A multi-language user matching recommendation method and system based on multi-dimensional label fusion

Abstract

The application relates to the technical field of artificial intelligence, and discloses a multi-language user matching recommendation method and system based on multi-dimensional label fusion, which comprises the following steps: extracting language feature symbols and interactive behavior features in multi-dimensional label data to construct multi-dimensional label vectors of users in a digital service platform; performing semantic alignment processing on the multi-dimensional label vectors by using a cross-language fusion layer in a deep learning model to obtain aligned semantic vectors; calculating information coverage corresponding to the aligned semantic vectors by using a coverage function in the deep learning model; determining core label vectors in the multi-dimensional label vectors to generate enhanced semantic representation vectors corresponding to the users; and calculating semantic correlation degrees between the enhanced semantic representation vectors to match out recommendation content corresponding to the users from the digital service platform. The application can improve the accuracy of multi-language user matching recommendation based on multi-dimensional label fusion.

Description

Technical Field

[0001] This invention relates to a multilingual user matching and recommendation method and system based on multidimensional tag fusion, belonging to the field of artificial intelligence technology. Background Technology

[0002] As digital service platforms expand globally, the number of multilingual users from different regions continues to grow. The behavioral data generated by these users on the platform exhibits multilingual and multi-dimensional characteristics, encompassing diverse information such as language preferences, interaction habits, and demand tendencies. These multilingual users have an increasingly urgent need for accurate recommendations. Precise matching and recommendations can not only improve the user experience but also enhance platform user stickiness and conversion efficiency.

[0003] However, most existing multilingual user matching and recommendation schemes rely solely on single-language tags or simple text translation features for matching, failing to fully explore the relationships within users' multi-dimensional tag data. Furthermore, the lack of effective cross-language semantic alignment mechanisms results in inaccurate mapping of information from users in different languages, leading to insufficient information coverage. Additionally, the recommendation process does not specifically enhance core content, ultimately resulting in low efficiency and insufficient recommendation accuracy in multilingual environments, making it difficult to meet the personalized needs of multilingual users. Therefore, a method to improve the accuracy of multilingual user matching and recommendation is needed. Summary of the Invention

[0004] This invention provides a multilingual user matching and recommendation method and system based on multidimensional tag fusion, the main purpose of which is to improve the accuracy of multilingual user matching and recommendation based on multidimensional tag fusion.

[0005] To achieve the above objectives, this invention provides a multilingual user matching and recommendation method based on multidimensional tag fusion, comprising:

[0006] Collect multidimensional tag data of users in the digital service platform, extract language feature symbols and interaction behavior features from the multidimensional tag data, and construct the multidimensional tag vector of users in the digital service platform;

[0007] The multidimensional label vector is input into a pre-trained deep learning model. The cross-language fusion layer in the deep learning model is used to perform semantic alignment processing on the multidimensional label vector to obtain an aligned semantic vector. The information coverage corresponding to the aligned semantic vector is calculated using the coverage function in the deep learning model.

[0008] Based on the information coverage, the core tag vector in the multi-dimensional tag vector is determined, and the enhanced semantic representation vector corresponding to the user is generated by combining the core tag vector and the interaction behavior features.

[0009] Calculate the semantic correlation between the enhanced semantic representation vectors, and based on the semantic correlation and the enhanced semantic representation vectors, match the recommended content corresponding to the user from the digital service platform.

[0010] Optionally, the extraction of linguistic features and interaction behavior features from the multidimensional label data includes:

[0011] The multidimensional label data is cleaned to obtain standardized label data;

[0012] The standardized label data is subjected to data classification processing to obtain language data and behavioral data;

[0013] The language data is parsed to obtain basic semantic units;

[0014] The basic semantic units are subjected to feature encoding processing to obtain language feature symbols;

[0015] The behavioral data is processed by behavioral sequence extraction to obtain interactive behavioral features.

[0016] Optionally, the step of performing semantic unit parsing on the language data to obtain basic semantic units includes:

[0017] The language data is subjected to grammatical structure analysis to obtain a grammatical dependency tree;

[0018] Semantic role annotation is performed on the grammatical dependency tree to obtain a semantic role framework;

[0019] The semantic role framework is subjected to entity linking processing to obtain a set of linked entities;

[0020] The set of linked entities is subjected to relation extraction processing to obtain a semantic relation graph;

[0021] Cluster analysis is performed on the semantic relationship graph to obtain semantic clusters;

[0022] The semantic clusters are processed by extracting core units to obtain basic semantic units.

[0023] Optionally, the step of extracting linguistic features and interaction behavior features from the multidimensional tag data to construct the user's multidimensional tag vector in the digital service platform includes:

[0024] Semantic graphs are constructed from the linguistic features to obtain a semantic relation network;

[0025] The interactive behavior features are analyzed and processed to obtain a behavior pattern map;

[0026] Analyze the structural correspondence between the semantic relationship network and the behavioral pattern graph to obtain cross-modal mapping relationships;

[0027] Based on the cross-modal mapping relationship, the semantic relationship network and the behavioral pattern graph are fused to obtain fused topological features;

[0028] The fused topological features are subjected to feature optimization processing to obtain optimized topological features;

[0029] Based on the optimized topological features, a multidimensional tag vector of the user in the digital service platform is constructed.

[0030] Optionally, the step of using the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multi-dimensional label vector to obtain an aligned semantic vector includes:

[0031] The multidimensional label vector is projected using the feature projection algorithm in the cross-language fusion layer to obtain a projected semantic representation.

[0032] The projected semantic representation is encoded using the attention encoder in the cross-language fusion layer to obtain a semantic feature vector;

[0033] Collect the context data corresponding to the multidimensional label vector, and use the feature-aware network in the cross-language fusion layer to mine the context features in the context data;

[0034] The semantic feature vector and the context features are fused using the semantic fusion network in the cross-language fusion layer to obtain an aligned semantic vector.

