Intelligent advertisement sentence generation and accurate delivery system fused with natural language processing

Abstract

This invention discloses an intelligent advertising slogan generation and precise targeting system integrating natural language processing, comprising the following steps: collecting user search keywords, browsing content, and advertising interaction data; processing this data to form user behavior semantic data; extracting user interest topics and their distribution based on topic modeling methods; constructing topic constraint functions based on user interest topics; generating a candidate set of advertising slogans under these constraints; semantically representing the advertising slogans; mapping user interest topics and advertising slogan semantics to a unified semantic space; calculating the semantic correlation between the two; constructing a topic coupling kernel matrix based on the semantic correlation; and using a determinant-based point process to filter the candidate set of advertising slogans to obtain the optimal set of advertising slogans; and matching the optimal set of advertising slogans to the target user group based on the distribution of user interest topics, thereby achieving precise targeting of advertising slogans. This invention relates to the fields of natural language processing and text semantic analysis technology.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent advertising slogan generation and precise delivery system that integrates natural language processing. Background Technology

[0002] With the development of internet advertising and intelligent recommendation technology, the automatic generation and precise targeting of advertising slogans based on user behavior data has become an important application scenario for digital marketing. Current systems typically extract simple features from user click, browsing and search data, and generate advertising slogans by combining rule templates or general text generation models. At the same time, they rely on basic similarity calculation or click-through rate prediction models to complete advertising content matching and delivery.

[0003] Existing technologies still have significant limitations in the generation and delivery of advertising copy. On the one hand, traditional methods are mostly based on keyword matching or shallow semantic modeling, making it difficult to extract deep thematic semantic structures from user behavior text. This leads to semantic discrepancies between the generated advertising copy and the user's actual needs, affecting content relevance. On the other hand, existing advertising copy selection methods typically employ single similarity ranking or simple deduplication strategies, lacking the ability to model the overall diversity of the advertising copy set. This easily results in semantic repetition or redundant expression, reducing the attractiveness of the advertising content. Furthermore, existing advertising matching methods usually process user interest representations and advertising copy representations separately, lacking a fine-grained matching mechanism within a unified semantic space. This makes it difficult to achieve high-precision semantic alignment between user needs and advertising content, thus affecting the accuracy and effectiveness of advertising delivery.

[0004] Therefore, how to provide an intelligent advertising slogan generation and precise delivery system that integrates natural language processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent advertising slogan generation and precise delivery system that integrates natural language processing. This invention constructs a collaborative processing flow from advertising slogan generation to precise delivery through user behavior semantic modeling, topic semantic modeling, topic constraint generation, advertising slogan generation, unified semantic space mapping, and a filtering mechanism based on row and column point processes. It has the advantages of high semantic matching accuracy, strong advertising slogan expression diversity, and strong delivery targeting.

[0006] The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to an embodiment of the present invention includes the following steps: The user behavior corpus construction module collects user search keyword data, user browsing page text data, and user historical advertising interaction text data, and performs text cleaning, word segmentation, and semantic embedding encoding to generate user behavior semantic vector data. The topic semantic modeling module inputs the user behavior semantic vector data into the BERTopic model and outputs user interest topic vectors and corresponding topic probability distributions. The topic constraint generation module constructs user interest topic constraint functions based on topic probability distribution. The ad slogan generation module generates a candidate set of ad slogans under the constraints of the user interest topic constraint function, and performs semantic encoding on the candidate set of ad slogans to obtain the semantic vector of the ad slogans; The unified semantic space mapping module maps user interest topic vectors and advertising slogan semantic vectors to the same semantic space, and calculates the semantic relevance matrix between advertising slogans and user interest topics based on the mapping results; The topic coupling kernel matrix construction module constructs a topic coupling kernel matrix based on the similarity relationship between the semantic relevance matrix and the semantic vector of the advertising slogan; The ad slogan selection and delivery module constructs a matrix-based process model using the theme coupling kernel matrix as input. It performs subset selection on the candidate ad slogan set to obtain the optimal ad slogan set, and matches the optimal ad slogan set to the target user group based on the theme probability distribution, outputting accurate ad slogan delivery results.

