An AI technology-based short message template classification and automatic matching method and system

Abstract

The application discloses a kind of short message template classification and automatic matching method and system based on AI technology, it is related to information and communication technology field, introduce pre-training language model to short message request and template text are separated type semantic modeling, respectively extract request side vector group and template side vector group;Again based on request side vector, construct semantic neighborhood graph, and with template vector as reference, dynamic correction is carried out, and request-template joint semantic graph is formed;Node feature propagation is carried out on the graph structure using graph convolutional neural network, and the migration probability matrix of request to template category and template category feedback information are generated, and then the semantic representation of request and template is optimized, combined with category response intensity and semantic deviation situation, the system generates category weight adjustment table, dynamically corrects template category representation, matches the optimized vector through joint matching module Calculation, determine the optimal template and complete parameter filling and message sending.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technology, and in particular to a method and system for classifying and automatically matching short message templates based on AI technology. Background Art

[0002] Currently, in business SMS sending scenarios, predefined template matching is usually used to bind content and select templates for user SMS requests. Even when deep learning or vectorized matching methods are introduced, one-way static matching is still the main method. This method lacks the ability to model the semantic interaction structure between requests and templates, and is difficult to handle problems such as request semantic ambiguity, overlap between template classes, or drastic changes within classes.

[0003] For example, Chinese patent publication number "CN120106036A" discloses a method for replacing SMS variables based on dynamic template editing. This method extracts template text through code parsing and replaces variable tags in the template with actual content based on stored variable mappings. The replaced complete SMS content is then transmitted to the SMS service module for delivery to the target user. However, the following issues remain:

[0004] Shallow semantic models cannot accurately capture the true intent of SMS requests. The template category center cannot dynamically respond to changes in the semantic distribution of actual requests, resulting in distorted semantic clustering and large matching deviations.

[0005] Requests and templates are usually regarded as two independent vectors, with no structural cascade and no semantic interaction, making it difficult to adapt to complex matching relationships; there are also deficiencies in semantic feedback and dynamic adjustment, making it impossible to continuously optimize matching accuracy during operation.

[0006] In view of this, the present invention provides an AI-based SMS template classification and automatic matching method and system to solve the above problems. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for SMS template classification and automatic matching based on AI technology, which solves the problems of weak semantic understanding ability, lack of structural interaction between requests and templates, static and non-adaptive template categories, and lack of closed-loop optimization mechanism in the matching process in existing SMS template matching methods.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] In a first aspect, the present invention provides a method for classifying and automatically matching SMS templates based on AI technology, comprising the following steps:

[0010] S101: Extracting a request-side vector group and a template-side vector group based on a pre-trained language model, wherein the request-side vector group is used to represent the request vector and user intent of SMS call request data, and the template-side vector group is used to represent the semantic vector set of all SMS templates corresponding to the original text data of the service SMS template;

[0011] S102: Build a request semantic neighborhood graph based on the request-side vector group, and use the template-side vector group as a reference to perform a first dynamic correction on the request semantic neighborhood graph based on node distance distribution to build a request-template joint semantic graph;

[0012] S103: Utilize the request-template joint semantic graph and the template-side vector group to jointly drive the graph convolutional neural network to perform node feature propagation, generate a category migration matrix and template category feature feedback, and optimize the request-side vector group and the template-side vector group based on the template category feature feedback information;

[0013] S104: generating a category weight adjustment table according to the category migration matrix and the optimized template side vector group, and dynamically adjusting the template category representation of the template side vector group;

[0014] S105: Input the optimized request-side vector group and the template-side vector group adjusted by category weights into a joint matching module to generate a target SMS template and complete message sending.

[0015] As a preferred technical solution of the first aspect of the present invention, the construction logic of the pre-trained language model is:

[0016] The input layer performs word segmentation on the original text data to generate a token sequence. The embedding layer converts the token sequence into a fixed-dimensional vector sequence. The vector sequence contains word vectors, position vectors, and segment vectors, which are used to identify word content, position structure, and semantic partition information.

[0017] Encoding layer: Use the Transformer structure to establish matching confidence analysis of long-distance dependencies and semantic alignment relationships between texts, and extract the matching confidence between any two vector sequences.

[0018] If the match confidence is lower than or equal to the match confidence threshold, the candidate fallback mechanism is triggered. In the candidate fallback stage, the vector sequence is classified by field function to extract the category label of the user intent. The vector sequence with the category label of the user intent is returned to the input layer for re-segmentation.

[0019] The vector sequence with the category label of the user intention is semantically expanded to obtain the expanded original text data, and the expanded original text data is returned to the input layer for re-segmentation.

[0020] The expanded original text data is introduced into a multimodal fusion matching method of rules, statistics and neural network models to obtain fused original text data, which is then returned to the input layer for re-segmentation.

[0021] Until the matching confidence is higher than the matching confidence threshold, the vector sequence and matching confidence are sent to the output layer;

[0022] The output layer, based on the type of original text data, outputs a vector sequence including a request-side semantic vector group and a template-side semantic vector group to complete the target template selection.

[0023] As a preferred technical solution of the first aspect of the present invention, the logic for obtaining the token sequence is as follows:

[0024] In the word segmentation stage, the original text data is identified by predefined field rules or named entity recognition models, and replaced with unified placeholder tags.

[0025] Retrain the subword vocabulary based on high-frequency words in the SMS field, update the category labels of user intentions in the subword vocabulary, the semantic expansion method in the original text data, and the multimodal fusion matching method in the original text data.

[0026] As a preferred technical solution of the first aspect of the present invention, the construction logic of the request semantic neighborhood graph is:

[0027] The request-side vector group includes the request semantic vector of a single request corresponding to the request node. The request semantic vector includes three granularities: sentence-level request, keyword request, and action-target request.