[0035] Optionally, calculating the information coverage corresponding to the aligned semantic vector using the coverage function in the deep learning model includes:

[0036] The variance of the alignment semantic vector is calculated to obtain the vector variance value;

[0037] Calculate the information entropy corresponding to each vector in the alignment semantic vector to obtain the vector information entropy;

[0038] Calculate the sparsity of each vector in the alignment semantic vector to obtain the vector sparsity;

[0039] By combining the vector variance, the vector information entropy, and the vector sparsity, the information coverage corresponding to the aligned semantic vector is calculated using the coverage function.

[0040] Optionally, the step of combining the core tag vector and the interaction behavior features to generate the enhanced semantic representation vector corresponding to the user includes:

[0041] Construct a heterogeneous information graph corresponding to the core label vector and the interaction behavior features;

[0042] The heterogeneous information graph is subjected to node feature aggregation processing to obtain aggregated node features;

[0043] Based on the features of the aggregation nodes, an aggregation semantic representation vector corresponding to the user is generated;

[0044] The aggregated semantic representation vector is subjected to semantic enhancement processing to obtain the enhanced semantic representation vector.

[0045] Optionally, the step of aggregating node features on the heterogeneous information graph to obtain aggregated node features includes:

[0046] Identify the node and edge types in the heterogeneous information graph;

[0047] Extract the initial node attributes corresponding to the node type and the edge weight features corresponding to the edge type, respectively.

[0048] Based on the edge weight features, construct the message passing path in the heterogeneous information graph;

[0049] Based on the message passing path, calculate the importance coefficient between the initial attributes of the nodes;

[0050] Based on the importance coefficient, the initial attributes of the nodes are weighted and fused to obtain preliminary fusion features;

[0051] The preliminary fusion features are subjected to nonlinear transformation to obtain aggregate node features.

[0052] Optionally, calculating the semantic association degree between the enhanced semantic representation vectors includes:

[0053] The enhanced semantic representation vector is subjected to co-normalization to obtain the standard semantic vector;

[0054] Multi-granularity feature extraction is performed on the standard semantic vector to obtain multi-scale semantic features;

[0055] Calculate the interactive attention corresponding to the multi-scale semantic features, and determine the cross-vector attention map of the multi-scale semantic features based on the interactive attention;

[0056] Based on the standard semantic vector and the cross-vector attention map, the semantic correlation between the enhanced semantic representation vectors is calculated.

[0057] To address the aforementioned problems, this invention also provides a multilingual user matching and recommendation system based on multidimensional tag fusion, the system comprising:

[0058] The tag vector construction module is used to collect multidimensional tag data of users in the digital service platform, extract language feature symbols and interaction behavior features from the multidimensional tag data, and construct the multidimensional tag vector of users in the digital service platform.

[0059] The information coverage calculation module is used to input the multi-dimensional label vector into a pre-trained deep learning model, use the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multi-dimensional label vector to obtain an aligned semantic vector, and use the coverage function in the deep learning model to calculate the information coverage corresponding to the aligned semantic vector.

[0060] An enhanced semantic generation module is used to determine the core tag vector in the multi-dimensional tag vector based on the information coverage, and to generate the enhanced semantic representation vector corresponding to the user by combining the core tag vector and the interaction behavior features.

[0061] The recommended content matching module is used to calculate the semantic correlation between the enhanced semantic representation vectors, and based on the semantic correlation and the enhanced semantic representation vectors, to match the recommended content corresponding to the user from the digital service platform.

[0062] Compared to the problems described in the background art, this invention, by extracting linguistic features and interactive behavior features from the multidimensional label data, can accurately capture users' language expression patterns and interactive behavior patterns on digital service platforms, avoiding interference from irrelevant features and improving the accuracy of subsequent multidimensional label vector construction. This invention utilizes a cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multidimensional label vectors, obtaining aligned semantic vectors. This eliminates the distribution differences of the multidimensional label vectors in the semantic space, establishes a standard semantic understanding benchmark, and avoids deviations in subsequent analysis due to language differences. Furthermore, this invention combines the core label vectors and the interactive behavior features to generate enhanced semantic representation vectors corresponding to the user, obtaining a more comprehensive and high-quality representation vector that integrates the user's core semantic preferences and real interactive behaviors. This provides a basis for subsequently matching recommended content corresponding to the user from the digital service platform. By calculating the semantic correlation between the enhanced semantic representation vectors, this invention can understand the strength of the correlation between the enhanced semantic representation vectors, thereby improving the accuracy of subsequent matching of recommended content corresponding to the user. Therefore, the multilingual user matching and recommendation method and system based on multidimensional tag fusion provided in this embodiment of the invention can improve the accuracy of multilingual user matching and recommendation based on multidimensional tag fusion. Attached Figure Description

[0063] Figure 1This is a flowchart illustrating a multilingual user matching and recommendation method based on multidimensional tag fusion, provided as an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the content matching process of a multilingual user matching and recommendation method based on multidimensional tag fusion, provided in an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram of a module for implementing a multilingual user matching and recommendation system based on multidimensional tag fusion, as provided in an embodiment of the present invention.

[0066] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0068] This application provides a multilingual user matching and recommendation method based on multidimensional tag fusion. The executing entity of this multilingual user matching and recommendation method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the multilingual user matching and recommendation method based on multidimensional tag fusion can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0069] Reference Figure 1 The diagram shown is a flowchart illustrating a multilingual user matching and recommendation method based on multidimensional tag fusion according to an embodiment of the present invention. In this embodiment, the multilingual user matching and recommendation method based on multidimensional tag fusion includes:

[0070] S1. Collect multidimensional tag data of users in the digital service platform, and extract language feature symbols and interaction behavior features from the multidimensional tag data to construct the multidimensional tag vector of users in the digital service platform.

