reconstructed output. The present invention adopts a feature coupling mechanism so that the autoencoder can not only reduce the dimension of the data, but also more finely capture and reconstruct the intrinsic correlation and structural characteristics in the input data, thereby optimizing the information flow of the network.

[0121] S401, initializing the network parameters of the autoencoder, including the weight and bias of the autoencoder. In one embodiment, the initialization method is expressed as:

[0122]

[0123]

[0124] Where W r is the weight parameter of the autoencoder, br is the bias parameter of the autoencoder, is the initial weight of the autoencoder; is the initial bias of the autoencoder; It means the mean is 0 and the variance is Normal distribution of is the number of neurons in the input layer of the autoencoder.

[0125] S402, input feature extracted data X r Propagate through the sequence of encoder layers and transform into a low-dimensional representation H r , expressed as:

[0126]

[0127] Where Re() is the ReLU activation function; H r is the low-dimensional representation of the encoder output; and are the weight and bias of the encoder respectively; and are the weights and biases of the decoder; Reconstructed output.

[0128] S403. In the hidden layer of the autoencoder, the weight is adjusted by the dynamic weight adjustment layer to achieve the purpose of adaptive feature refinement. The calculation method of the weight update amount of the autoencoder is expressed as:

[0129]

[0130] In the formula, ΔW r is the weight update amount; tanh( ) is the hyperbolic tangent function; λ r is the learning rate; r is the adjustment factor; is the loss function L r About weight W r The gradient of L r is the reconstruction error. Preferably, λ r Set to 0, 01, γ r Set to 3, 14.

[0131] Furthermore, the gradient Solve the loss function L by chain rule r The gradient with respect to the weight is calculated and expressed as:

[0132]

[0133] In the formula, is the partial derivative of the loss function with respect to the reconstructed output.

[0134] In one embodiment, the calculation method of the mean square error of the loss function is expressed as:

[0135]

[0136] In the formula, || || is the L2 norm, α r is the regularization coefficient, is the L1 norm of the encoder weights. Preferably, α r Set to 0, 3.

[0137] S404: Update the parameters of the encoder and decoder using the gradient descent method. The updating method is expressed as:

[0138]

[0139] Where η r is the learning rate; is the weight parameter of the autoencoder at the t+1th iteration; is the weight parameter of the autoencoder at the tth iteration; is the bias parameter of the autoencoder at the t+1th iteration; is the bias parameter of the autoencoder at the tth iteration; t represents the number of iterations; and are the gradients of the loss function with respect to weights and biases, respectively.

[0140] In one embodiment, the learning rate η r It is set by adaptive adjustment to dynamically adjust according to different stages of network training, expressed as:

[0141]

[0142] In the formula, is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, δ r is the sensitivity parameter of the learning rate adjustment. Preferably, δ r Set to 0, 5.

[0143] S405, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0144] S5. Generative AI model training

[0145] The generative AI model of the present invention adopts a Transformer model based on high-order tensor accumulation. The Transformer model based on high-order tensor accumulation optimizes the self-attention mechanism of the Transformer on the basis of the basic Transformer model. This is achieved by decomposing the traditional attention score matrix into multiple smaller matrices, each of which focuses on the specific relationship between different modalities, thereby improving the ability of text generation.

[0146] Specifically, the training process of the Transformer model based on high-order tensor accumulation is as follows:

[0147] S501. Initialize the parameters of the Transformer model, including the weights and biases of the attention mechanism. In one embodiment, the initialization method uses an approximate normal distribution for random initialization.

[0148] S502, define the original feature vector as W uI , the feature vector after feature extraction is W uT , the feature vector after feature dimension reduction is W uP Furthermore, a self-attention operation is performed on each input feature to calculate the weights between the features. The self-attention calculation method of the present invention is expressed as:

[0149]

[0150] In the formula, Q u , K u and V u are query, key, and value matrices respectively; K u The transpose of d k is the dimension of the key vector; A u is the output of the attention weight; softmax() is the Softmax function.

[0151] Furthermore, the query, key, and value matrices are composed of the input data X u =[V u ; T u ;P u ] is obtained through weight conversion, and the calculation method is expressed as:

[0152] Q u =W uQ X u

[0153] K u =W uK X u

[0154] V u =WuV X u

[0155] Where, X u =[V u ; T u ;P u ] is the merge V u 、T u and P u The feature vectors of W are stacked; uQ , W uK and W uV is the transformation matrix corresponding to query, key, and value.