[0028] Based on the semantic similarity threshold, functional relevance or context transfer relationship between the request semantic vectors, the adjacent edges between the request nodes are constructed to form a preliminary semantic neighborhood graph structure;

[0029] Edge type labels are set for different request semantic vector relationships, including semantic rewriting edges, functional similarity edges, contextual transition edges, and antonym edges, to form a heterogeneous semantic graph structure.

[0030] Perform graph clustering, introduce skip connections, and adjust edge weights on the preliminary semantic neighborhood graph structure and heterogeneous semantic graph structure.

[0031] The K-nearest neighbor algorithm is used to establish edges between all request semantic vectors. , find the K most semantically similar request semantic vectors, connect them into adjacent edges, and build the initial semantic neighborhood graph. The length of each edge represents the semantic similarity between request nodes.

[0032] As a preferred technical solution of the first aspect of the present invention, the dynamic correction logic for the first dynamic correction of the request semantic neighborhood graph is:

[0033] Obtaining a preset template-side vector group, and calculating the semantic distance of the adjacent nodes of each request node in the request semantic neighborhood graph relative to the template-side vector group;

[0034] Based on the mean and standard deviation of the semantic distance between the node corresponding to the adjacent edge of the current request node and the template side vector group, a dynamic correction threshold is set to determine whether the adjacent edge meets the semantic rationality constraint;

[0035] An edge culling operation is performed on the adjacent edges that do not meet the dynamic correction threshold constraint. After culling, if the number of current node neighbors is lower than the set K value, new adjacent nodes with reasonable semantics are re-screened based on the template semantic distance to replace the culled edges, completing the first dynamic correction of the requested semantic neighborhood graph.

[0036] As a preferred technical solution of the first aspect of the present invention, the construction logic of the request-template joint semantic graph is:

[0037] For the dynamically modified neighborhood graph structure, a heterogeneous joint semantic graph of requests and templates is constructed; where:

[0038] Node structure: includes request node and template node, and labels node type;

[0039] Edge type: The edge between requesting nodes is a semantic proximity edge; the edge between requesting nodes and template nodes is a semantic matching edge;

[0040] Connection strategy: Semantic matching edges with matching confidence higher than a preset confidence threshold are connected to achieve cross-side graph connection.

[0041] As a preferred technical solution of the first aspect of the present invention, the optimization logic of the request-side vector group and the template-side vector group is:

[0042] Using the request-template joint semantic graph as graph structure input and the template side vector group as initial node feature input, driving the graph convolutional neural network to perform at least two rounds of node feature propagation;

[0043] Generate the category migration matrix based on the similarity calculation between the node features after the propagation is completed and the preset category vector, which is used to represent the mapping probability distribution of the request node to the template category space;

[0044] Based on the category transfer matrix and the category to which the template node belongs, the feature representation of each category template node after graph convolution propagation is extracted, and the corresponding category semantic center vector is generated and transmitted to the requesting node as a feedback signal;

[0045] The requesting node compares the similarity between its own propagated semantic vector and the semantic centers of each category, and selects the most relevant category center as the target feedback vector, which represents the deviation degree and gradient direction of the template node in the category representation space.

[0046] The target feedback vector in the category migration matrix is ​​utilized, and a back-propagation mechanism is adopted to jointly optimize the node features of the request-side vector group and the template-side vector group.

[0047] As a preferred technical solution of the first aspect of the present invention, the logic for dynamically adjusting the template category representation of the template side vector group is:

[0048] Using the generated category migration matrix, the semantic response strength of all request nodes to each template category is counted;

[0049] For each template category, calculate the offset of its semantic center vector before and after propagation optimization;

[0050] The semantic response strength and offset metrics are fused into a category weight adjustment coefficient according to the formula to measure whether the representation weight of each category in the semantic space needs to be strengthened or weakened.

[0051] The category weight adjustment coefficients are organized into a structured mapping to form the final category weight adjustment table: with the template category as the index, each category corresponds to an adjustment weight , used to drive the update of its category representation in the semantic space;

[0052] The template category representation of the template side vector group is dynamically adjusted using the category weight adjustment coefficient of each module category in the category weight adjustment table.

[0053] As a preferred technical solution of the first aspect of the present invention, the target SMS template and the logic for completing message sending are:

[0054] Input the optimized request-side vector group and template-side vector group;

[0055] Use vector similarity or fusion feature similarity algorithms to match each request vector with all template vectors one by one to obtain a similarity score. Select the template with the highest score as the target SMS template.

[0056] Based on the target SMS template, the parameter information carried in the request is used to fill the placeholders in the template to generate a message. The generated SMS is sent to the user through the connected SMS platform or message push service.

[0057] In a second aspect, the application provides an SMS template classification and automatic matching system based on AI technology, which is based on the implementation of the first aspect and comprises a vector extraction module, a semantic graph construction module, a graph propagation and optimization module, a category weight adjustment module, and a matching and sending module, wherein data transmission is performed between the various modules through wired and / or wireless transmission.

[0058] The vector extraction module extracts a request-side vector group and a template-side vector group based on a pre-trained language model, wherein the request-side vector group is used to represent the request vector and user intent of the SMS call request data, and the template-side vector group is used to represent the semantic vector set corresponding to all SMS templates of the original text data of the business SMS template.

[0059] The semantic graph construction module constructs a request semantic neighborhood graph based on the request-side vector group, and performs a first dynamic correction on the request semantic neighborhood graph according to the node distance distribution and with reference to the template-side vector group, thereby constructing a request-template joint semantic graph.

[0060] The graph propagation and optimization module drives the node feature propagation of the graph convolutional neural network using the request-template joint semantic graph and the template-side vector group, generates a category migration matrix and a template category feature feedback, and optimizes the request-side vector group and the template-side vector group according to the template category feature feedback information.

[0061] The category weight adjustment module generates a category weight adjustment table according to the category migration matrix and the optimized template-side vector group, and dynamically adjusts the template category representation of the template-side vector group.

[0062] The matching and sending module inputs the optimized request-side vector group and the template-side vector group adjusted by the category weight into a joint matching module, generates a target SMS template, and completes message sending.