[0071] This invention, by extracting linguistic features and interactive behavior features from the multidimensional tag data, can accurately capture users' language expression patterns and interactive behavior patterns on digital service platforms, avoiding interference from irrelevant features and improving the accuracy of subsequent multidimensional tag vector construction. The multidimensional tag data refers to data generated by users through various interaction channels on the digital service platform, such as user-generated content, clickstream data, and browsing history. The linguistic features are semantic units extracted from the linguistic data, such as keywords, sentiment polarity, and topic tags. The interactive behavior features are quantitative indicators characterizing user interaction patterns, such as click frequency, session duration, and page navigation paths. Furthermore, the collection of multidimensional tag data can be achieved through data collection tools, such as log file analysis systems for recording user interaction behavior, natural language processing interfaces for parsing user-generated content, and user behavior tracking SDKs for capturing real-time interaction events. Both the log file analysis system and the natural language processing interface are compiled using programming languages, such as Java.

[0072] As an embodiment of the present invention, the extraction of language feature symbols and interaction behavior features from the multidimensional label data includes:

[0073] The multidimensional label data is cleaned to obtain standardized label data;

[0074] The standardized label data is subjected to data classification processing to obtain language data and behavioral data;

[0075] The language data is parsed to obtain basic semantic units;

[0076] The basic semantic units are subjected to feature encoding processing to obtain language feature symbols;

[0077] The behavioral data is processed by behavioral sequence extraction to obtain interactive behavioral features.

[0078] The standardized tag data refers to consistent and complete data that has been cleaned and formatted; the language data refers to modal data containing user-generated text content, such as comments and query statements; the behavioral data refers to modal data that records user operation sequences, such as clickstreams and browsing history; the basic semantic unit refers to the smallest semantic element obtained after decomposition, such as lexical units and phrase fragments; the language feature is a feature vector with semantic discriminative power obtained after encoding; and the interactive behavior feature is a numerical feature that quantifies user behavior patterns, such as sequence patterns and frequent itemsets.

[0079] Furthermore, the multidimensional label data can be cleaned using Z-score-based anomaly detection combined with regularization methods to identify and remove outliers and handle missing values, resulting in standardized label data. The standardized label data can be modally segmented using rule-based stratified sampling, dynamically dividing the data into linguistic and behavioral data based on data attributes. The linguistic data can be parsed using dependency parsing combined with named entity recognition to extract core lexical units and entity phrases from sentences, yielding basic semantic units. The basic semantic units can be feature-encoded using word frequency-inverse document frequency combined with word embedding models to generate word vectors and aggregate them into document-level features, obtaining linguistic feature symbols. The behavioral data can be processed using hidden Markov models combined with frequent pattern mining algorithms to extract behavioral sequences, identify state transition probabilities and common behavioral paths, extract sequence statistical features, and obtain interactive behavioral features. For example, when analyzing user behavior on e-commerce platforms, behavioral sequence modeling can identify typical patterns such as "browsing-adding to cart-paying," thereby extracting interactive behavioral features representing purchase intent.

[0080] Furthermore, as an optional embodiment of the present invention, the step of performing semantic unit parsing processing on the language data to obtain basic semantic units includes:

[0081] The language data is subjected to grammatical structure analysis to obtain a grammatical dependency tree;

[0082] Semantic role annotation is performed on the grammatical dependency tree to obtain a semantic role framework;

[0083] The semantic role framework is subjected to entity linking processing to obtain a set of linked entities;

[0084] The set of linked entities is subjected to relation extraction processing to obtain a semantic relation graph;

[0085] Cluster analysis is performed on the semantic relationship graph to obtain semantic clusters;

[0086] The semantic clusters are processed by extracting core units to obtain basic semantic units.

[0087] The grammatical dependency tree is a tree-like data structure representing the grammatical dependencies between words; the semantic role framework is a set of labeled predicate-argument structures; the linked entity set is a set of consistent entities after disambiguation; the semantic relationship graph is a graph structure with entities as nodes and relationships as edges; the semantic cluster is a group of semantic units with high cohesion in the graph; and the core unit extraction process is the process of selecting key semantic elements from the community.

[0088] Furthermore, the language data can be processed using the Stanford parser to perform syntactic structure analysis and construct a syntactic dependency tree representing subject-verb-object relationships. The syntactic dependency tree can be processed using the PropBank framework to perform semantic role annotation, annotating the core arguments of the predicates to obtain a semantic role framework. Entity disambiguation algorithms combined with a knowledge base can be used to link entities within the semantic role framework, resolving entity ambiguities and obtaining a set of linked entities, such as a local context-aware entity disambiguation algorithm. Open information extraction techniques can be used to extract relationships from the set of linked entities, extracting semantic relationships between entities and constructing a semantic relationship graph, such as rule-based phrase-level open extraction techniques. The semantic relationship graph can be clustered using the xxLouvain graph clustering algorithm to obtain semantic clusters. The semantic community can be processed using the PageRank algorithm to extract core units, calculating node importance scores and selecting entities or words with high scores as basic semantic units. For example, when processing user comments such as "This phone has a long battery life and excellent photo quality," semantic parsing can yield core units such as "battery life" and "photo quality," providing a foundation for subsequent feature encoding.

[0089] This invention extracts linguistic features and interactive behavior features from the multidimensional tag data to construct a multidimensional tag vector for the user on the digital service platform. This can comprehensively characterize the user's semantic preferences and behavioral patterns, avoiding a one-sided understanding of user features during the user profile construction process, thereby improving the accuracy of personalized services and user satisfaction. The multidimensional tag vector is a hierarchical structure object constructed based on the linguistic features and interactive behavior features.

