An AI-based multi-language precise user demand prediction and marketing decision method

Abstract

The application discloses a kind of based on AI multilingual precise user demand prediction and marketing decision method, it is related to business data processing technical field, including, collection multilingual user data, and pre-processing, extract the initial text features of multilingual user data after pre-processing;Marketing element set is input into multilingual generative AI model, generates marketing content, and according to semantic alignment vector, marketing content is adaptively reconstructed and consistency check, and output personalized marketing content;Through the preference channel of user, personalized marketing content is distributed to user, and real-time tracking user feedback behavior data, user feedback behavior data is used as new multilingual user data.The application is projected to unified culture neutral space by calculating the correlation of semantic representation vector and preset culture dimension vector, effectively eliminates the interference of cultural background to semantic expression, improves the quality of demand prediction model input feature.

Description

Technical Field

[0001] This invention relates to the field of business data processing technology, and in particular to an AI-based multilingual method for accurate user demand prediction and marketing decision-making. Background Technology

[0002] With the deepening development of the globalized business environment, multilingual user data mining and cross-cultural marketing strategy optimization have become core aspects of enterprises' international operations. Existing technologies generally use machine translation combined with semantic analysis to process multilingual data and build user demand prediction based on deep learning models. In recent years, attention-based cross-cultural semantic representation learning and generative AI marketing content automatic generation technologies have gradually matured, providing multinational enterprises with data-driven marketing decision support tools.

[0003] However, existing methods have shortcomings in fine-grained alignment and dynamic adaptability of multilingual cultural semantics. Existing technologies usually rely on static cultural dimension models or simple semantic embedding spatial projections, which fail to effectively quantify the dynamic impact of different cultural backgrounds on user demand representation. This results in insufficient accuracy in predicting user demand in cross-cultural scenarios and makes it difficult to adapt to the semantic expression diversity caused by cultural differences. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based multilingual method for accurate user demand prediction and marketing decision-making to solve the prediction bias problem caused by insufficient alignment of cultural semantics in cross-cultural user demand prediction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an AI-based multilingual method for accurate user demand prediction and marketing decision-making. The method includes: collecting multilingual user data and preprocessing it to extract initial text features from the preprocessed multilingual user data; inputting the initial text features into a cultural semantic alignment network to generate semantic representation vectors; calculating the semantic correlation vector between the semantic representation vectors and preset cultural dimension vectors; projecting the semantic representation vectors onto a unified cultural-neutral semantic space based on the semantic correlation vectors to generate semantic alignment vectors; constructing a demand prediction model; inputting the semantic alignment vectors into the demand prediction model for multi-dimensional analysis to output accurate user demand prediction results; combining user background data with the accurate user demand prediction results to generate a marketing element set; inputting the marketing element set into a multilingual generative AI model to generate marketing content; adaptively reconstructing and verifying the consistency of the marketing content based on the semantic alignment vectors to output personalized marketing content; distributing the personalized marketing content to users through their preferred channels; and tracking user feedback behavior data in real time, using the user feedback behavior data as new multilingual user data.

[0008] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for collecting multilingual user data, preprocessing it, and extracting the initial text features of the preprocessed multilingual user data are as follows.

[0009] The multilingual user data includes user text data, user behavior data, and user background data;

[0010] The preprocessing includes data cleaning, deduplication, format standardization, language recognition, and tagging.

[0011] Multi-level feature extraction is performed on the preprocessed multilingual user data to extract surface language pattern features and local semantic association features, and then merge them into intermediate features.

[0012] The decision tree model is used to enhance intermediate features and remove redundancy to generate initial text features.

[0013] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for generating the semantic alignment vector are as follows:

[0014] The initial text features are subjected to a nonlinear transformation to generate a semantic representation vector.

[0015] The semantic correlation vector between the semantic representation vector and the preset cultural dimension vector is calculated using a multi-head cross-cultural attention mechanism.

[0016] By using a gated residual cultural fusion network, the semantic relevance vector and the semantic representation vector are fused with cultural context to generate a cultural context enhancement vector.

[0017] A cultural bias subspace is constructed based on the sparsity constraint of the semantic association degree vector, and an orthogonal complement projection matrix is ​​constructed in the cultural bias subspace through an incremental Gram-Schmidt orthogonalization process.

[0018] The cultural context enhancement vector is projected onto a unified cultural neutral semantic space using an orthogonal complement projection matrix, generating a semantic alignment vector.

[0019] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for constructing the demand prediction model are as follows:

[0020] A static feature encoding layer is constructed based on a convolutional neural network, a dynamic behavior encoding layer is constructed based on a long short-term memory network, and a relation encoding layer is constructed based on a graph convolutional network.

[0021] By fusing the static feature encoding layer, dynamic behavior encoding layer, and relation encoding layer with gated attention, a demand prediction model is formed.

[0022] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method of this invention, the specific steps for outputting the accurate user demand prediction result are as follows:

[0023] The semantic alignment vector is input into the static feature encoding layer for multi-scale feature extraction and enhancement to generate enhanced static features.

[0024] The enhanced static features are input into the dynamic behavior encoding layer to perform temporal pattern mining and generate spatiotemporal dynamic features.

[0025] Spatiotemporal dynamic features are input into the relation encoding layer for graph structure propagation and aggregation to generate higher-order relation features;

[0026] Multi-head cross-modal attention fusion is performed to enhance static features, spatiotemporal dynamic features, and higher-order relation features to generate multi-dimensional user features.

[0027] Based on multi-dimensional user characteristics, the probability distribution of each demand dimension is calculated in parallel through a multi-task prediction network, and accurate user demand prediction results are output.

[0028] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the accurate user demand prediction result includes probability distribution vectors of multiple demand dimensions, wherein each demand dimension corresponds to a demand category, and the probability value of each demand dimension represents the strength of the demand category.