[0063] In the above technical solution, the application provides the following technical effects and advantages:

[0064] The application realizes efficient, accurate, and self-adaptive matching and pushing between SMS requests and templates by constructing a request-template semantic joint graph, introducing a graph neural network, and introducing a category feedback mechanism. The various modules cooperate with each other, the logic chain is clear, and the application is easy to implement and can be flexibly extended to multi-language and multi-business field scenarios.

[0065] Through separate semantic modeling and joint optimization of the request side and the template side, the user intent and key field semantics in the request can be accurately captured, and the semantic alignment effect between the request and the template can be significantly improved. Compared with traditional rule matching or one-way models, the matching accuracy is higher and the fault tolerance is stronger.

[0066] By constructing a request-template joint semantic graph and introducing template vectors to dynamically modify the request semantic graph structure, we achieve interactive and collaborative modeling of request structure and template categories in the semantic space, breaking through the limitations of existing methods that are based only on unilateral static mapping.

[0067] For the first time, template category feedback is introduced into the request vector optimization process. The category migration matrix and gradient feedback mechanism are used to make the request semantic representation adjustable and self-learning, significantly enhancing the system's generalization ability in multiple SMS template scenarios.

[0068] By utilizing the category response distribution and template semantic offset, a category weight adjustment table is constructed, and the template category center vector is dynamically adjusted to make the template representation closer to the actual request semantic clustering, solving the problems of rigid template semantic expression and large intra-class heterogeneity. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0070] Figure 1 This is a flow chart of the SMS template classification and automatic matching method of the present invention.

[0071] Figure 2 This is a framework diagram of the SMS template classification and automatic matching system of the present invention. DETAILED DESCRIPTION

[0072] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be more comprehensive and complete, and will fully convey the concepts of the example embodiments to those skilled in the art. The accompanying drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures indicate identical or similar parts, and thus repeated descriptions thereof will be omitted.

[0073] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more example embodiments. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure can be practiced while omitting one or more of the specific details, or methods, components, steps, etc. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0074] Embodiment 1

[0075] As Figure 1 shown, the present application provides a short message template classification and automatic matching method based on AI technology, comprising the following steps:

[0076] S101: Extracting request side vector group and template side vector group based on pre-trained language model, the request side vector group is used to represent the request vector and user intention of short message calling request data, and the template side vector group is used to represent the semantic vector set corresponding to all short message templates of business short message template original text data;

[0077] It should be noted that: based on the pre-trained language model, the business short message template original text data and the short message calling request data are separately modeled; two types of semantic embedding vectors are generated in parallel, which are respectively the request side vector group and the template side vector group independently represented, laying a foundation for subsequent double-channel semantic comparison and optimization, ensuring clear data structure, separated input path, and being conducive to forming subsequent bidirectional semantic interaction.

[0078] Short message calling request data refers to the original text or request content in the real-time short message sending request initiated from the business system or third-party interface; short message request generated in actual business scenarios such as system background, business API interface, and user operation. Exemplarily, request R1: "System detects that your account is at risk, please pay attention to verification.", request R2: "Bill has been generated, please pay in time.", and request R3: "Thank you for using this service, your account has been successfully activated."

[0079] Business short message template original text data refers to standardized short message template text pre-set by enterprises or platforms and used for business push. It comes from the template management module of enterprise short message platform, is usually configured and generated by business personnel or product managers, and is stored in template database. Exemplarily, template T1: "You have a new bill to pay, please handle it in time.", template T2: "Dear customer, your account is abnormal, please contact customer service.", and template T3: "Your verification code is: {code}, please do not disclose."

[0080] Specifically, the construction logic of the pre-trained language model is:

[0081] The input layer performs word segmentation on the original text data to generate Token sequence; the Token sequence is converted into a fixed-dimensional vector sequence through the embedding layer, and the vector sequence includes word vector, position vector and segmentation vector, which are used to identify word content, position structure and semantic partition information;

[0082] Encoding layer: the matching confidence between the long-distance dependence and semantic alignment relationship between the text is established by the Transformer structure, and the matching confidence between any two vector sequences is extracted,

[0083] If the matching confidence is lower than or equal to the matching threshold, the candidate fallback mechanism is triggered, and in the candidate fallback stage, the field function classification recognition is performed on the vector sequence to extract the category label of the user intent, and the vector sequence with the category label of the user intent is returned to the input layer for rewording processing.

[0084] The vector sequence with the category label of the user intent is obtained by the semantic expansion method, and the expanded original text data is returned to the input layer for rewording processing.

[0085] The expanded original text data is introduced into the multi-modal fusion matching mode of rule, statistics and neural network model, the fused original text data is obtained, and the fused original text data is returned to the input layer for rewording processing.

[0086] Until the matching confidence is higher than the matching threshold, the vector sequence and the matching confidence are sent to the output layer.

[0087] The output layer, based on the type of the original text data, outputs the vector sequence including the request side semantic vector group and the template side semantic vector group, and completes the target template selection.

[0088] Further, there are a large number of dynamic variable fields (such as verification code, user name, amount, etc.) in the SMS call request, which causes the problem of semantic deviation, and the acquisition logic of the Token sequence is:

[0089] In the word segmentation stage, the field recognition is performed on the original text data by the pre-defined field rule or named entity recognition model, and the uniform placeholder mark is used for replacement, such as processing "verification code is 123456" into "verification code is <code>This method can significantly improve the semantic alignment effect between the request-side text and the template text, and reduce the noise interference of actual variable values ​​on semantic modeling. In addition, the present invention retrains the subword vocabulary based on high-frequency words in the SMS field, updates the category label of the user's intention in the subword vocabulary, the semantic expansion method in the original text data, and the multimodal fusion matching method in the original text data, ensuring that the domain vocabulary participates in the encoding as a whole word, and further improving the word segmentation quality and matching effect.