[0090] As an embodiment of the present invention, the step of extracting language feature symbols and interaction behavior features from the multidimensional tag data to construct the multidimensional tag vector of the user in the digital service platform includes:

[0091] Semantic graphs are constructed from the linguistic features to obtain a semantic relation network;

[0092] The interactive behavior features are analyzed and processed to obtain a behavior pattern map;

[0093] Analyze the structural correspondence between the semantic relationship network and the behavioral pattern graph to obtain cross-modal mapping relationships;

[0094] Based on the cross-modal mapping relationship, the semantic relationship network and the behavioral pattern graph are fused to obtain fused topological features;

[0095] The fused topological features are subjected to feature optimization processing to obtain optimized topological features;

[0096] Based on the optimized topological features, a multidimensional tag vector of the user in the digital service platform is constructed.

[0097] The semantic relation network is a topological model representing the semantic associations between linguistic features through a graph structure, including semantic similarity and association strength between nodes; the behavior pattern graph is a behavior feature evolution graph constructed through a temporal graph network, reflecting the temporal dynamics and spatial distribution of user behavior; the cross-modal mapping relationship is a mapping matrix that establishes the correspondence between semantic network nodes and behavior graph nodes; the fused topological feature is a unified graph structure representation after fusing multimodal features, retaining the topological characteristics of the original features; the optimized topological feature is a stable topological architecture after structural simplification and feature enhancement; and the multidimensional label vector is a hierarchically organized user feature representation constructed based on the optimized topology.

[0098] Furthermore, a semantic graph can be constructed from the linguistic features using a graph attention network, calculating importance weights between nodes using an attention mechanism to generate a semantic relationship network with differentiated connection strengths; behavioral trajectory analysis can be performed on the interactive behavior features using a temporal graph convolutional network to capture the spatiotemporal evolution patterns in the behavior sequence and construct a behavior pattern graph reflecting the dynamic changes in user behavior; the structural correspondence between the semantic relationship network and the behavior pattern graph can be analyzed using a graph neural network alignment algorithm, calculating the cosine similarity of node embedding vectors and establishing a cross-modal mapping relationship matrix; and graph structure fusion technology can be used to deeply integrate the semantic relationship network and the behavior pattern graph based on the cross-modal mapping relationship, using graph convolution operations to achieve feature propagation and... Aggregation yields fused topological features that retain bimodal characteristics. These fused topological features can be optimized using graph sparsification algorithms combined with feature importance assessment to remove redundant connections, strengthen critical paths, and identify core feature clusters through dynamic community discovery algorithms, resulting in optimized topological features. Based on these optimized topological features, multidimensional label vectors with multi-scale feature representations can be constructed using hierarchical graph encoding. For example, in e-commerce platform user profile construction, language features are constructed into a three-level semantic network of "brand preference - product attributes - usage scenarios," and behavioral features are modeled as a behavioral graph of "browsing path - purchase decision - after-sales interaction." A mapping relationship between "brand preference and purchase decision" is established through graph alignment, ultimately generating a multidimensional label vector that fully reflects the user's consumption profile.

[0099] S2. Input the multidimensional label vector into a pre-trained deep learning model, use the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multidimensional label vector to obtain an aligned semantic vector, and use the coverage function in the deep learning model to calculate the information coverage corresponding to the aligned semantic vector.

[0100] This invention utilizes a cross-language fusion layer in a deep learning model to perform semantic alignment processing on the multi-dimensional label vectors, obtaining aligned semantic vectors. This eliminates the distribution differences of the multi-dimensional label vectors in the semantic space, establishes a standard semantic understanding benchmark, and avoids biases in subsequent analysis caused by language differences. The pre-trained deep learning model is based on a large-scale multilingual corpus. This model learns the universal semantic representations of dozens of languages, possessing cross-language semantic understanding and feature extraction capabilities, and provides reliable basic model support for the semantic alignment processing of the cross-language fusion layer, such as the XLM-RoBERTa model. The cross-language fusion layer is a multilingual semantic alignment network based on an attention mechanism, capable of mapping label vectors from different languages ​​to a unified semantic space. The aligned semantic vectors are vectors with consistent semantic representations after semantic alignment processing, such as cross-language unified vectors or semantically standardized vectors.

[0101] As an embodiment of the present invention, the step of using the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multi-dimensional label vector to obtain an aligned semantic vector includes:

[0102] The multidimensional label vector is projected using the feature projection algorithm in the cross-language fusion layer to obtain a projected semantic representation.

[0103] The projected semantic representation is encoded using the attention encoder in the cross-language fusion layer to obtain a semantic feature vector;

[0104] Collect the context data corresponding to the multidimensional label vector, and use the feature-aware network in the cross-language fusion layer to mine the context features in the context data;

[0105] The semantic feature vector and the context features are fused using the semantic fusion network in the cross-language fusion layer to obtain an aligned semantic vector.

[0106] The projected semantic representation is a low-dimensional semantic representation obtained by projecting the multidimensional label vector using a feature projection algorithm in the cross-language fusion layer. The feature projection algorithm includes canonical correlation analysis and t-SNE dimensionality reduction. The semantic feature vector is an expression vector obtained by encoding the projected semantic representation using an attention encoder in the cross-language fusion layer. The attention encoder consists of a multi-head self-attention layer and a positional feedforward network. The context data consists of the source and target language context word vector sequences and syntactic dependency relation data corresponding to the multidimensional label vector. The context features are cross-language context features mined using a feature-aware network in the cross-language fusion layer. The feature-aware network consists of a bidirectional GRU layer. The aligned semantic vector is a feature obtained by gating and fusing the semantic feature vector and the context features using a semantic fusion network in the cross-language fusion layer. The semantic fusion network consists of gated recurrent units and fully connected layers.