[0007] Optionally, the user behavior corpus construction module includes: Collect user search keyword data, which includes the sequence of query terms entered by the user, query time information, and query frequency information; Collect user browsing page text data, which includes webpage title text, body text text, and page tag information corresponding to keyword sequences; Collect user's historical advertising interaction text data, which includes the advertising title text, advertising description text, and interaction behavior markers triggered by the user during browsing. User search keyword data, user page browsing text data, and user historical advertising interaction text data are processed in a unified structure and then concatenated in chronological order to generate the original user behavior text set. The keyword sequence set, browsing text sequence and advertising interaction text set are uniformly converted into structured records containing user identifier, timestamp, text content and behavior type fields, and sorted and aligned according to timestamp to form time-series text data in a unified format. The original user behavior text set is subjected to text cleaning processing, which includes removing stop words, invalid symbols, HTML tags and unified character encoding, to generate a standardized text set; Perform word segmentation on the standardized text set, dividing the standardized text into word sequence and constructing a word list; Semantic embedding encoding is performed on the word sequence based on the word lexicon, mapping the word sequence to a semantic vector representation, and the word vector is aggregated based on text granularity to generate user behavior semantic vector data corresponding to user behavior.

[0008] Optionally, the topic semantic modeling module includes: User behavior semantic vector data is used as text corpus. The text corpus is then semantically encoded based on a pre-trained language model to generate corresponding high-dimensional text semantic embedding representations and to construct a set of text semantic embeddings. The topic information in historical corpora is extracted based on the set of text semantic embeddings, a set of topic prior vectors is constructed, and the similarity between each text semantic embedding and each topic prior vector is calculated to obtain the text topic association score. Based on the text topic association score, the text semantic embedding is reconstructed directionally to generate a topic-constrained semantic embedding representation; The semantic embedding representation guided by the topic constraint is used as input to perform nonlinear dimensionality reduction processing to obtain a set of low-dimensional semantic representations; The semantic similarity between texts is calculated based on a low-dimensional semantic representation set, a semantic similarity matrix is ​​generated, and a text topic relevance matrix is ​​constructed based on the text topic association score. A topic coupling kernel matrix is ​​constructed based on the semantic similarity matrix and the text topic relevance matrix. The topic coupling kernel matrix is ​​then used as clustering input to perform density clustering based on the kernel matrix on the low-dimensional semantic representation set, generating a set of topic clusters. A class-level text set is constructed based on the topic cluster set. The texts within each topic cluster are merged to generate topic texts. The term weights are calculated based on the topic texts to obtain the topic keyword representation. Topic vectors are generated based on topic keywords, and the probability distribution of text under each topic is calculated based on the similarity between topic vectors and text semantic embeddings, thus obtaining user interest topic vectors and corresponding topic probability distributions.

[0009] Optionally, the topic constraint generation module includes: Obtain the topic probability distribution and construct the user topic distribution vector according to the user identifier; Based on the user topic distribution vector, the probabilities of each topic are sorted to obtain a topic sequence arranged in descending order of weight, and topics that meet the preset threshold conditions are selected from the topic sequence to form an effective topic set. The topic probabilities in the effective topic set are normalized to generate normalized topic weights corresponding to each topic vector, and the normalized topic weights are bound to the corresponding topic vectors to form a topic weight mapping set. A user interest topic constraint function is constructed based on the topic weight mapping set. The normalized topic weight is used as a modulation coefficient to act on the corresponding topic vector, forming a multi-topic weighted constraint structure. The multi-topic weighted constraint structure is applied to the input semantic vector. The semantic correlation between the input semantic vector and each topic vector is calculated, and the semantic correlation is weighted and integrated with the corresponding normalized topic weights to obtain the constraint result of the input semantic vector under the user interest topic constraint. Based on the constraint results, the semantic expression of the input semantic vector is constrained and adjusted so that the input semantic vector converges in the semantic space towards the direction of high-weight topics, generating a semantic constraint representation that conforms to the distribution characteristics of user topics, and the semantic constraint representation is used as the output of the user interest topic constraint function.

[0010] Optionally, the advertising slogan generation module includes: Obtain the user interest topic constraint function, and input the user interest topic constraint function as the generation constraint condition into the advertising slogan generation model; Based on the user interest topic constraint function, the conditional generation process is performed. During the generation process, the user interest topic constraint function is applied to the semantic representation of the generation model, so that the semantic expression in the generation process is modulated and constrained by the topic weight, and an advertising text sequence that satisfies the distribution of user interest topics is generated. The advertising text sequence is generated multiple times. Under the condition that the user interest topic constraint function remains unchanged, multiple advertising texts are generated by adjusting the decoding path of the generation model to obtain an initial advertising candidate set. The initial set of candidate advertising slogans is input into the user interest topic constraint function one by one. The semantic correlation between the semantic representation of each advertising slogan text and the topic vector is calculated, and the normalized topic weights are combined for weighted integration to obtain the topic constraint results of each advertising slogan text. Based on the topic constraint results, the initial candidate set of advertising slogans is filtered, and advertising slogan texts that satisfy the user interest topic distribution constraints are retained to generate a constrained candidate set of advertising slogans. The constrained candidate set of advertising slogans is input into a pre-trained language model for semantic encoding. Each advertising slogan text is then subjected to contextual semantic encoding to generate an advertising slogan semantic vector corresponding to the advertising slogan text.