[0090] S102: Build a request semantic neighborhood graph based on the request-side vector group, and use the template-side vector group as a reference to perform a first dynamic correction on the request semantic neighborhood graph based on node distance distribution to build a request-template joint semantic graph;

[0091] It should be noted that the local semantic relationship between requests is expressed based on the request semantic neighborhood graph. The semantic adjacency between requests is not only determined by the request text itself, but also the constraint effect of the clustering structure of the existing template in the semantic space on the request side vector group should be considered. Therefore, the template side vector group is used as a reference to dynamically intervene in and optimize the graph structure on the request side.

[0092] The graph structure integrates node information from requests and templates, breaking the limitations of traditional one-way graph construction. Unlike existing one-way modeling methods, the request-template joint semantic graph breaks the boundaries of the data side to form a heterogeneous semantic graph with a dynamically variable structure. This enables dynamic coordination between the request-side graph structure and template features.

[0093] Specifically, the construction logic of the request semantic neighborhood graph is:

[0094] The request side vector group is denoted as , is a positive integer; each It is a request semantic vector that represents a single request (such as a user query, intent, input phrase, etc.) corresponding to a request node in the neighborhood graph. The request semantic vector includes three granularities: sentence-level request, keyword request, and action-goal request.

[0095] Based on the semantic similarity threshold, functional relevance or context transfer relationship between the request semantic vectors, the adjacent edges between the request nodes are constructed to form a preliminary semantic neighborhood graph structure;

[0096] Edge type labels are set for different request semantic vector relationships, including semantic rewriting edges, functional similarity edges, contextual transition edges, and antonym edges, to form a heterogeneous semantic graph structure.

[0097] Perform graph clustering, introduce skip connections, and adjust edge weights on the preliminary semantic neighborhood graph structure;

[0098] The K-nearest neighbor algorithm is used to establish edges between all request semantic vectors. , find the K most semantically similar request semantic vectors, connect them into adjacent edges, and build the initial semantic neighborhood graph. The length of each edge represents the semantic similarity between request nodes.

[0099] In other words, this is the first time that a template-embedded-driven dynamic modification mechanism for the request semantic neighborhood graph and a request-template joint semantic graph construction method have been introduced. This allows template-side vectors to participate in the structural adjustment of the request-side graph, breaking the technical limitations of existing technologies that only build graphs on the request side or only build static graphs based on single-side data. This allows for the dynamic expression of deep semantic relationships between requests and templates, significantly improving the semantic adaptation capability between requests and templates.

[0100] More specifically, the dynamic correction logic for the first dynamic correction of the requested semantic neighborhood graph is:

[0101] Obtaining a preset template-side vector group, and calculating the semantic distance of the adjacent nodes of each request node in the request semantic neighborhood graph relative to the template-side vector group;

[0102] Based on the mean and standard deviation of the semantic distance between the node corresponding to the adjacent edge of the current request node and the template side vector group, a dynamic correction threshold is set to determine whether the adjacent edge meets the semantic rationality constraint;

[0103] An edge culling operation is performed on the adjacent edges that do not meet the dynamic correction threshold constraint. After culling, if the number of current node neighbors is lower than the set K value, new adjacent nodes with reasonable semantics are re-screened based on the template semantic distance to replace the culled edges, completing the first dynamic correction of the requested semantic neighborhood graph.

[0104] Furthermore, the construction logic of the request-template joint semantic graph is as follows:

[0105] For the dynamically modified neighborhood graph structure, a heterogeneous joint semantic graph of requests and templates is constructed; where:

[0106] Node structure: includes request node and template node, and labels node type;

[0107] Edge type: The edge between requesting nodes is a semantic proximity edge; the edge between requesting nodes and template nodes is a semantic matching edge;

[0108] Connection strategy: Semantic matching edges with matching confidence higher than the preset confidence threshold are connected to achieve cross-side graph connection; breaking the traditional limitation of building graphs only on the request side or template side, dynamic interaction and collaborative expression between the request side graph structure and the template structure are achieved, which is conducive to enhancing semantic adaptation capabilities and providing structural support for subsequent matching optimization.

[0109] S103: driving the graph convolutional neural network to propagate node features using the request-template joint semantic graph and the template side vector group, generating a category transfer matrix and template category feature feedback, and optimizing the request side vector group and the template side vector group according to the template category feature feedback information;

[0110] It should be noted that: based on the request-template joint semantic graph and the template side vector group, the graph convolutional neural network is driven to propagate; the output is a reverse optimization double-sided semantic embedding, which synchronously outputs: a category transfer matrix (representing a classification mapping) and a template category feature feedback (category gradient information). Realize the recursive optimization closed loop of request and template embedding, break the single propagation paradigm of traditional static graph convolution, form a category feedback closed loop; the template category feature feedback is input reversely, which optimizes the embedding weight of the template side and the embedding vector of the request side in real time. Support dynamic embedding evolution, improve the quality of matching representation.

[0111] Specifically, the step S103 includes the following steps:

[0112] The request-template joint semantic graph is input as a graph structure, and the template side vector group is input as a template node feature, and the graph convolutional neural network is driven to propagate the request node features for at least two rounds;

[0113] According to the category labels of each template node in the template side vector group, the semantic similarity between each request node and each category template node in the propagation process is counted, so as to construct a category transfer matrix to generate a category transfer matrix. The category transfer matrix is used to depict the semantic response strength between each request node and each category template node after graph convolution propagation;

[0114] More specifically, the category transfer matrix M is a two-dimensional weight matrix,

[0115] ;

[0116] Wherein: represents the request node i; represents the template category j; represents the probability that the request node is mapped or belongs to the template category; is the similarity calculation result between the vector representation of the request node after semantic propagation and the category semantic center vector corresponding to the template category, which is converted into a probability form after normalization processing, and is used to represent the distribution mapping relationship of the request semantic in the category space.