[0107] This invention utilizes the coverage function in the deep learning model to calculate the information coverage corresponding to the aligned semantic vector, thereby understanding the information integrity of the aligned semantic vector and providing a quantitative basis for the subsequent determination of the core label vector. The coverage function is a composite function used to calculate the information coverage corresponding to the aligned semantic vector; the information coverage is a quantitative indicator that measures the information richness and distribution uniformity in the aligned semantic vector, such as information entropy value, feature diversity, etc.

[0108] As an embodiment of the present invention, the step of calculating the information coverage corresponding to the aligned semantic vector using the coverage function in the deep learning model includes:

[0109] The variance of the alignment semantic vector is calculated to obtain the vector variance value;

[0110] Calculate the information entropy corresponding to each vector in the alignment semantic vector to obtain the vector information entropy;

[0111] Calculate the sparsity of each vector in the alignment semantic vector to obtain the vector sparsity;

[0112] By combining the vector variance, the vector information entropy, and the vector sparsity, the information coverage corresponding to the aligned semantic vector is calculated using the coverage function.

[0113] Wherein, the vector variance value is a quantitative representation of the degree of dispersion of the numerical values ​​of each dimension of the alignment semantic vector; the vector information entropy is a measure of the uniformity of the information distribution of the alignment semantic vector; and the vector sparsity is a characterization of the density of the element distribution of the alignment semantic vector.

[0114] Furthermore, the variance of the aligned semantic vector can be calculated using an unbiased estimation algorithm to obtain the vector variance value. This method uses n-1 as the denominator, which can more accurately reflect the dispersion of the vector values. The information entropy of each vector in the aligned semantic vector can be calculated using a probability distribution transformation method based on softmax. This method first converts the vector values ​​into a probability distribution and then calculates the entropy value, which can effectively handle negative values ​​and accurately assess the uniformity of information distribution. The sparsity of each vector in the aligned semantic vector can be calculated using the vector norm ratio method. This method calculates the ratio of the absolute value of each dimension of the vector to the Euclidean length, which can stably reflect the density characteristics of the vector.

[0115] Furthermore, as another optional embodiment of the present invention, the step of combining the vector variance value, the vector information entropy, and the vector sparsity to calculate the information coverage corresponding to the aligned semantic vector using the coverage function includes:

[0116] The coverage function is calculated using the following formula:

[0117]

[0118] Where C represents the information coverage corresponding to the alignment semantic vector. Let E represent the vector variance, and E represent the vector information entropy. Let S denote the minimum perturbation constant, and let S denote the vector sparsity.

[0119] In this formula, the molecule part It maintains the product relationship between variance and information entropy, reflecting the synergistic effect of distribution dispersion and information richness. It is a minimal constant used to prevent division by zero errors, and its value is typically taken as [value missing]. , Maintain the original key characteristics and provide smooth decay when sparsity approaches 0.

[0120] S3. Based on the information coverage, determine the core tag vector in the multi-dimensional tag vector, and combine the core tag vector and the interaction behavior features to generate the enhanced semantic representation vector corresponding to the user.

[0121] This invention generates an enhanced semantic representation vector corresponding to the user by combining the core tag vector and the interaction behavior features. This results in a high-quality representation vector that comprehensively integrates the user's core semantic preferences and real interaction behaviors, providing a basis for matching recommended content corresponding to the user from the digital service platform. The core tag vector is a feature vector composed of key tags with high information coverage selected from the multi-dimensional tag vectors, representing the user's core attribute features. The preset graph neural network algorithm is a fusion algorithm based on an attention-enhanced graph network. The enhanced semantic representation vector is a deep semantic representation that integrates core tag features and interaction behavior features, enabling a more comprehensive representation of the user's interests, preferences, and behavioral characteristics. Furthermore, based on the information coverage, the core tag vector in the multi-dimensional tag vector is determined by setting an information coverage threshold to filter vectors higher than the threshold, or by selecting the top N vectors in descending order of information coverage.

[0122] As an embodiment of the present invention, the step of combining the core tag vector and the interaction behavior features to generate the enhanced semantic representation vector corresponding to the user includes:

[0123] Construct a heterogeneous information graph corresponding to the core label vector and the interaction behavior features;

[0124] The heterogeneous information graph is subjected to node feature aggregation processing to obtain aggregated node features;

[0125] Based on the features of the aggregation nodes, an aggregation semantic representation vector corresponding to the user is generated;

[0126] The aggregated semantic representation vector is subjected to semantic enhancement processing to obtain the enhanced semantic representation vector.

[0127] The heterogeneous information graph refers to a graph structure data containing multiple types of nodes and edges, with core label vectors as nodes and interactive behavior features as edges; the aggregated node features are the updated feature representations of nodes in the heterogeneous information graph obtained by fusing their multi-order neighbor node information through the message passing mechanism of a graph neural network; the aggregated semantic representation vector is an initial vector representation specifically for user nodes, generated based on their aggregated node features and incorporating local graph structure information.

[0128] Furthermore, a heterogeneous information graph corresponding to the core label vector and the interaction behavior features can be constructed using a meta-path-based graph construction method. Specifically, the user core label vector is used as the initial feature of the user node, the interaction behavior feature is used as the attribute of the corresponding edge, and item nodes and behavior type nodes are introduced to construct a complete heterogeneous information graph. Through feature mapping and normalization, the aggregated node features are projected onto a semantic space of a unified dimension. After optimizing the feature distribution through the Sigmoid activation function, an aggregated semantic representation vector is generated. The aggregated semantic representation vector can be semantically enhanced through a multi-head self-attention mechanism. This mechanism can capture the semantic dependencies within the vector from different subspaces and stabilize the training process through residual connections and layer normalization, ultimately outputting an enhanced semantic representation vector with rich semantic information and strong discriminative power.