[0029] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for generating the marketing element set are as follows:

[0030] Based on multiple demand dimensions and corresponding demand categories, corresponding marketing element template sets are mapped from a predefined marketing strategy knowledge base;

[0031] By fusing user background data with the probability distribution vectors of multiple demand dimensions in the accurate user demand prediction results, joint features are generated.

[0032] Based on the joint features, the adaptation weight of each template in the marketing element template set is calculated;

[0033] The marketing element templates are combined according to the adaptation weight to generate a set of marketing elements.

[0034] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for generating marketing content are as follows:

[0035] The marketing element set is converted into cue vectors recognizable by a multilingual generative AI model through an encoder network, and the initial marketing content is generated by decoding the data through the multilingual generative AI model.

[0036] The initial marketing content undergoes multilingual grammar validation and style unification processing to generate marketing content.

[0037] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for outputting personalized marketing content are as follows:

[0038] Based on semantic alignment vectors, marketing content is deconstructed and reconstructed through a multi-granularity integral attention mechanism to generate culturally adaptive content fragments;

[0039] By using a cybernetics-based feedback optimization mechanism, dynamic consistency checks are performed on culturally adaptable content fragments to generate optimized marketing content.

[0040] By integrating optimized marketing content with user background data through cross-modal gating, personalized marketing content can be generated.

[0041] As a preferred embodiment of the AI-based multilingual accurate user demand prediction and marketing decision-making method described in this invention, the specific steps for using feedback behavior data as new multilingual user data are as follows:

[0042] Personalized marketing content is dynamically allocated to and distributed to user-preferred channels through multi-agent reinforcement learning;

[0043] The user feedback behavior data generated after distribution is tracked and adaptively adjusted in real time using cybernetics methods.

[0044] By using multimodal coding and incremental learning, the tracked user feedback behavior data is used as new multilingual user data.

[0045] The beneficial effects of this invention are as follows: By generating semantic alignment vectors through a cultural semantic alignment network, culturally neutral semantic representation vectors for multilingual user data are achieved; by calculating the correlation between the semantic representation vectors and preset cultural dimension vectors, and projecting them onto a unified culturally neutral space, the interference of cultural background on semantic expression is effectively eliminated, improving the quality of input features for demand prediction models; as a culturally independent standardized representation, the semantic alignment vector provides a purer and more comparable semantic foundation for demand prediction, thereby improving the accuracy and consistency of cross-cultural user demand analysis; through cultural dimension vectors and semantic projection mechanisms, the adaptability to multicultural scenarios is enhanced, providing reliable technical support for global marketing decisions. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of an AI-based multilingual method for accurate user demand prediction and marketing decision-making.

[0048] Figure 2 This is a flowchart for data preprocessing and initial feature extraction.

[0049] Figure 3 A flowchart for generating semantic alignment vectors.

[0050] Figure 4 A flowchart for building a demand forecasting model and outputting accurate user demand forecasting results. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for accurate user demand prediction and marketing decision-making based on AI multilingualism, including the following steps:

[0055] S1. Collect multilingual user data, preprocess it, and extract the initial text features of the preprocessed multilingual user data.

[0056] S1.1: Multilingual user data includes user text data, user behavior data, and user background data;

[0057] It should be noted that user text data refers to user-generated text content, such as social media posts and customer service conversation logs;

[0058] User behavior data refers to user action records, such as page browsing history and button click sequences;

[0059] User background data refers to user attribute information, such as age range and geographical location.

[0060] S1.2: Preprocessing includes data cleaning, deduplication, format standardization, language recognition, and tagging;

[0061] Specifically, the process involves data cleaning of multilingual user data, including removing special characters (such as commas, periods, and @ symbols) and irrelevant symbols from user text data and correcting spelling errors; filtering invalid or abnormal records from user behavior data; filling missing values ​​in user background data (e.g., using the mean or mode); deduplication, removing duplicate user records based on the user's unique identifier; format standardization of multilingual user data, converting date fields to "year-month-day" format, time fields to "hour:minute:second" format, and numeric fields to decimal or integer representation; language recognition of user text data, using a character n-gram statistical language recognition algorithm to identify the language type; and tagging of multilingual user data, adding language tags such as "English" or "Chinese" to user text data, adding behavior type tags such as "browsing" or "purchasing" to user behavior data, and adding attribute tags such as "age range_25-35" to user background data.

[0062] S1.3: Perform multi-level feature extraction on the preprocessed multilingual user data, extract surface language pattern features and local semantic association features, and fuse them into intermediate features;

[0063] Specifically, surface language pattern features are extracted, including word frequency statistics, average sentence length, and punctuation density of user text data; local semantic association features are extracted, and text vectors are generated from local semantic association features using the word frequency-inverse document frequency method to capture word co-occurrence relationships; surface language pattern features and local semantic association features are concatenated into a multi-dimensional feature vector to form intermediate features.

[0064] S1.4: The intermediate features are enhanced and redundant features are removed using a decision tree model to generate initial text features.

[0065] The pre-training process of the decision tree model is as follows: The decision tree model is pre-trained using historical data, which consists of multiple intermediate features and corresponding sample labels. The sample labels are known user demand category identifiers. The decision tree model iterates through all intermediate features, evaluates the classification ability of each intermediate feature for the sample labels, and selects the intermediate feature with the strongest classification ability as the root node for data splitting. The selection process is recursively repeated at each child node, continuing to select the best intermediate feature for splitting, with the goal of making the sample labels in each leaf node as consistent as possible. The splitting process continues until a predefined stopping condition is met (a set rule for terminating the continued growth of the tree structure, such as the number of samples in a leaf node falling below a certain value, or the maximum depth of the decision tree from the root node to the leaf node reaching a certain value), thus obtaining the pre-trained decision tree model.