[0117] For example, if a request node has a high degree of similarity with a "payment reminder" category template node after propagation, the value corresponding to the "payment reminder" column in the category transition matrix will be high. The category transition matrix reflects the distribution trend of requests in the template category semantic space, providing a basis for request category determination and semantic modification.

[0118] Based on the category transfer matrix and the category to which the template node belongs, the feature representation of each category template node after graph convolution propagation is extracted, and the corresponding category semantic center vector is generated and transmitted to the requesting node as a feedback signal;

[0119] The requesting node compares the similarity between its own propagated semantic vector and the semantic centers of each category, and selects the most relevant category center as the target feedback vector;

[0120] Based on the target feedback vector, the semantic representations of the request-side vector group and the template-side vector group are optimized. Optimization methods include, but are not limited to, linear weighted adjustment, feedforward network mapping, or end-to-end training updates based on loss functions, enabling the updated semantic representations to more accurately represent their semantic categorical attributes. This breaks the limitation of traditional semantic modeling, where template information serves only as a static reference, and establishes a closed-loop semantic bidirectional optimization mechanism, whereby the template guides the request and, in turn, optimizes the template through request feedback, thereby improving the accuracy, stability, and domain adaptability of the overall semantic matching system.

[0121] By introducing template category nodes as supervisory anchors within the graph structure, and leveraging graph convolutional neural networks to generate a category transfer matrix and obtain feedback on template category features, this approach achieves a closed-loop, dynamic update of semantic representations. Compared to traditional static graph neural networks that only support a one-time propagation approach, this approach forms an embedding optimization loop through category feedback, improving the generalization and semantic adaptability of the semantic matching model to diverse request intents.

[0122] S104: generating a category weight adjustment table according to the category migration matrix and the optimized template side vector group, and dynamically adjusting the template category representation of the template side vector group;

[0123] It should be noted that during the matching and output stages, a dynamic optimization mechanism for category weights was proposed. By generating a category weight adjustment table based on the category migration matrix, the granularity and weight distribution of the category space on the template side are dynamically adjusted. The category weight adjustment table is then incorporated into the matching candidate screening and output node filtering process, achieving multiple rounds of matching iterations and dynamic optimization. This solves the problem of fixed matching rules and single output control in existing technologies, and effectively improves the accuracy and flexibility of template matching in different business scenarios. The template side and the request side are viewed as dual input channels for dynamic collaboration, and a nonlinear technical framework with multi-path cross-dependency and dynamic feedback optimization is adopted to achieve continuous evolution and automatic adaptation of the template library. This breaks through the limitations of static classifiers and rule matching dependence, and meets the needs of intelligent classification and automatic matching of SMS templates under complex application conditions such as multi-business, multi-scenario, and template cold start.

[0124] The logic for dynamically adjusting the template category representation of the template side vector group is:

[0125] Using the generated category migration matrix, the semantic response strength of all request nodes to each template category is counted;

[0126] Specifically, for each template category , all request nodes Accumulate the probability values ​​of the category in the migration matrix to obtain its semantic response strength, which represents the global request response strength of the template category in the current propagation state:

[0127] ;

[0128] in: represents the probability that the request node is mapped or belongs to the template category; is the total number of requesting nodes; is the semantic response strength of the template category in the current propagation round.

[0129] For each template category, the offset of its semantic center vector before and after propagation optimization is calculated.

[0130] Specifically, we obtain the original semantic center vector of the template category and the new semantic center vector obtained after propagation optimization, and calculate the distance between the two as the offset metric:

[0131] ;

[0132] in: Calculated by embedding the initial template; It is obtained by taking the average representation of all template nodes belonging to the template category after graph convolution propagation; Represents the offset metric of the template category in the semantic space.

[0133] The semantic response strength and offset metrics are fused into a category weight adjustment coefficient according to the formula to measure whether the representation weight of each category in the semantic space needs to be strengthened or weakened.

[0134] The weighted normalization strategy is used to construct the following adjustment factor:

[0135] ;

[0136] in: is the adjustable fusion weight; is the total number of template categories; The larger the value, the more active or unstable the template is in the current propagation state, and the more dynamic adjustment is needed.

[0137] The category weight adjustment coefficients are organized into a structured mapping to form the final category weight adjustment table: with the template category as the index, each category corresponds to an adjustment weight , used to drive the update of its category representation in the semantic space;

[0138] ;

[0139] Using the category weight adjustment coefficients of each module category in the category weight adjustment table, the template category representation of the template side vector group is dynamically adjusted to enhance its clustering expression ability. The update can be performed in the following form:

[0140] ;

[0141] in: is the category offset direction vector; is the step size parameter; is the final adjusted category center vector.

[0142] It should be noted that: through the introduction of the category weight adjustment table, the system can dynamically correct the position of the template category center according to the actual semantic propagation status, so that it is more in line with the semantic distribution of the request node, and improve the clarity of the category boundary and the matching discriminability. It is especially suitable for business scenarios where the distribution of request samples changes dynamically or the boundaries between multiple categories are blurred.

[0143] S105: Input the optimized request-side vector group and the template-side vector group adjusted by category weights into a joint matching module to generate a target SMS template and complete message sending.

[0144] Specifically, the target SMS template and the logic for sending the message are as follows:

[0145] Input the optimized request-side vector group and template-side vector group;

[0146] It's important to note that the request-side vector group has already undergone graph convolutional propagation, category feedback adjustment, and semantic embedding optimization, enabling it to more accurately represent the request intent. The template-side vector group has also been dynamically modified using the "category transfer" and "category weight adjustment" mechanisms, bringing it closer to the true semantic center. These semantically optimized vectors are input to ensure semantic consistency and discriminability during the final match.

[0147] Use vector similarity or fusion feature similarity algorithm to match each request vector with all template vectors one by one to obtain a similarity score, and select the template with the highest score as the target SMS template.

[0148] For example, for a request like "Thank you for signing up for this service. Your account has been activated," the highest-scoring template might be "Dear user, your account has been successfully activated." A higher score indicates closer semantics. Ultimately, the system selects the template with the highest score as the "successful matching target template."