[0011] Optionally, the unified semantic space mapping module includes: Obtain user interest topic vectors and construct a topic vector set; simultaneously obtain advertising slogan semantic vectors and construct an advertising slogan vector set. The topic vector set and the advertising slogan vector set are taken as input and input into the semantic embedding layer of the same pre-trained language model. A unified mapping process is performed on the topic vector and the advertising slogan semantic vector to generate the mapped topic representation vector and advertising slogan representation vector. A semantic space is constructed based on the mapped topic representation vector and the advertising slogan representation vector. The semantic space is a vector space defined by the semantic embedding layer. The topic representation vector and the advertising slogan representation vector have the same vector dimension and a unified semantic measurement method in the semantic space. Based on the topic representation vector and the advertising slogan representation vector in the semantic space, the semantic similarity between each advertising slogan representation vector and each topic representation vector is calculated to obtain the semantic relevance between the advertising slogan and each topic. The semantic relevance results are arranged according to the correspondence between the advertising slogan representation vector and the topic representation vector to construct a semantic relevance matrix. The rows of the semantic relevance matrix correspond to the advertising slogan representation vector, and the columns correspond to the topic representation vector. Each element in the semantic relevance matrix represents the semantic relevance between the corresponding advertising slogan and the topic.

[0012] Optionally, the topic coupling kernel matrix construction module includes: Obtain the semantic relevance matrix, extract the topic relevance vector corresponding to each advertising slogan from the semantic relevance matrix, and obtain the set of semantic vectors for each advertising slogan; Based on the set of semantic vectors of advertising slogans, calculate the semantic similarity between any two semantic vectors of advertising slogans, and generate a set of semantic similarity relationships of advertising slogans; Associate and bind the topic relevance vector with the corresponding advertising slogan semantic vector, so that each advertising slogan semantic vector corresponds to a unique topic relevance vector; Based on the semantic similarity relationship set of advertising slogans and the topic relevance vector, the semantic similarity between any two advertising slogans is modulated so that the semantic similarity is adjusted according to the difference in the topic relevance vectors of the corresponding two advertising slogans, thereby generating advertising slogan similarity relationship after topic constraint; Based on the similarity relationship of advertising slogans constrained by the theme, all advertising slogans are combined in pairs and arranged to construct a theme coupling kernel matrix. Each element in the theme coupling kernel matrix is ​​determined by the modulated similarity relationship between the corresponding advertising slogans.

[0013] Optionally, the advertising slogan screening and delivery module includes: Obtain the topic coupling kernel matrix and the set of advertising slogan candidates corresponding to each index position in the topic coupling kernel matrix; Based on the index correspondence between the topic coupling kernel matrix and the advertising slogan candidate set, a deterministic point process model is constructed. The topic coupling kernel matrix is ​​used as the kernel function matrix, and a mapping relationship is established between any subset of the advertising slogan candidate set and the corresponding kernel submatrix in the topic coupling kernel matrix. A candidate subset set is generated based on the candidate set of advertising slogans, and the kernel submatrix corresponding to each candidate subset is extracted from the topic coupling kernel matrix according to the mapping relationship; The determinant of the nucleus matrix is ​​calculated to obtain the determinant value of the corresponding candidate subset, and the determinant value is used as the selection metric for the candidate subset; Based on the selection metric, a subset selection process is performed. Candidate advertising slogans are selected from the candidate set of advertising slogans and added to the current subset. After each addition, the kernel matrix is ​​updated according to the newly added advertising slogans until the preset subset size condition is met, and the optimal set of advertising slogans is generated. Obtain the probability distribution of user interest topics, and construct a user topic distribution vector based on the topic probability distribution; The advertising slogans in the optimal set are associated with the topic relevance vectors corresponding to the semantic relevance matrix, and the topic relevance vectors of each advertising slogan are matched and calculated based on the user topic distribution vector to obtain the matching relationship between the advertising slogans and the target user group. Based on the matching relationship, the advertising slogans in the optimal set are assigned to the corresponding target user groups, and the advertising slogans are accurately delivered.