[0066] Specifically, the decision tree model assigns an importance score to each intermediate feature based on the degree of correlation between the intermediate feature and the sample label; it sets an importance score threshold, retains intermediate features with importance scores higher than the threshold, and removes intermediate features with importance scores lower than the threshold; it then performs a normalization operation on the retained intermediate features, adjusting the values ​​of each intermediate feature to the same range, thus enhancing the intermediate features and obtaining the initial text features.

[0067] It should be noted that the importance score threshold is set based on the statistical distribution of the contribution of intermediate features in historical data to the classification of user demand categories. It usually takes a value in the range of (0,1). The value is determined by balancing the accuracy and redundancy of intermediate feature selection on the validation set through cross-validation to ensure that it retains the ability to significantly distinguish demand predictions.

[0068] S2. Input the initial text features into the cultural semantic alignment network to generate a semantic representation vector. Calculate the semantic correlation vector between the semantic representation vector and the preset cultural dimension vector. Based on the semantic correlation vector, project the semantic representation vector into a unified cultural neutral semantic space to generate a semantic alignment vector.

[0069] S2.1: Perform a nonlinear transformation on the initial text features to generate a semantic representation vector;

[0070] It should be noted that the pre-training process of the fully connected layer is completed using historical data, which consists of multiple intermediate features and corresponding sample labels. The intermediate features from the historical data are fed into the fully connected layer, and a prediction result is generated through forward operation. The prediction result is compared with the sample label to obtain the difference. Based on the difference, the weight matrix and bias vector of the fully connected layer are successively adjusted through the backpropagation algorithm to gradually reduce the difference between the prediction result and the sample label. When the difference between the prediction result and the sample label is less than a preset difference value (the preset difference value is set based on the convergence target of the difference between the model prediction result and the sample label during the pre-training process of the fully connected layer, and its purpose is to stop training when the difference is small enough to indicate that the fully connected layer has learned sufficiently; the example value is 0.001), the pre-training process stops, and at this time the weight matrix and bias vector of the fully connected layer are determined.

[0071] Specifically, the initial text features are combined with the weight matrix of the fully connected layer and a bias vector is added to obtain the linear transformation result. The ReLU nonlinear activation function is applied to the linear transformation result, setting negative values ​​to zero and retaining positive values, thereby completing the nonlinear transformation and generating the semantic representation vector.

[0072] S2.2: Calculate the semantic relevance vector between the semantic representation vector and the preset cultural dimension vector using a multi-head cross-cultural attention mechanism. The expression is:

[0073] ;

[0074] In the formula, Represents the semantic relevance vector. Represents a semantic representation vector. This represents a matrix composed of predefined cultural dimension vectors. This represents the matrix transpose operation. This represents the dimension of the semantic representation vector and each preset cultural dimension vector.

[0075] It should be noted that the preset cultural dimension vector is constructed based on the cultural dimensions in cross-cultural psychology theory (such as Hofstede's cultural dimensions). Each dimension corresponds to a vector with a fixed direction and semantic meaning, representing a set of vectors with different cultural tendencies, such as a vector representing the "individualism-collectivism" cultural dimension.

[0076] S2.3: By using a gated residual cultural fusion network, the semantic relevance vector and the semantic representation vector are fused with cultural context to generate a cultural context-enhanced vector;

[0077] It should be noted that the pre-training process of the gated residual cultural fusion network is as follows: Pre-training is completed using historical data; the historical data consists of historical semantic representation vectors, corresponding semantic relevance vectors, and sample labels; the historical semantic representation vectors and corresponding semantic relevance vectors are input into the gated residual cultural fusion network to generate prediction results of cultural context enhancement vectors; the prediction results are compared with the sample labels, and the parameters of the fully connected layers in the gated residual cultural fusion network are adjusted through the backpropagation algorithm so that the prediction results continuously approach the sample labels until they are less than the preset convergence threshold (the preset convergence threshold is set based on the decreasing trend of the difference between the prediction results and the sample labels (such as cross-entropy or mean square error) in multiple pre-training on historical data, and the value range is usually between 0 and 1, and the value is selected based on balancing training efficiency and network generalization ability). At this point, the pre-training process is completed, and the pre-trained gated residual cultural fusion network is obtained.

[0078] Specifically, semantic relevance vectors are used as weight coefficients to weight and combine preset cultural dimension vectors to generate cultural context vectors. A gated residual cultural fusion network is used to concatenate the semantic representation vectors and cultural context vectors, and a gating value is generated through a fully connected layer and a sigmoid activation layer in the gating mechanism. The gating value controls the degree of fusion of the cultural context vectors. The gating value is then combined element-wise with the cultural context vectors to obtain adjusted cultural context information. The gated residual cultural fusion network connects the adjusted cultural context information with the semantic representation vectors through residual connections to generate a cultural context enhancement vector.

[0079] S2.4: Construct a cultural bias subspace based on the sparsity constraint of the semantic association degree vector, and construct an orthogonal complement projection matrix in the cultural bias subspace through an incremental Gram-Schmidt orthogonalization process;

[0080] It should be noted that the sparsity constraint of the semantic relevance vector means that only a few elements in the semantic relevance vector have significantly non-zero values, thus allowing the selection of preset cultural dimension vectors with high relevance.

[0081] The incremental Gram-Schmidt orthogonalization process refers to processing a sequence of vectors one by one, removing the projection of each vector onto the previously orthogonalized vectors and normalizing it, thereby generating a set of standard orthogonal vectors.