[0149] Based on the target SMS template, the parameter information carried in the request (such as name, verification code, amount, time, etc.) is filled in the placeholders in the template to generate an SMS text with complete structure, accurate semantics, and compliance with business regulations.

[0150] For example, if the template is: "Hello, your verification code is: {CODE}, please do not disclose it." The request parameter is CODE=123456, and the generated SMS message is: "Hello, your verification code is: 123456, please do not disclose it."

[0151] Finally, the system sends the generated SMS to the user through the connected SMS platform or message push service, completing the entire closed-loop process.

[0152] Example 2

[0153] like Figure 2 As shown, the parts not described in detail in this embodiment are as shown in Example 1. This embodiment provides that the present invention provides an AI-based SMS template classification and automatic matching system, including a vector extraction module, a semantic graph construction module, a graph propagation and optimization module, a category weight adjustment module, and a matching and sending module. Data is transmitted between each module via wired and / or wireless communication.

[0154] A vector extraction module extracts a request-side vector group and a template-side vector group based on a pre-trained language model. The request-side vector group is used to represent the request vector and user intent of the SMS call request data, and the template-side vector group is used to represent the semantic vector set of all SMS templates corresponding to the original text data of the service SMS template;

[0155] The semantic graph construction module constructs a request semantic neighborhood graph based on the request side vector group, and performs a first dynamic correction on the request semantic neighborhood graph according to a node distance distribution, with the template side vector group as a reference, to construct a request-template joint semantic graph.

[0156] The graph propagation and optimization module drives a node feature propagation of a graph convolutional neural network by using the request-template joint semantic graph and the template side vector group together to generate a category migration matrix and a template category feature feedback, and optimizes the request side vector group and the template side vector group according to the template category feature feedback information.

[0157] The category weight adjustment module generates a category weight adjustment table according to the category migration matrix and the optimized template side vector group, and dynamically adjusts the template category representation of the template side vector group.

[0158] The matching and sending module inputs the optimized request side vector group and the template side vector group adjusted by the category weight into a joint matching module to generate a target short message template and complete message sending.

[0159] The construction logic of the pre-trained language model is as follows:

[0160] The input layer generates a Token sequence by performing word segmentation on original text data; and the Token sequence is converted into a fixed-dimension vector sequence by an embedding layer, the vector sequence including a word vector, a position vector and a segmentation vector, which are used to identify word content, position structure and semantic partition information.

[0161] The encoding layer: the matching confidence between texts is analyzed by establishing a long-distance dependency and semantic alignment relationship of the texts in a Transformer structure, and the matching confidence between any two vector sequences is extracted,

[0162] If the matching confidence is lower than or equal to a matching confidence threshold, a candidate rollback mechanism is triggered, and in the candidate rollback phase, the vector sequence is classified and identified by field function to extract a category label of the user intent, and the vector sequence with the category label of the user intent is rolled back to the input layer for re-word segmentation processing.

[0163] The vector sequence with the category label of the user intent is obtained by a semantic expansion method to obtain expanded original text data, and the expanded original text data is rolled back to the input layer for re-word segmentation processing.

[0164] The expanded original text data is introduced into a multi-modal fusion matching mode of rule, statistics and neural network model to obtain fused original text data, and the fused original text data is rolled back to the input layer for re-word segmentation processing.

[0165] Until the matching confidence is higher than the matching confidence threshold, the vector sequence and the matching confidence are sent to the output layer.

[0166] The output layer, based on the type of original text data, outputs a vector sequence including a request-side semantic vector group and a template-side semantic vector group to complete the target template selection.

[0167] The logic for obtaining the Token sequence:

[0168] In the word segmentation stage, the original text data is identified by predefined field rules or named entity recognition models, and replaced with unified placeholder tags.

[0169] Retrain the subword vocabulary based on high-frequency words in the SMS field, update the category labels of user intentions in the subword vocabulary, the semantic expansion method in the original text data, and the multimodal fusion matching method in the original text data.

[0170] The construction logic of the request semantic neighborhood graph is:

[0171] The request-side vector group includes the request semantic vector of a single request corresponding to the request node. The request semantic vector includes three granularities: sentence-level request, keyword request, and action-target request.

[0172] Based on the semantic similarity threshold, functional relevance or context transfer relationship between the request semantic vectors, the adjacent edges between the request nodes are constructed to form a preliminary semantic neighborhood graph structure;

[0173] Edge type labels are set for different request semantic vector relationships, including semantic rewriting edges, functional similarity edges, contextual transition edges, and antonym edges, to form a heterogeneous semantic graph structure.

[0174] Perform graph clustering, introduce skip connections, and adjust edge weights on the preliminary semantic neighborhood graph structure and heterogeneous semantic graph structure.

[0175] The K-nearest neighbor algorithm is used to establish edges between all request semantic vectors. , find the K most semantically similar request semantic vectors, connect them into adjacent edges, and build the initial semantic neighborhood graph. The length of each edge represents the semantic similarity between request nodes.

[0176] The dynamic correction logic for the first dynamic correction of the requested semantic neighborhood graph is:

[0177] Obtaining a preset template-side vector group, and calculating the semantic distance of the adjacent nodes of each request node in the request semantic neighborhood graph relative to the template-side vector group;

[0178] Based on the mean and standard deviation of the semantic distance between the node corresponding to the adjacent edge of the current request node and the template side vector group, a dynamic correction threshold is set to determine whether the adjacent edge meets the semantic rationality constraint;

[0179] Performing edge pruning operation on the adjacent edges that do not meet the dynamic correction threshold constraint, and if the number of adjacent nodes of the current node is lower than the set K value after pruning, re-screening new adjacent nodes with reasonable semantics based on the template semantic distance to replace the pruned edges, and completing the first dynamic correction of the request semantic neighborhood graph.