[0156] S503, using high-order tensor operations to merge the features of each mode, and enhancing it through dual decomposition technology to optimize the interaction and dependence between different features, expressed as:

[0157]

[0158] In the formula, Represents the tensor product operation, which is used to fuse the features of different modalities; A ui is the attention weight of the i-th feature; Ng is the number of features.

[0159] S504, calculating the error based on the loss function and updating the parameters of the Transformer model through the back propagation algorithm. The calculation method of the loss function is expressed as:

[0160]

[0161] Where, L u is the loss function of the Transformer model; N is the number of samples; y n is the onehot encoding of the label; is the output of the model, calculated by the preset Softmax function; u is the regularization coefficient. Preferably, λ u Set to 0, 3.

[0162] Furthermore, the parameters of the Transformer model are expressed as:

[0163]

[0164] Where η u is the learning rate; W ui is the weight of the i-th neuron of the Transformer model; is the weight of the i-th neuron of the Transformer model at the t-th iteration; is the weight of the i-th neuron of the Transformer model at the t+1th iteration; is the loss function L u About weight W ui gradient.

[0165] S505, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0166] S6. Product recommendation copy generation

[0167] The trained model is used to generate recommendation texts for new products. In one embodiment, the collected raw data is vectorized and then input into the trained feature extraction and feature dimensionality reduction model for feature processing. Furthermore, the processed features are input into the Transformer model for training the Transformer model, thereby generating product recommendation texts. Based on the technology of the present invention, merchants can dynamically generate targeted recommendation texts based on the real-time data of products, and more effectively achieve the purpose of marketing.

[0168] When using:

[0169] 1. Data collection and annotation (S1):

[0170] Collect multimodal information such as product patterns, parameter descriptions, etc. from multiple e-commerce platforms.

[0171] The data is stored in JSON format and manually annotated to generate recommendation text.

[0172] The text information is converted into text vectors through a pre-trained natural language processing model, and the image is converted into image vectors through a convolutional neural network.

[0173] The text vector is concatenated with the image vector to form the original data vector feature.

[0174] 2. Data expansion (S2):

[0175] Use a generative adversarial network (GAN) based on the quantum tunneling effect to generate samples and increase data diversity.

[0176] It includes a generator and a discriminator. The generator simulates the super-position state in quantum computing, and the discriminator evaluates the statistical fit of the sample.

[0177] 3. Feature extraction model training (S3):

[0178] Adopting the adaptive biological influence optimization neural network algorithm, we simulate the biological adaptive behavior, optimize the network parameters and avoid the gradient problem.

[0179] 4. Feature dimensionality reduction model training (S4):

[0180] Using an autoencoding neural network algorithm based on feature adaptive refinement, the data dimension is reduced through the encoder and decoder while capturing the intrinsic correlation of the data.

[0181] 5. Generative AI model training (S5):

[0182] A Transformer model based on high-order tensor accumulation is used to optimize the self-attention mechanism and improve text generation capabilities.

[0183] 6. Product recommendation copy generation (S6):

[0184] The processed features are input into the trained Transformer model to generate recommendation copy for new products.

[0185] The entire process combines multimodal data processing, advanced data augmentation technology, adaptive optimization algorithms and high-order tensor Transformer models to dynamically generate targeted recommendation copy and improve marketing effectiveness.

[0186] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the definition of the claims.

Claims

1. A method for generating product recommendation copy based on a multimodal artificial intelligence large model, characterized in that: The following steps are involved: S1. Data collection and labeling: Collect multimodal information of products from multiple e-commerce platforms, including product images, parameter descriptions, etc., and manually label them; S2. Data expansion: Generate samples using a generative adversarial network based on the quantum tunneling effect to achieve data expansion; S3, feature extraction: a 3-layer fully connected neural network is used to extract features based on the adaptive biological impact optimization algorithm; S4, feature dimensionality reduction: feature dimensionality reduction is performed using an autoencoder neural network based on feature adaptive refinement; S5. Generative AI model training: training is performed using the Transformer model based on high-order tensor accumulation; S6. Product recommendation copy generation: Use the trained model to generate recommendation copy for new products.

2. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 1, characterized in that: The generation of an adversarial network based on the quantum tunneling effect in step S2 includes: S201, initializing the weight parameters of the generator and the discriminator; S202, converting input data into a high-dimensional quantum state through multi-state encoding of quantum bit superposition; S203, the generator simulates the data generation process using a quantum tunneling algorithm to generate new data; S204, the discriminator adopts a quantum state estimation strategy to evaluate the statistical fit between the generated samples and the real samples; S205, iterative training until the preset conditions are met.

3. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 2, characterized in that: The generator and the weight of the generator in step S201 are subject to a specific distribution, which has a mean of 0. is a normal distribution with a variance; in the step S202, the input data is converted into a high-dimensional quantum state through multi-state encoding of quantum bit superposition, so that each data point is represented as a quantum state probability distribution; in the step S203, the generator simulates the data generation process through a quantum tunneling algorithm, combines quantum error correction coding to improve robustness, and uses quantum logic gate serialization to represent; in the steps S204 and 205, during the adversarial training stage, the generator deceives the discriminator by minimizing the loss function, and the discriminator evaluates the similarity between the generated samples and the real samples; the loss function includes the logarithmic loss of the generated samples and the real samples, and a regularization term; the output of the discriminator is based on the measurement probability of the quantum state, and the weights of the generator and the discriminator are updated by a quantum derivative optimization method; the step S2 iterates the training until the maximum number of iterations is met.

4. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 1, characterized in that: The parameters of the neural network in step S3 are optimized by simulating the survival competition and resource adaptation process of the organism.

5. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 4, characterized in that: The weights and biases are initialized using a normal distribution during the training of the neural network; After evaluating the performance of network parameters in each iteration, the loss and environmental fitness are calculated; weights and biases are updated according to resource competition, and the momentum term is used to smooth the learning process; learning rate and sensitivity parameters are adjusted to adapt to environmental changes; resource allocation sensitivity hyperparameters are dynamically adjusted to avoid local optimality; step S3 is iterated until a preset maximum number of times or a stop condition is reached.

6. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 1, characterized in that: In step S4, the autoencoder neural network realizes low-dimensional conversion and high-dimensional reconstruction of features through an encoder and a decoder, and at the same time uses a feature coupling mechanism to capture the intrinsic correlation and structural characteristics of the data.

7. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 6, characterized in that: During the training process of the autoencoder, the network parameters are initialized using a normal distribution; the input data is encoded into a low-dimensional representation and reconstructed by forward propagation; dynamic weight adjustment is applied in the hidden layer to achieve adaptive feature refinement; the loss function (including mean square error and L1 regularization) is calculated; the parameters are updated using the gradient descent method; the learning rate is dynamically adjusted to adapt to different training stages; the step S4 is iterated until a preset maximum number of times or a stopping condition is reached.

8. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 1, characterized in that: The training process of the Transformer model based on high-order tensor accumulation in step S5 includes: S501, initializing the parameters of the Transformer model; S502, performing a self-attention operation on the input features; S503, using high-order tensor operations to combine the features of each mode; S504, calculating the error based on the loss function and updating the model parameters through the back propagation algorithm; S505: Iterate training until preset conditions are met.

9. The method for generating product recommendation copy based on a multimodal artificial intelligence large model according to claim 8, characterized in that: In the steps S502-S505, the attention score matrix is ​​decomposed into multiple smaller matrices, each matrix focuses on the relationship between different modalities, thereby improving the ability of text generation; the step S501 uses an approximate normal distribution to randomly initialize the parameters of the Transformer model; the step S502 defines and processes feature vectors, including feature extraction and dimensionality reduction, and calculates the weights between features through a self-attention mechanism; the step S503 optimizes the interaction between features through a dual decomposition technique, and uses a tensor product operation to achieve feature fusion; in the step S504, the cross entropy loss and the regularization term constitute a loss function, the cross entropy loss calculates the error between the model output and the true label, the regularization term is used to control the model complexity and prevent overfitting, and the back propagation algorithm is used to calculate the gradient of the loss function to the model parameters, and update the parameters through the gradient descent method.

10. A method for generating product recommendation copy based on a multimodal artificial intelligence large model according to any one of claims 1 to 9, characterized in that: The Transformer model generates recommendation text based on the processed feature values ​​of the input.

Images