[0014] The beneficial effects of this invention are: This invention achieves a deep semantic association between user needs and advertising content by integrating user behavior semantic modeling, topic-based semantic analysis methods, and a unified semantic space mapping mechanism. This enables the system to extract structured topic information from user behavior data and generate advertising content that conforms to the distribution of user interests, thereby improving the matching degree between advertising content generation and user needs.

[0015] This invention modulates the advertising slogan generation process by introducing a topic constraint mechanism, so that the generated advertising slogan is guided by the user's interest topic in the semantic expression process. At the same time, it combines semantic relevance calculation under a unified semantic space to achieve consistent expression between advertising slogan and user interests, and reduce the deviation between advertising slogan and user needs.

[0016] This invention constructs a topic coupling kernel matrix and combines it with a determinant point process to screen the candidate set of advertising slogans. This allows the advertising slogan selection process to consider both semantic similarity and topic distribution differences. In this way, while ensuring that the advertising slogans are relevant to user needs, semantic repetition between advertising slogans is avoided, and the differentiated expression of the advertising slogan set is achieved.

[0017] This invention achieves the corresponding allocation between advertising slogans and target user groups by matching the optimal set of advertising slogans with the distribution of user interest topics. This enables the advertising slogan delivery process to be executed based on semantic matching results, thereby forming a collaborative processing flow from advertising slogan generation and screening to delivery, improving the targeting and consistency of advertising slogan delivery. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent advertising slogan generation and precise delivery system integrating natural language processing proposed in this invention; Figure 2 This is a schematic diagram illustrating the process of constructing user behavior corpus and modeling topic semantics in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 The intelligent advertising slogan generation and precise targeting system, which integrates natural language processing, includes the following steps: The user behavior corpus construction module collects user search keyword data, user browsing page text data, and user historical advertising interaction text data, and performs text cleaning, word segmentation, and semantic embedding encoding to generate user behavior semantic vector data. The topic semantic modeling module inputs the user behavior semantic vector data into the BERTopic model and outputs user interest topic vectors and corresponding topic probability distributions. The topic constraint generation module constructs user interest topic constraint functions based on topic probability distribution. The ad slogan generation module generates a candidate set of ad slogans under the constraints of the user interest topic constraint function, and performs semantic encoding on the candidate set of ad slogans to obtain the semantic vector of the ad slogans; The unified semantic space mapping module maps user interest topic vectors and advertising slogan semantic vectors to the same semantic space, and calculates the semantic relevance matrix between advertising slogans and user interest topics based on the mapping results; The topic coupling kernel matrix construction module constructs a topic coupling kernel matrix based on the similarity relationship between the semantic relevance matrix and the semantic vector of the advertising slogan; The ad slogan selection and delivery module constructs a matrix-based process model using the theme coupling kernel matrix as input. It performs subset selection on the candidate ad slogan set to obtain the optimal ad slogan set, and matches the optimal ad slogan set to the target user group based on the theme probability distribution, outputting accurate ad slogan delivery results.

[0021] In this embodiment, the user behavior corpus construction module includes: Collect user search keyword data, which includes the sequence of query terms entered by the user, query time information, and query frequency information; Collect user browsing page text data, which includes webpage title text, body text text, and page tag information corresponding to keyword sequences; Collect user's historical advertising interaction text data, which includes the advertising title text, advertising description text, and interaction behavior markers triggered by the user during browsing. User search keyword data, user page browsing text data, and user historical advertising interaction text data are processed in a unified structure and then concatenated in chronological order to generate the original user behavior text set. The keyword sequence set, browsing text sequence and advertising interaction text set are uniformly converted into structured records containing user identifier, timestamp, text content and behavior type fields, and sorted and aligned according to timestamp to form time-series text data in a unified format. The original user behavior text set is subjected to text cleaning processing, which includes removing stop words, invalid symbols, HTML tags and unified character encoding, to generate a standardized text set; Stop word removal involves deleting non-semantic functional words from the text based on a preset stop word list; invalid symbol removal involves deleting punctuation marks, special characters, and non-text characters; HTML tag removal involves parsing and removing tag content from the webpage source code, retaining only readable text; and unified character encoding involves converting text in different encoding formats into a unified encoding standard to ensure text consistency. Perform word segmentation on the standardized text set, dividing the standardized text into word sequence and constructing a word list; Semantic embedding encoding is performed on the word sequence based on the word vocabulary, mapping the word sequence to a semantic vector representation, and the word vector is aggregated based on text granularity to generate user behavior semantic vector data corresponding to user behavior. Based on the pre-trained language model, each word in the word list is mapped to a corresponding semantic vector. Then, word sequence vector representations are generated according to the contextual relationship of words in the sequence. Finally, word vectors are sequentially combined and weighted on a text-by-text basis to obtain the overall semantic vector representation of the corresponding text.