[0082] Specifically, based on the sparsity constraint of semantic relevance vectors, cultural dimension vectors with semantic relevance vectors higher than a preset threshold are selected from the preset cultural dimension vectors to form a vector group; the linear space spanned by the vector group is defined as the cultural bias subspace; within the cultural bias subspace, a set of linearly independent vectors is selected sequentially from the vector group as the basis vector sequence to be orthogonalized; each basis vector in the basis vector sequence to be orthogonalized is processed sequentially through an incremental Gram-Schmidt orthogonalization process. For the first basis vector in the basis vector sequence to be orthogonalized, ... The first standard orthogonal basis vector is generated by directly performing a normalization operation. For each subsequent basis vector in the sequence to be orthogonalized, the projection components of the basis vector in all previously generated standard orthogonal basis vector directions are removed to generate an orthogonalization result. Then, the orthogonalization result is normalized to generate a new standard orthogonal basis vector. This process is continued sequentially until all basis vectors in the sequence to be orthogonalized have been processed, resulting in a complete set of standard orthogonal bases. The standard orthogonal basis matrix formed by this set of standard orthogonal bases is the orthogonal complement projection matrix.

[0083] It should be noted that the semantic relevance vector threshold is obtained by running a multi-head cross-cultural attention mechanism on historical multilingual user data to obtain the semantic relevance vector of each sample, statistically analyzing the relevance distribution, and statistically analyzing the semantic relevance values ​​of each cultural dimension across all samples. Inflection point regions that significantly distinguish between "relevant" and "irrelevant" cultural dimensions are selected, and the construction effect of the cultural bias subspace under different thresholds is tested on the validation set. Downstream tasks (such as the accuracy of culturally adaptive content generation) are used as evaluation metrics to select subspace dimensions that ensure stable performance for downstream tasks and reasonable subspace dimensions. The value is set based on the range of 0.3-0.45. The value is determined by statistical analysis of the semantic relevance values ​​of each cultural dimension in a large-scale historical multilingual user data (covering more than 10 languages ​​and more than 20 cultural regions). The semantic relevance can effectively filter random noise and weakly correlated dimensions, while retaining signals with actual cultural representation significance. At the same time, in the downstream cultural adaptive content generation task, this range can achieve the best balance between accuracy and subspace dimension - when it is below 0.30, redundancy increases significantly, and when it is above 0.45, key cultural signals begin to be lost.

[0084] S2.5: Project the cultural context enhancement vector onto a unified cultural neutral semantic space using an orthogonal complement space projection matrix to generate a semantic alignment vector.

[0085] It should be noted that the culturally neutral semantic space refers to the semantic space in which the cultural context enhancement vector is spatially mapped by the orthogonal complement projection matrix. This eliminates the differences in semantic expression caused by different cultural backgrounds, making the semantic representations from multilingual users comparable under a unified benchmark.

[0086] Specifically, the orthogonal complement projection matrix corresponds to a subspace perpendicular to the cultural bias direction; the cultural context enhancement vector is spatially mapped to the orthogonal complement projection matrix, so that the information related to cultural bias in the components of the cultural context enhancement vector is suppressed; the mapped cultural context enhancement vector is located in a unified cultural neutral semantic space, retaining cross-cultural consistent semantic information, forming a semantic alignment vector.

[0087] S3. Construct a demand forecasting model, input the semantic alignment vector into the demand forecasting model for multi-dimensional analysis, and output accurate user demand forecasting results.

[0088] S3.1: Construct a static feature encoding layer based on a convolutional neural network, a dynamic behavior encoding layer based on a long short-term memory network, and a relation encoding layer based on a graph convolutional network;

[0089] Specifically, by stacking multiple convolutional layers and non-linear activation layers, and using convolutional kernels with different receptive fields, hierarchical extraction of local semantic patterns is performed on the semantic alignment vector. After each convolutional operation, feature regions with high response intensity are retained, thus forming a static feature encoding layer. User behavior data is arranged in chronological order, and through the state propagation mechanism of forget gate and update gate, long-term memory of key behavioral events is maintained during sequence traversal, thus forming a dynamic behavior encoding layer. A user relationship network is constructed with users as nodes and user interaction behaviors as edges. By aggregating the feature information of the directly connected neighbors of each node, the node representation is updated layer by layer, thus forming a relationship encoding layer.

[0090] It should be noted that key behavioral events refer to user actions with clear intent in a time sequence, such as staying on a page for more than a preset time, completing a form, adding items to the cart, making a payment, and visiting the same product details page multiple times consecutively.

[0091] S3.2: Gated attention is used to fuse the static feature encoding layer, dynamic behavior encoding layer and relation encoding layer to form a demand prediction model.

[0092] Specifically, the feature representations in the static feature encoding layer, dynamic behavior encoding layer, and relation encoding layer are aligned. The importance of the three is dynamically adjusted through a gating attention mechanism. Attention weights are used to highlight feature sources that are more relevant to the current demand prediction task. This enables the static, dynamic, and relation features to be synergistically integrated in the channel and temporal dimensions to form a unified fusion feature structure, thereby forming a demand prediction model.

[0093] It should be noted that the pre-training process of the demand prediction model is as follows: Historical user data is used as samples, which include semantic alignment vectors, user background data, and corresponding real user demand category identifiers; the semantic alignment vectors are sequentially passed through a static feature encoding layer, a dynamic behavior encoding layer, and a relation encoding layer to generate their respective encoding structures; the three types of encoding structures are integrated through a gating attention fusion mechanism; the integrated feature structure is used to perform matching learning on known demand categories, and the parameters of each layer are adjusted to improve matching accuracy, thereby completing the pre-training of the demand prediction model.

[0094] S3.3: Input the semantic alignment vector into the static feature encoding layer for multi-scale feature extraction and enhancement to generate enhanced static features;

[0095] Specifically, the semantic alignment vector is sequentially passed through multiple convolutional kernels with different receptive fields in the static feature encoding layer to capture short-range and long-range patterns in local semantic segments, forming multi-scale local feature responses. Nonlinear activation operations are used to enhance the expressive differences of multi-scale local feature responses and retain response regions that are discriminative to user needs. The multi-scale convolution results are stacked hierarchically and integrated into channels to generate enhanced static features.