[0180] The construction logic of the request-template combined semantic graph is:

[0181] For the dynamically corrected neighborhood graph structure, a heterogeneous combined semantic graph of requests and templates is constructed, wherein:

[0182] Node structure: contains request nodes and template nodes, and marks node types;

[0183] Edge type: the edge between the request nodes is a semantic adjacent edge; the edge between the request node and the template node is a semantic matching edge;

[0184] Connection strategy: constructing a connection edge for the semantic matching edge with a matching confidence higher than a preset confidence threshold, and realizing cross-side graph connection.

[0185] The optimization logic of the request-side vector group and the template-side vector group is:

[0186] Taking the request-template combined semantic graph as the graph structure input and the template-side vector group as the initial node feature input, driving the graph convolutional neural network to perform at least two rounds of node feature propagation;

[0187] According to the similarity calculation of the node features after propagation and the preset category vector, the category migration matrix is generated, which is used to represent the mapping probability distribution of the request node to the template category space;

[0188] Based on the category migration matrix, the feature representation of each category template node after graph convolution propagation is extracted based on the category to which the template node belongs, and the corresponding category semantic center vector is generated as a feedback signal to the request node;

[0189] The request node compares the similarity of its own propagated semantic vector and each category semantic center, and selects the most relevant category center as the target feedback vector, which represents the deviation degree and gradient direction of the template node in the category representation space;

[0190] Using the target feedback vector in the category migration matrix, the node features of the request-side vector group and the template-side vector group are jointly optimized by using the backpropagation mechanism.

[0191] The logic for dynamically adjusting the template category representation of the template-side vector group is:

[0192] Using the generated category migration matrix, the semantic response strength of all request nodes to each template category is counted;

[0193] For each template category, calculate the offset of its semantic center vector before and after propagation optimization;

[0194] The semantic response strength and offset metrics are fused into a category weight adjustment coefficient according to the formula to measure whether the representation weight of each category in the semantic space needs to be strengthened or weakened.

[0195] The category weight adjustment coefficients are organized into a structured mapping to form the final category weight adjustment table: with the template category as the index, each category corresponds to an adjustment weight , used to drive the update of its category representation in the semantic space;

[0196] The template category representation of the template side vector group is dynamically adjusted using the category weight adjustment coefficient of each module category in the category weight adjustment table.

[0197] The target SMS template and the logic for completing message sending are as follows:

[0198] Input the optimized request-side vector group and template-side vector group;

[0199] Use vector similarity or fusion feature similarity algorithms to match each request vector with all template vectors one by one to obtain a similarity score. Select the template with the highest score as the target SMS template.

[0200] Based on the target SMS template, the parameter information carried in the request is used to fill the placeholders in the template to generate a message. The generated SMS is sent to the user through the connected SMS platform or message push service.

[0201] Example 3

[0202] This embodiment further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in embodiment 1 are implemented.

[0203] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, or magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0204] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0205] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0206] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only one type. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0207] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0208] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0209] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.< / code>

Claims

1. A method for classifying and automatically matching SMS templates based on AI technology, characterized in that: The following steps are involved: S101: Extracting a request-side vector group and a template-side vector group based on a pre-trained language model, wherein the request-side vector group is used to represent the request vector and user intent of SMS call request data, and the template-side vector group is used to represent the semantic vector set of all SMS templates corresponding to the original text data of the service SMS template; S102: Build a request semantic neighborhood graph based on the request-side vector group, and use the template-side vector group as a reference to perform a first dynamic correction on the request semantic neighborhood graph based on node distance distribution to build a request-template joint semantic graph; The construction logic of the request semantic neighborhood graph is: The request-side vector group includes the request semantic vector of a single request corresponding to the request node. The request semantic vector includes three granularities: sentence-level request, keyword request, and action-target request. Based on the semantic similarity threshold, functional relevance or context transfer relationship between the request semantic vectors, the adjacent edges between the request nodes are constructed to form a preliminary semantic neighborhood graph structure; Edge type labels are set for different request semantic vector relationships, including semantic rewriting edges, functional similarity edges, contextual transition edges, and antonym edges, to form a heterogeneous semantic graph structure. Perform graph clustering, introduce skip connections, and adjust edge weights on the preliminary semantic neighborhood graph structure and heterogeneous semantic graph structure. The K-nearest neighbor algorithm is used to establish edges between all request semantic vectors. , find the K most semantically similar request semantic vectors, connect them into adjacent edges, and build the initial semantic neighborhood graph. The length of each edge represents the semantic similarity between request nodes; The dynamic correction logic for the first dynamic correction of the requested semantic neighborhood graph is: Obtaining a preset template-side vector group, and calculating the semantic distance of the adjacent nodes of each request node in the request semantic neighborhood graph relative to the template-side vector group; Based on the mean and standard deviation of the semantic distance between the node corresponding to the adjacent edge of the current request node and the template side vector group, a dynamic correction threshold is set to determine whether the adjacent edge meets the semantic rationality constraint; Perform edge culling on adjacent edges that do not meet the dynamic correction threshold constraint. After culling, if the number of current node neighbors is lower than the set K value, re-screen semantically reasonable new adjacent nodes based on the template semantic distance to replace the culled edges, completing the first dynamic correction of the requested semantic neighborhood graph. The construction logic of the request-template joint semantic graph is as follows: For the dynamically modified neighborhood graph structure, a heterogeneous joint semantic graph of requests and templates is constructed; where: Node structure: includes request node and template node, and labels node type; Edge type: The edge between requesting nodes is a semantic proximity edge; the edge between requesting nodes and template nodes is a semantic matching edge; Connection strategy: Semantic matching edges with matching confidence higher than a preset confidence threshold are connected to achieve cross-side graph connection; S103: Utilize the request-template joint semantic graph and the template-side vector group to jointly drive the graph convolutional neural network to perform node feature propagation, generate a category migration matrix and template category feature feedback, and optimize the request-side vector group and the template-side vector group based on the template category feature feedback information; S104: generating a category weight adjustment table according to the category migration matrix and the optimized template side vector group, and dynamically adjusting the template category representation of the template side vector group; S105: Input the optimized request-side vector group and the template-side vector group adjusted by category weights into a joint matching module to generate a target SMS template and complete message sending.