[0129] Furthermore, as an optional embodiment of the present invention, the step of aggregating node features of the heterogeneous information graph to obtain aggregated node features includes:

[0130] Identify the node and edge types in the heterogeneous information graph;

[0131] Extract the initial node attributes corresponding to the node type and the edge weight features corresponding to the edge type, respectively.

[0132] Based on the edge weight features, construct the message passing path in the heterogeneous information graph;

[0133] Based on the message passing path, calculate the importance coefficient between the initial attributes of the nodes;

[0134] Based on the importance coefficient, the initial attributes of the nodes are weighted and fused to obtain preliminary fusion features;

[0135] The preliminary fusion features are subjected to nonlinear transformation to obtain aggregate node features.

[0136] Wherein, the node type is a multilingual tag node; the edge type includes user-tag association edges and product-tag matching edges; the node initial attribute is the inherent attribute parameter corresponding to the node type, which is the original feature expression of the node type before participating in feature aggregation; the edge weight feature is the weight value on the edge, representing the strength of the relationship, such as the frequency of cross-language interaction between users; the message transmission path is the path of information propagation between nodes based on edge weights; the importance coefficient is a coefficient representing the degree of mutual influence between nodes during message transmission; and the preliminary fusion feature is the intermediate representation feature obtained through weighted fusion.

[0137] Furthermore, node and edge types in the heterogeneous information graph can be identified through rule-based type matching mechanisms. For example, node types and equilateral types can be distinguished by parsing the "type" field in node attributes and the "relation" field in edge attributes. Pre-trained graph embedding models can be used to extract the initial node attributes corresponding to each node type and the edge weight features corresponding to each edge type. For example, the Node2Vec model can be used to learn the profile features of user nodes and the content features of product nodes, and the TransR model can be used to calculate the edge weight features. Based on the edge weight features, the heterogeneous information graph can be constructed using a meta-path-guided random walk algorithm. The message passing path in the graph is randomly walked according to the meta-path of "user-purchase-product-label-tag" to generate a semantically coherent path sequence; based on the message passing path, the importance coefficient between nodes is calculated, for example, through a heterogeneous graph attention network; according to the importance coefficient, the initial attributes of the nodes can be weighted and fused by a weighted summation operation to obtain preliminary fusion features; the original attributes of the target node and the features of its associated nodes are multiplied by their corresponding importance coefficients and then summed to obtain the preliminary fusion features of the target node; the preliminary fusion features can be nonlinearly transformed by the ReLU activation function to obtain aggregated node features.

[0138] S4. Calculate the semantic correlation degree between the enhanced semantic representation vectors, and based on the semantic correlation degree and the enhanced semantic representation vectors, match the recommended content corresponding to the user from the digital service platform.

[0139] This invention calculates the semantic correlation degree between the enhanced semantic representation vectors to understand the strength of the association between them, thereby improving the accuracy of subsequent matching of recommended content for the user. The semantic correlation degree is a quantitative indicator that measures the semantic similarity and relevance between the enhanced semantic representation vectors.

[0140] As an embodiment of the present invention, calculating the semantic association degree between the enhanced semantic representation vectors includes:

[0141] The enhanced semantic representation vector is subjected to co-normalization to obtain the standard semantic vector;

[0142] Multi-granularity feature extraction is performed on the standard semantic vector to obtain multi-scale semantic features;

[0143] Calculate the interactive attention corresponding to the multi-scale semantic features, and determine the cross-vector attention map of the multi-scale semantic features based on the interactive attention;

[0144] Based on the standard semantic vector and the cross-vector attention map, the semantic correlation between the enhanced semantic representation vectors is calculated.

[0145] Wherein, the standard semantic vector is a vector representation with unified dimensions and comparability obtained by co-normalizing the enhanced semantic representation vector; the multi-scale semantic feature is a feature set containing multi-level semantic information such as word level, phrase level and sentence level obtained by multi-granularity feature extraction of the standard semantic vector; the interactive attention is the bidirectional attention weight distribution corresponding to the multi-scale semantic feature; and the cross-vector attention graph is the semantic unit association matrix constructed by the multi-scale semantic feature based on the interactive attention.

[0146] Furthermore, the enhanced semantic representation vector can be co-normalized using linear projection combined with L2 normalization to obtain a standard semantic vector. For example, a linear projection network consisting of three fully connected layers can be used to project the 128-dimensional and 256-dimensional enhanced semantic representation vectors to 64 dimensions, followed by L2 normalization to constrain the vector magnitude to 1, thus obtaining a standard semantic vector. The standard semantic vector can then be processed using a convolutional neural network combined with a self-attention mechanism to extract multi-granularity features, resulting in multi-scale semantic features. For instance, a 1×1 convolutional kernel can be used to extract fine-grained features at the word level, a 3×3 convolutional kernel to extract medium-grained features at the phrase level, and a 5×5 convolutional kernel to extract coarse-grained features at the topic level. This is combined with a self-attention mechanism to strengthen the weights of key semantic features, outputting multi-scale semantic features. Finally, a multi-head cross-attention mechanism can be used to calculate the interactive attention corresponding to the multi-scale semantic features, such as setting six independent attention heads. The interaction focuses on different dimensions such as semantic matching, sentiment tendency, and topic association. The attention weights output by each head are concatenated to obtain a comprehensive interactive attention. Based on the interactive attention, a cross-vector attention map of the multi-scale semantic features is reconstructed through a weight matrix. For example, the interactive attention weights are multiplied by a dot product with a randomly initialized and trained feature mapping matrix to reconstruct a cross-vector attention map with dimensions consistent with the multi-scale semantic features, which intuitively reflects the correlation strength between different vector features. Based on the standard semantic vector and the cross-vector attention map, the semantic correlation between the enhanced semantic representation vectors can be calculated using the attention-weighted cosine similarity method. For example, the standard semantic vector and the cross-vector attention map are multiplied element-wise to obtain a weighted vector, and then the cosine similarity between the weighted vectors is calculated. If the result is 0.82, it indicates that the two enhanced semantic representation vectors have a high semantic correlation.