[0096] It should be noted that short-range and long-range patterns refer to the co-occurrence relationship between adjacent words within a local segment and the semantic association between words at a distance across sentences or paragraphs in the semantic alignment vector.

[0097] S3.4: Input the enhanced static features into the dynamic behavior encoding layer to perform temporal pattern mining and generate spatiotemporal dynamic features;

[0098] Specifically, the static features are aligned with the records of user behavior data arranged in time series. In the dynamic behavior coding layer, the state information is transmitted in time period by time through the gating mechanism. The memory of key behavioral events is retained and the influence of irrelevant operations is suppressed. The sequence and interval patterns of behavior occurrence are captured to reflect the evolution of user interests. Combined with the behavior clustering pattern in the time dimension, a spatiotemporal dynamic feature that reflects the temporal dependence of behavior and the spatial distribution characteristics of behavior is formed.

[0099] It should be noted that state information refers to the internal representation that reflects the user's current interests and historical behavior memory, which is retained by the dynamic behavior coding layer through a gating mechanism during the process of traversing the user's behavior sequence.

[0100] Behavioral clustering patterns refer to the repetitive patterns in which similar behaviors of users occur in a concentrated manner within a specific time period or in a specific scenario.

[0101] S3.5: Input spatiotemporal dynamic features into the relation encoding layer for graph structure propagation and aggregation to generate higher-order relation features;

[0102] Specifically, spatiotemporal dynamic features are used as the initial representation of each node in the user relationship network, which is constructed with users as nodes and user interactions as edges. In the relationship encoding layer, each user's spatiotemporal dynamic features are exchanged with the spatiotemporal dynamic features of its directly connected neighboring users, allowing the current user's initial representation to absorb the behavioral patterns of its neighboring users. Through multiple rounds of exchange, each user's initial representation gradually integrates the behavioral features of first-order, second-order, and even higher-order neighboring users, forming high-order relationship features that include social influence, group preferences, and network location characteristics.

[0103] S3.6: Enhanced static features, spatiotemporal dynamic features, and higher-order relationship features are fused using multi-head cross-modal attention to generate multi-dimensional user features;

[0104] Specifically, the system aligns enhanced static features, spatiotemporal dynamic features, and higher-order relational features. Enhanced static features reflect stable preferences in user text, spatiotemporal dynamic features embody the temporal evolution of user behavior, and higher-order relational features characterize the user's group influence in the relational network. Through a multi-head cross-modal attention mechanism, the system establishes associations between different components of the three types of features. Each attention head focuses on feature matching relationships at different semantic granularities, capturing the interaction patterns between static attributes, dynamic behaviors, and social associations. The results of each attention head are spliced ​​and integrated to generate multi-dimensional user features.

[0105] S3.7: Based on multi-dimensional user characteristics, the probability distribution of each demand dimension is calculated in parallel through a multi-task prediction network, and the accurate user demand prediction results are output.

[0106] Specifically, based on multi-dimensional user characteristics, a multi-task prediction network is used to synchronously map multiple demand dimensions. Each demand dimension corresponds to an independent mapping path. Each path uses the corresponding weight matrix and bias vector to convert the user's multi-dimensional feature vector into an intermediate representation that matches the number of demand dimension categories. The intermediate representation reflects the user's preference for each category under the corresponding demand dimension. Then, through normalization, each preference degree forms a complete distribution structure, forming the probability distribution vector of the k-th demand dimension. The probability distribution vectors of all demand dimensions together constitute the accurate user demand prediction result.

[0107] Based on multi-dimensional user characteristics, the probability distribution of each demand dimension is calculated in parallel through a multi-task prediction network, and the expression is:

[0108] ;

[0109] ;

[0110] ;

[0111] In the formula, Indicates the first A probability distribution vector for each demand dimension. Indicates the demand dimension index. Indicates the first The weight matrix corresponding to each demand dimension Represents a user's multi-dimensional feature vector. Indicates the first The bias vector corresponding to each demand dimension Represents the set of real numbers. Indicates the first The number of categories in each demand dimension Represents the user's multi-dimensional feature vector Dimensions.

[0112] S3.8: The accurate user demand prediction results include probability distribution vectors for multiple demand dimensions, where each demand dimension corresponds to a demand category, and the probability value of each demand dimension represents the strength of the demand category.

[0113] It should be noted that the accurate user demand prediction results are composed of probability distribution vectors of multiple demand dimensions. Each demand dimension corresponds to a type of user demand, such as power performance, appearance design, or price sensitivity. Each demand dimension contains several specific demand categories. For example, "long battery life" and "fast charging support" are different categories under the power performance demand dimension. Each value in the probability distribution vector corresponds to the strength of a demand category. The higher the value, the stronger the user's preference for the demand category.

[0114] S4. Combine user background data with accurate user demand prediction results to generate a set of marketing elements.

[0115] S4.1: Based on multiple demand dimensions and corresponding demand categories, map the corresponding marketing element template set from the predefined marketing strategy knowledge base;

[0116] Specifically, based on the multiple demand dimensions and the demand categories with high probability values ​​under each dimension from the accurate user demand prediction results, the data is matched with the rule entries in the predefined marketing strategy knowledge base. The predefined marketing strategy knowledge base stores the preset association relationships between each demand category and the marketing element template. By searching for the marketing element templates corresponding to each high probability demand category, the multiple marketing element templates found are collected to form a marketing element template set that matches the current user demand.