2. The method for classifying and automatically matching SMS templates based on AI technology according to claim 1, characterized in that: The construction logic of the pre-trained language model is: The input layer performs word segmentation on the original text data to generate a token sequence. The embedding layer converts the token sequence into a fixed-dimensional vector sequence. The vector sequence contains word vectors, position vectors, and segment vectors, which are used to identify word content, position structure, and semantic partition information. Encoding layer: Use the Transformer structure to establish matching confidence analysis of long-distance dependencies and semantic alignment relationships between texts, and extract the matching confidence between any two vector sequences. If the match confidence is lower than or equal to the match confidence threshold, the candidate fallback mechanism is triggered. In the candidate fallback stage, the vector sequence is classified by field function to extract the category label of the user intent. The vector sequence with the category label of the user intent is returned to the input layer for re-segmentation. The vector sequence with the category label of the user intention is semantically expanded to obtain the expanded original text data, and the expanded original text data is returned to the input layer for re-segmentation. The expanded original text data is introduced into a multimodal fusion matching method of rules, statistics and neural network models to obtain fused original text data, which is then returned to the input layer for re-segmentation. Until the matching confidence is higher than the matching confidence threshold, the vector sequence and matching confidence are sent to the output layer; The output layer, based on the type of original text data, outputs a vector sequence including a request-side semantic vector group and a template-side semantic vector group to complete the target template selection.

3. The method for classifying and automatically matching SMS templates based on AI technology according to claim 2, characterized in that: The logic for obtaining the Token sequence: In the word segmentation stage, the original text data is identified by predefined field rules or named entity recognition models, and replaced with unified placeholder tags. Retrain the subword vocabulary based on high-frequency words in the SMS field, update the category labels of user intentions in the subword vocabulary, the semantic expansion method in the original text data, and the multimodal fusion matching method in the original text data.

4. The method for classifying and automatically matching SMS templates based on AI technology according to claim 1, characterized in that: The optimization logic of the request side vector group and the template side vector group is: Using the request-template joint semantic graph as graph structure input and the template side vector group as initial node feature input, driving the graph convolutional neural network to perform at least two rounds of node feature propagation; Generate the category migration matrix based on the similarity calculation between the node features after the propagation is completed and the preset category vector, which is used to represent the mapping probability distribution of the request node to the template category space; Based on the category transfer matrix and the category to which the template node belongs, the feature representation of each category template node after graph convolution propagation is extracted, and the corresponding category semantic center vector is generated and transmitted to the requesting node as a feedback signal; The requesting node compares the similarity between its own propagated semantic vector and the semantic centers of each category, and selects the most relevant category center as the target feedback vector, which represents the deviation degree and gradient direction of the template node in the category representation space. The target feedback vector in the category migration matrix is ​​utilized, and a back-propagation mechanism is adopted to jointly optimize the node features of the request-side vector group and the template-side vector group.

5. The method for classifying and automatically matching SMS templates based on AI technology according to claim 1, characterized in that: The logic for dynamically adjusting the template category representation of the template side vector group is: Using the generated category migration matrix, the semantic response strength of all request nodes to each template category is counted; For each template category, calculate the offset of its semantic center vector before and after propagation optimization; The semantic response strength and offset metrics are fused into a category weight adjustment coefficient according to the formula to measure whether the representation weight of each category in the semantic space needs to be strengthened or weakened. The category weight adjustment coefficients are organized into a structured mapping to form the final category weight adjustment table: with the template category as the index, each category corresponds to an adjustment weight , used to drive the update of its category representation in the semantic space; The template category representation of the template side vector group is dynamically adjusted using the category weight adjustment coefficient of each module category in the category weight adjustment table.

6. The method for classifying and automatically matching SMS templates based on AI technology according to claim 1, characterized in that: The target SMS template and the logic for completing message sending are as follows: Input the optimized request-side vector group and template-side vector group; Use vector similarity or fusion feature similarity algorithms to match each request vector with all template vectors one by one to obtain a similarity score. Select the template with the highest score as the target SMS template. Based on the target SMS template, the parameter information carried in the request is used to fill the placeholders in the template to generate a message. The generated SMS is sent to the user through the connected SMS platform or message push service.

7. A SMS template classification and automatic matching system based on AI technology, based on the implementation of the SMS template classification and automatic matching method based on AI technology according to any one of claims 1 to 6, characterized in that: It includes a vector extraction module, a semantic graph construction module, a graph propagation and optimization module, a category weight adjustment module, and a matching and sending module. Data is transmitted between each module via wired and / or wireless communication. A vector extraction module extracts a request-side vector group and a template-side vector group based on a pre-trained language model. The request-side vector group is used to represent the request vector and user intent of the SMS call request data, and the template-side vector group is used to represent the semantic vector set of all SMS templates corresponding to the original text data of the service SMS template; The semantic graph construction module builds a request semantic neighborhood graph based on the request-side vector group, and uses the template-side vector group as a reference to perform the first dynamic correction on the request semantic neighborhood graph based on the node distance distribution to construct a request-template joint semantic graph. The graph propagation and optimization module uses the request-template joint semantic graph and the template-side vector group to drive the graph convolutional neural network to propagate node features, generate a category migration matrix and template category feature feedback, and optimize the request-side vector group and template-side vector group based on the template category feature feedback information; The category weight adjustment module generates a category weight adjustment table based on the category migration matrix and the optimized template side vector group, and dynamically adjusts the template category representation of the template side vector group; The matching and sending module inputs the optimized request-side vector group and the template-side vector group adjusted by category weights into the joint matching module to generate the target SMS template and complete the message sending.

Images