[0147] This invention, by matching recommended content corresponding to a user from a digital service platform based on the semantic relevance and the enhanced semantic representation vector, can effectively improve the accuracy of content recommendation, thereby significantly improving the user experience and satisfaction on the digital service platform. Furthermore, the specific steps for matching recommended content corresponding to a user from the digital service platform based on the semantic relevance and the enhanced semantic representation vector are as follows: Based on the semantic relevance, identify highly relevant content clusters in the user's interest graph, and filter out a set of candidate content that highly matches the user's historical preferences. Simultaneously, based on the enhanced semantic representation vector, calculate the vector space similarity between the platform content and the user profile to obtain the most relevant potential interest content. Combine the corresponding content popularity to optimize the ranking of the candidate content set and the potential interest content, and finally match the recommended content corresponding to the user from the digital service platform. For a more intuitive understanding of the content matching process of the multilingual user matching and recommendation method based on multidimensional tag fusion in this application, please refer to [link to relevant documentation]. Figure 2 The diagram shown is a schematic representation of the content matching process of a multilingual user matching and recommendation method based on multidimensional tag fusion provided by this invention. It should be noted that in this invention, Figure 2 The flowchart presented is only for the content matching process of a multilingual user matching and recommendation method based on multidimensional tag fusion, and is not limited to the content matching process of a multilingual user matching and recommendation method based on multidimensional tag fusion in different actual application scenarios.

[0148] Compared to the problems described in the background art, this invention, by extracting linguistic features and interactive behavior features from the multidimensional label data, can accurately capture users' language expression patterns and interactive behavior patterns on digital service platforms, avoiding interference from irrelevant features and improving the accuracy of subsequent multidimensional label vector construction. This invention utilizes a cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multidimensional label vectors, obtaining aligned semantic vectors. This eliminates the distribution differences of the multidimensional label vectors in the semantic space, establishes a standard semantic understanding benchmark, and avoids deviations in subsequent analysis due to language differences. Furthermore, this invention combines the core label vectors and the interactive behavior features to generate enhanced semantic representation vectors corresponding to the user, obtaining a more comprehensive and high-quality representation vector that integrates the user's core semantic preferences and real interactive behaviors. This provides a basis for subsequently matching recommended content corresponding to the user from the digital service platform. By calculating the semantic correlation between the enhanced semantic representation vectors, this invention can understand the strength of the correlation between the enhanced semantic representation vectors, thereby improving the accuracy of subsequent matching of recommended content corresponding to the user. Therefore, the multilingual user matching and recommendation method and system based on multidimensional tag fusion provided in this embodiment of the invention can improve the accuracy of multilingual user matching and recommendation based on multidimensional tag fusion.

[0149] like Figure 3 The diagram shown is a functional block diagram of a multilingual user matching and recommendation system based on multidimensional tag fusion according to the present invention.

[0150] The multilingual user matching and recommendation system 300 based on multidimensional tag fusion described in this invention can be installed in an electronic device. Depending on the functions implemented, the multilingual user matching and recommendation system based on multidimensional tag fusion may include a tag vector construction module 301, an information coverage calculation module 302, an enhanced semantic generation module 303, and a recommendation content matching module 304. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0151] In this embodiment of the invention, the functions of each module / unit are as follows:

[0152] The tag vector construction module 301 is used to collect multidimensional tag data of users in the digital service platform, extract language feature symbols and interaction behavior features from the multidimensional tag data, and construct the multidimensional tag vector of users in the digital service platform.

[0153] The information coverage calculation module 302 is used to input the multidimensional label vector into a pre-trained deep learning model, use the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multidimensional label vector to obtain an aligned semantic vector, and use the coverage function in the deep learning model to calculate the information coverage corresponding to the aligned semantic vector.

[0154] The enhanced semantic generation module 303 is used to determine the core tag vector in the multi-dimensional tag vector based on the information coverage, and generate the enhanced semantic representation vector corresponding to the user by combining the core tag vector and the interaction behavior features.

[0155] The recommended content matching module 304 is used to calculate the semantic correlation degree between the enhanced semantic representation vectors, and based on the semantic correlation degree and the enhanced semantic representation vectors, to match the recommended content corresponding to the user from the digital service platform.

[0156] In detail, the modules in the multilingual user matching and recommendation system 300 based on multidimensional tag fusion described in this embodiment of the invention employ the same methods as described above. Figure 1 This method employs the same technical means as the multilingual user matching and recommendation method based on multidimensional tag fusion described above, and can produce the same technical effect, so it will not be elaborated here.

[0157] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0158] Finally, it should be noted that in the above embodiments, each embodiment can be combined with each other or independent. Deleting any one of them will not affect the technical implementation of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not 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.

Claims

1. A multilingual user matching and recommendation method based on multidimensional tag fusion, characterized in that, The method includes: Collect multidimensional tag data of users in the digital service platform, extract language features and interaction behavior features from the multidimensional tag data, and construct the multidimensional tag vector of users in the digital service platform. The language features are semantic units extracted from language data, and the interaction behavior features are quantitative indicators that characterize user interaction patterns. The multidimensional label vector is input into a pre-trained deep learning model. The cross-language fusion layer in the deep learning model is used to perform semantic alignment processing on the multidimensional label vector to obtain an aligned semantic vector. The information coverage corresponding to the aligned semantic vector is calculated using the coverage function in the deep learning model. Based on the information coverage, the core tag vector in the multi-dimensional tag vector is determined, and the enhanced semantic representation vector corresponding to the user is generated by combining the core tag vector and the interaction behavior features. Calculate the semantic correlation between the enhanced semantic representation vectors, and based on the semantic correlation and the enhanced semantic representation vectors, match the recommended content corresponding to the user from the digital service platform.

2. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The extraction of linguistic features and interaction behavior features from the multidimensional label data includes: The multidimensional label data is cleaned to obtain standardized label data; The standardized label data is subjected to data classification processing to obtain language data and behavioral data; The language data is parsed to obtain basic semantic units; The basic semantic units are subjected to feature encoding processing to obtain language feature symbols; The behavioral data is processed by behavioral sequence extraction to obtain interactive behavioral features.

3. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 2, characterized in that, The process of parsing the language data to obtain basic semantic units includes: The language data is subjected to grammatical structure analysis to obtain a grammatical dependency tree; Semantic role annotation is performed on the grammatical dependency tree to obtain a semantic role framework; The semantic role framework is subjected to entity linking processing to obtain a set of linked entities; The set of linked entities is subjected to relation extraction processing to obtain a semantic relation graph; Cluster analysis is performed on the semantic relationship graph to obtain semantic clusters; The semantic clusters are processed by extracting core units to obtain basic semantic units.

4. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The step of extracting linguistic features and interaction behavior features from the multidimensional tag data to construct the user's multidimensional tag vector in the digital service platform includes: Semantic graphs are constructed from the linguistic features to obtain a semantic relation network; The interactive behavior features are analyzed and processed to obtain a behavior pattern map; Analyze the structural correspondence between the semantic relationship network and the behavioral pattern graph to obtain cross-modal mapping relationships; Based on the cross-modal mapping relationship, the semantic relationship network and the behavioral pattern graph are fused to obtain fused topological features; The fused topological features are subjected to feature optimization processing to obtain optimized topological features; Based on the optimized topological features, a multidimensional tag vector of the user in the digital service platform is constructed.

5. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The step of using the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multi-dimensional label vector to obtain an aligned semantic vector includes: The multidimensional label vector is projected using the feature projection algorithm in the cross-language fusion layer to obtain a projected semantic representation. The projected semantic representation is encoded using the attention encoder in the cross-language fusion layer to obtain a semantic feature vector; Collect the context data corresponding to the multidimensional label vector, and use the feature-aware network in the cross-language fusion layer to mine the context features in the context data; The semantic feature vector and the context features are fused using the semantic fusion network in the cross-language fusion layer to obtain an aligned semantic vector.

6. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The step of calculating the information coverage corresponding to the aligned semantic vector using the coverage function in the deep learning model includes: The variance of the alignment semantic vector is calculated to obtain the vector variance value; Calculate the information entropy corresponding to each vector in the alignment semantic vector to obtain the vector information entropy; Calculate the sparsity of each vector in the alignment semantic vector to obtain the vector sparsity; By combining the vector variance, the vector information entropy, and the vector sparsity, the information coverage corresponding to the aligned semantic vector is calculated using the coverage function.

7. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The step of combining the core tag vector and the interaction behavior features to generate the enhanced semantic representation vector corresponding to the user includes: Construct a heterogeneous information graph corresponding to the core label vector and the interaction behavior features; The heterogeneous information graph is subjected to node feature aggregation processing to obtain aggregated node features; Based on the features of the aggregation nodes, an aggregation semantic representation vector corresponding to the user is generated; The aggregated semantic representation vector is subjected to semantic enhancement processing to obtain the enhanced semantic representation vector.

8. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 7, characterized in that, The step of aggregating node features in the heterogeneous information graph to obtain aggregated node features includes: Identify the node and edge types in the heterogeneous information graph; Extract the initial node attributes corresponding to the node type and the edge weight features corresponding to the edge type, respectively. Based on the edge weight features, construct the message passing path in the heterogeneous information graph; Based on the message passing path, calculate the importance coefficient between the initial attributes of the nodes; Based on the importance coefficient, the initial attributes of the nodes are weighted and fused to obtain preliminary fusion features; The preliminary fusion features are subjected to nonlinear transformation to obtain aggregate node features.

9. The multilingual user matching and recommendation method based on multidimensional tag fusion as described in claim 1, characterized in that, The calculation of the semantic association degree between the enhanced semantic representation vectors includes: The enhanced semantic representation vector is subjected to co-normalization to obtain the standard semantic vector; Multi-granularity feature extraction is performed on the standard semantic vector to obtain multi-scale semantic features; Calculate the interactive attention corresponding to the multi-scale semantic features, and determine the cross-vector attention map of the multi-scale semantic features based on the interactive attention; Based on the standard semantic vector and the cross-vector attention map, the semantic correlation between the enhanced semantic representation vectors is calculated.

10. A multilingual user matching and recommendation system based on multidimensional tag fusion, characterized in that, The system includes: The tag vector construction module is used to collect multidimensional tag data of users in the digital service platform, extract language features and interaction behavior features from the multidimensional tag data, and construct the multidimensional tag vector of the user in the digital service platform. The language features are semantic units extracted from language data, and the interaction behavior features are quantitative indicators that characterize the user's interaction pattern. The information coverage calculation module is used to input the multi-dimensional label vector into a pre-trained deep learning model, use the cross-language fusion layer in the deep learning model to perform semantic alignment processing on the multi-dimensional label vector to obtain an aligned semantic vector, and use the coverage function in the deep learning model to calculate the information coverage corresponding to the aligned semantic vector. An enhanced semantic generation module is used to determine the core tag vector in the multi-dimensional tag vector based on the information coverage, and to generate the enhanced semantic representation vector corresponding to the user by combining the core tag vector and the interaction behavior features. The recommended content matching module is used to calculate the semantic correlation between the enhanced semantic representation vectors, and based on the semantic correlation and the enhanced semantic representation vectors, to match the recommended content corresponding to the user from the digital service platform.

Images