[0117] It should be noted that the predefined marketing strategy knowledge base refers to a collection that stores the preset relationships between various demand categories and marketing element templates, as well as stored cultural adaptation and alternative expression rules. Each marketing element template is configured with a corresponding preset structured feature vector, which is used to match the corresponding marketing strategy template according to user needs. The preset relationship is a mapping rule or index structure that indicates which marketing element templates should be triggered for a certain demand category.

[0118] S4.2: Fuse user background data with the probability distribution vectors of multiple demand dimensions in the accurate user demand prediction results to generate joint features;

[0119] Specifically, the probability distribution vectors of multiple demand dimensions in the accurate user demand prediction results are concatenated along the feature dimension. The user background data includes attribute information such as the user's age range and geographical location. The probability distribution vectors of multiple demand dimensions reflect the user's preference intensity in each demand category. The concatenated user background data and the probability distribution vectors of multiple demand dimensions in the accurate user demand prediction results also include the user's demographic attributes and demand tendency information, forming a joint feature that can comprehensively reflect the user's individual attributes and demand structure.

[0120] S4.3: Based on joint features, calculate the adaptation weight of each template in the marketing element template set;

[0121] ;

[0122] In the formula, This indicates the first marketing element template in the collection. The adaptation weight of each marketing element template This indicates the marketing element template index. This represents the adjustment coefficient. Represents the joint eigenvector With the Preset structured feature vectors of each marketing element template Cosine similarity between them

[0123] Represents the joint eigenvector. Indicates the first Feature vector of each marketing element template The complement of the adjustment coefficient, This represents the summation index, used to iterate through all templates in the marketing element template set. This indicates the total number of templates in the marketing element template set. Indicates the first The pre-defined structured feature vectors of each marketing element template, and Similarly, it is used for iterations in summation calculations.

[0124] It should be noted that the adjustment coefficient comes from a predefined marketing strategy knowledge base. The example value is 0.5, which is based on balancing the contribution ratio of feature matching degree and demand probability in the weight calculation.

[0125] S4.4: Combine marketing element templates according to the adaptation weight to generate a marketing element set.

[0126] Specifically, based on the adaptation weight of each template in the marketing element template set, all marketing element templates are sorted from high to low according to their adaptation weight. Marketing element templates whose adaptation weight reaches the preset reference level threshold are selected. The selected marketing element templates are then arranged and combined according to their priority order in user needs to form a structured marketing element set.

[0127] It should be noted that the reference level threshold was determined through validation on historical user data. Multiple representative user samples were selected, and the distribution characteristics of their marketing element template adaptation weights were statistically analyzed. Combined with the requirements for content conciseness and coverage in actual marketing scenarios, a weight value that retains marketing elements while excluding low-relevance templates was set as the reference level threshold. The reference level threshold ranges from 0.4 to 0.8, determined based on the statistical distribution of marketing element template adaptation weights in a large number of historical user samples. The value of 0.4 is based on the fact that, in the validation set, the content corresponding to templates with adaptation weights below 0.4 mostly failed to attract user attention or generate conversions, indicating a weak correlation with user needs. The value of 0.8 is based on the fact that the number of templates with adaptation weights above 0.8 is small, and their frequency of occurrence is low, resulting in most users being unable to generate an effective set of marketing elements.

[0128] S5. Input the set of marketing elements into the multilingual generative AI model to generate marketing content, and perform adaptive reconstruction and consistency verification of the marketing content based on semantic alignment vectors to output personalized marketing content.

[0129] S5.1: The marketing element set is converted into cue vectors that can be recognized by a multilingual generative AI model through an encoder network, and the initial marketing content is generated by decoding through the multilingual generative AI model;

[0130] It should be noted that the pre-training process of the multilingual generative AI model is as follows: A large-scale historical multilingual text corpus is accessed, containing publicly available text data in multiple languages, such as web page content, news reports, and encyclopedia entries. Through a masked language modeling task, the multilingual generative AI model learns to recover the original content based on context while masking some words, mastering cross-lingual vocabulary, grammar, and semantic rules. Building on this, a next-sentence prediction task enhances the multilingual generative AI model's understanding of text coherence. After pre-training, the multilingual generative AI model possesses the ability to generate fluent, grammatically correct multilingual text.

[0131] Specifically, the elements in the marketing element set are arranged into a natural language description sequence according to a preset structure. The vocabulary and grammatical structure in the natural language description sequence are feature-encoded by an encoder network to form a vector representation containing semantic structure and intent information. The vector representation is passed to the multilingual generative AI model as a cue vector. The multilingual generative AI model generates text content that conforms to grammar and context word by word according to the semantic information contained in the cue vector, forming initial marketing content covering the key marketing points.

[0132] It should be noted that semantic information refers to the intent, conceptual associations, and contextual logical relationships of each element in the marketing element set contained in the prompt vector, reflecting the theme and key points of the content to be generated.

[0133] S5.2: Perform multilingual grammar checks and style unification on the initial marketing content to generate marketing content.

[0134] Specifically, the initial marketing content undergoes multilingual grammar verification. A multilingual grammar checker identifies syntactic errors, incorrect word order, and collocation abnormalities in texts of different languages, and corrects them according to language norms. Style unification is achieved by comparing the initial marketing content with the expression patterns in the preset brand style, adjusting word choice, tone intensity, and sentence complexity to ensure that the content in different languages ​​maintains consistency in formality, emotional tone, and information density, thus forming marketing content that aligns with the brand tone.

[0135] It should be noted that the preset brand style refers to the pre-defined expression pattern that the brand should maintain in marketing content, including the unified standard of language features such as word choice, tone intensity, sentence complexity, and emotional coloring.

[0136] S5.3: Based on the semantic alignment vector, marketing content is deconstructed and reconstructed through a multi-granularity integral attention mechanism to generate culturally adaptive content fragments;

[0137] Specifically, based on the cross-cultural consistency semantics in the culturally neutral semantic space represented by the semantic alignment vector, a multi-granularity integral attention mechanism is used to scan the marketing content layer by layer at the lexical, phrase, and sentence levels. This identifies expressions that conflict with or are incompatible with the target language's cultural habits. Combining these with the culturally appropriate alternative expression rules stored in the predefined marketing strategy knowledge base, the identified expressions that conflict with or are incompatible with the target language's cultural habits are replaced with semantically equivalent expressions that conform to the target culture's acceptance habits. Without changing the core information carried by the original marketing element set, an expression structure that conforms to the target language's cultural habits is generated, forming culturally adapted content fragments.

[0138] S5.4: Dynamically verify the consistency of culturally adaptable content fragments through a cybernetics feedback optimization mechanism to generate optimized marketing content;

[0139] Specifically, a cybernetics feedback optimization mechanism is used to perform multiple rounds of verification on culturally adaptable content fragments. These fragments are compared with the language features in the preset brand style to identify deviations in tone intensity, word choice, or emotional expression. Based on the cross-cultural consistency expression benchmark indicated by the semantic alignment vector, the deviations are adjusted to ensure that the content remains culturally adaptable without deviating from the brand's unified expression norms. After multiple rounds of comparison and correction, optimized marketing content that conforms to the brand's tone and is culturally compatible is formed.

[0140] It should be noted that the cybernetics feedback optimization mechanism is an iterative optimization method based on continuous comparison and dynamic adjustment. By repeatedly comparing the current output results with the preset target standards, it identifies the existing deviations and makes targeted corrections to the content according to the type and degree of deviations. In the process of generating culturally adaptive content, the preset brand style and semantic alignment vector are used as reference benchmarks to conduct multiple rounds of verification on the language features, expression tendencies and semantic consistency of culturally adaptive content fragments, ensuring that the modified expression not only conforms to the acceptance habits of the target culture, but also does not deviate from the brand's unified communication positioning.

[0141] S5.5: By integrating optimized marketing content and user background data through cross-modal gating, personalized marketing content is generated.

[0142] Specifically, cross-modal gating integrates the language structure of optimized marketing content with user background data, including the user's age range, geographical location, and language preferences. Based on this data, cross-modal gating dynamically adjusts the expression, title selection, and example scenarios in the optimized marketing content, ensuring that the usage context and recommendation reasons in the content match the user's living environment and identity characteristics, thus creating personalized marketing content that is tailored to each individual.

[0143] It should be noted that the cross-modal gating adopts a gated attention structure, which includes a fully connected layer and a sigmoid activation layer. It receives optimized marketing content and user background data, and generates gating weights between 0 and 1 to adjust the retention strength of each word or phrase in the optimized marketing content. The parameters of the cross-modal gating are obtained through end-to-end training on historical personalized marketing samples. The training objective is to improve the click-through rate or conversion rate of users for personalized marketing content.

[0144] S6. Distribute personalized marketing content to users through their preferred channels and track user feedback behavior data in real time, using this data as new multilingual user data.

[0145] S6.1: Dynamically allocate and distribute personalized marketing content to user-preferred channels through multi-agent reinforcement learning;

[0146] Specifically, personalized marketing content is handed over to the agents in the multi-agent reinforcement learning framework for collaborative decision-making. Each agent corresponds to an optional user preference channel, such as social media, email, or mobile application push. Based on historical distribution experience, each agent evaluates the degree of user response that personalized marketing content may generate after being delivered through its assigned channel. Through multiple rounds of strategy negotiation, the channel combination most likely to generate positive feedback is determined, and the personalized marketing content is distributed through the channel combination.

[0147] It should be noted that each agent in the multi-agent reinforcement learning framework adopts a deep Q-network structure. The state space consists of the historical click rate of the user's preferred channel, recent activity, and channel load. The action space is whether to select the channel for distribution. The reward signal is the click or conversion event in the user's feedback behavior data after distribution. Each agent is jointly trained by sharing an experience replay buffer. Policy negotiation is achieved through a weighted voting mechanism, and the weights are dynamically adjusted by the historical response effects of each channel.

[0148] S6.2: Real-time feedback tracking and adaptive adjustment of user feedback behavior data generated after distribution using cybernetics methods;

[0149] Specifically, the system continuously monitors user feedback behavior data generated after distribution using cybernetics methods, periodically compares the user feedback behavior data with the preset response targets, identifies the deviation between actual feedback and expectations, and makes targeted adjustments to subsequent content generation and distribution strategies based on the direction and degree of deviation. This includes optimizing the expression of marketing content, changing distribution channels, or adjusting the timing of distribution, so that the overall marketing process gradually converges to a high-response state.

[0150] It should be noted that the preset response target refers to collecting user feedback behavior data from historical marketing activities, classifying and tagging the user feedback behavior data, determining the core business focus behavior indicators, statistically analyzing the performance distribution of the selected behavior indicators in historical data, and combining the response target value range set by the business expectations to solidify the response target value range into the preset response target; including observable user behavior standards such as click-through rate, conversion rate, or dwell time.

[0151] S6.3: Use the tracked user feedback behavior data as new multilingual user data through multimodal coding and incremental learning.

[0152] Specifically, the tracked user feedback behavior data is associated with the text, language type, and distribution context of the corresponding marketing content to form a complete record that includes behavioral responses and multilingual contexts. Through multimodal coding, the text content, user behavior patterns, and language attributes are uniformly represented as structured data. After cleaning and alignment, the structured data is added to the multilingual user dataset as new multilingual user data that reflects the actual user reactions, and participates in the incremental learning process of the subsequent demand prediction model.

[0153] In summary, this invention achieves culturally neutral semantic representation vectors for multilingual user data by generating semantic alignment vectors through a cultural semantic alignment network; by calculating the correlation between the semantic representation vectors and preset cultural dimension vectors and projecting them onto a unified culturally neutral space, it effectively eliminates the interference of cultural background on semantic expression and improves the quality of input features for demand prediction models; the semantic alignment vectors, as culturally independent standardized representations, provide a purer and more comparable semantic foundation for demand prediction, thereby improving the accuracy and consistency of cross-cultural user demand analysis; and through cultural dimension vectors and semantic projection mechanisms, it enhances adaptability to multicultural scenarios and provides reliable technical support for global marketing decisions.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based multi-language precise user demand prediction and marketing decision method, characterized in that: include, Collect multilingual user data, preprocess it, and extract the initial text features of the preprocessed multilingual user data; The initial text features are input into the cultural semantic alignment network to generate semantic representation vectors. The semantic relevance vector between the semantic representation vector and the preset cultural dimension vector is calculated. Based on the semantic relevance vector, the semantic representation vector is projected into a unified cultural neutral semantic space to generate semantic alignment vectors. Construct a demand forecasting model, input semantic alignment vectors into the demand forecasting model for multi-dimensional analysis, and output accurate user demand forecasting results. By combining user background data with accurate user demand prediction results, a set of marketing elements is generated; The marketing elements are input into a multilingual generative AI model to generate marketing content. Based on the semantic alignment vector, the marketing content is adaptively reconstructed and its consistency is verified to output personalized marketing content. Personalized marketing content is distributed to users through their preferred channels, and user feedback behavior data is tracked in real time, which is then used as new multilingual user data.

2. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The process of collecting multilingual user data, preprocessing it, and extracting the initial text features from the preprocessed multilingual user data involves the following specific steps. The multilingual user data includes user text data, user behavior data, and user background data; The preprocessing includes data cleaning, deduplication, format standardization, language recognition, and tagging. Multi-level feature extraction is performed on the preprocessed multilingual user data to extract surface language pattern features and local semantic association features, and then merge them into intermediate features. The decision tree model is used to enhance intermediate features and remove redundancy to generate initial text features.

3. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for generating the semantic alignment vector are as follows: The initial text features are subjected to a nonlinear transformation to generate a semantic representation vector. The semantic correlation vector between the semantic representation vector and the preset cultural dimension vector is calculated using a multi-head cross-cultural attention mechanism. By using a gated residual cultural fusion network, the semantic relevance vector and the semantic representation vector are fused with cultural context to generate a cultural context enhancement vector. A cultural bias subspace is constructed based on the sparsity constraint of the semantic association degree vector, and an orthogonal complement projection matrix is ​​constructed in the cultural bias subspace through an incremental Gram-Schmidt orthogonalization process. The cultural context enhancement vector is projected onto a unified cultural neutral semantic space using an orthogonal complement projection matrix, generating a semantic alignment vector.

4. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for constructing the demand forecasting model are as follows: A static feature encoding layer is constructed based on a convolutional neural network, a dynamic behavior encoding layer is constructed based on a long short-term memory network, and a relation encoding layer is constructed based on a graph convolutional network. By fusing the static feature encoding layer, dynamic behavior encoding layer, and relation encoding layer with gated attention, a demand prediction model is formed.

5. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for outputting accurate user demand prediction results are as follows. The semantic alignment vector is input into the static feature encoding layer for multi-scale feature extraction and enhancement to generate enhanced static features. The enhanced static features are input into the dynamic behavior encoding layer to perform temporal pattern mining and generate spatiotemporal dynamic features. Spatiotemporal dynamic features are input into the relation encoding layer for graph structure propagation and aggregation to generate higher-order relation features; Multi-head cross-modal attention fusion is performed to enhance static features, spatiotemporal dynamic features, and higher-order relation features to generate multi-dimensional user features. Based on multi-dimensional user characteristics, the probability distribution of each demand dimension is calculated in parallel through a multi-task prediction network, and accurate user demand prediction results are output.

6. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 5, characterized in that: The accurate user demand prediction results include probability distribution vectors for multiple demand dimensions, where each demand dimension corresponds to a demand category, and the probability value of each demand dimension represents the strength of the demand category.

7. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for generating the marketing element set are as follows: Based on multiple demand dimensions and corresponding demand categories, corresponding marketing element template sets are mapped from a predefined marketing strategy knowledge base; By fusing user background data with the probability distribution vectors of multiple demand dimensions in the accurate user demand prediction results, joint features are generated. Based on the joint features, the adaptation weight of each template in the marketing element template set is calculated; The marketing element templates are combined according to the adaptation weight to generate a set of marketing elements.

8. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for generating marketing content are as follows: The marketing element set is converted into cue vectors recognizable by a multilingual generative AI model through an encoder network, and the initial marketing content is generated by decoding the data through the multilingual generative AI model. The initial marketing content undergoes multilingual grammar validation and style unification processing to generate marketing content.

9. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for outputting personalized marketing content are as follows: Based on semantic alignment vectors, marketing content is deconstructed and reconstructed through a multi-granularity integral attention mechanism to generate culturally adaptive content fragments; By using a cybernetics-based feedback optimization mechanism, dynamic consistency checks are performed on culturally adaptable content fragments to generate optimized marketing content. By integrating optimized marketing content with user background data through cross-modal gating, personalized marketing content can be generated.

10. The AI-based multilingual accurate user demand prediction and marketing decision-making method as described in claim 1, characterized in that: The specific steps for using feedback behavior data as new multilingual user data are as follows. Personalized marketing content is dynamically allocated to and distributed to user-preferred channels through multi-agent reinforcement learning; The user feedback behavior data generated after distribution is tracked and adaptively adjusted in real time using cybernetics methods. By using multimodal coding and incremental learning, the tracked user feedback behavior data is used as new multilingual user data.

Images