[0022] In this embodiment, the topic semantic modeling module includes: User behavior semantic vector data is used as text corpus. The text corpus is then semantically encoded based on a pre-trained language model to generate corresponding high-dimensional text semantic embedding representations and to construct a set of text semantic embeddings. The text corpus is input into a pre-trained language model. The text is converted into a sequence of words and mapped to an initial vector by a word segmenter. The multi-layer self-attention structure in the model is used to encode the contextual relationship between words and output a text embedding representation containing contextual semantic information. The topic information in historical corpora is extracted based on the set of text semantic embeddings, a set of topic prior vectors is constructed, and the similarity between each text semantic embedding and each topic prior vector is calculated to obtain the text topic association score. The vector similarity calculation method is adopted to match each text semantic embedding vector with each topic prior vector one by one. By calculating the cosine similarity or inner product value between the vectors, the similarity score between the corresponding text and each topic is obtained. Based on the text topic association score, the text semantic embedding is directionally reconstructed to generate a topic-constrained semantic embedding representation, wherein the reconstruction process is represented as follows: ; in, Indicates the first A semantic embedding vector after text reconstruction Indicates the first The original semantic embedding vector of each text. Indicates the first Each topic prior vector Indicates the first The first text semantic embedding and the first Similarity scores between topic prior vectors This represents the semantic reconstruction intensity adjustment coefficient. This represents the total number of topic prior vectors. This represents the operation of the exponential function; The semantic embedding reconstruction formula is derived from the Softmax normalization function in probability theory and the linear combination and residual update method in vector space. It constructs a weight distribution through the similarity score between the text semantic embedding vector and the topic prior vector, and uses the Softmax function to exponentially normalize the similarity score, forming weight coefficients that satisfy probability distribution constraints. Vector difference is then introduced. This represents the offset from the original semantic embedding to the topic semantic direction. The overall semantic transfer direction is obtained by weighted summation. This transfer direction is scaled by an adjustment coefficient and superimposed on the original semantic embedding vector to form the reconstructed semantic embedding representation. This ensures that both sides of the formula are vectors in the same semantic vector space, satisfying the consistency of vector operations. This formula is used to introduce the guiding role of topic semantics on the basis of the original text semantic embedding. Its significance lies in dynamically adjusting the representation direction of the text in the semantic space by modeling the competitive relationship between the text and multiple topics. The specific improvement process is to introduce a Softmax normalization mechanism to probabilistically model the text-topic similarity on the basis of traditional linear weighted fusion, so that different topics form a competitive constraint relationship, and at the same time, by constructing The semantic offset replaces the direct superposition of topic vectors, thereby achieving a controllable transfer from the original semantics to the most relevant topic, so that the text representation can enhance its topic discrimination ability while maintaining the original semantic information. The semantic embedding representation guided by the topic constraint is used as input to perform nonlinear dimensionality reduction processing to obtain a set of low-dimensional semantic representations; The UMAP dimensionality reduction method is adopted to construct the high-dimensional semantic embedding into an adjacency graph structure. By maintaining local neighborhood relationships, the data is mapped through manifold learning, which compresses the high-dimensional semantic vector into a low-dimensional space while preserving the semantic similarity structure. The semantic similarity between texts is calculated based on a low-dimensional semantic representation set, a semantic similarity matrix is ​​generated, and a text topic relevance matrix is ​​constructed based on the text topic association score. The low-dimensional semantic representation vectors are matched pairwise, and the semantic similarity between texts is obtained by calculating the inverse function value of the cosine similarity or Euclidean distance between the vectors, and the corresponding similarity numerical relationship is constructed. A topic coupling kernel matrix is ​​constructed based on the semantic similarity matrix and the text topic relevance matrix. The topic coupling kernel matrix is ​​then used as the clustering input to perform density clustering based on the kernel matrix on the low-dimensional semantic representation set. During the clustering process, the topic consistency of texts within the same cluster is constrained based on the text topic relevance matrix to generate a set of topic clusters. Using the topic coupling kernel matrix as similarity input, high-density regions are identified based on the density reachability principle. Core samples and boundary samples are determined by calculating the kernel similarity neighborhood relationship between samples, and the text is clustered based on similarity connectivity. A class-level text set is constructed based on the topic cluster set. The texts within each topic cluster are merged to generate topic texts. The term weights are calculated based on the topic texts to obtain the topic keyword representation. Texts within the same topic cluster are concatenated in a preset order, and duplicate or highly similar sentences are removed. After unifying the text format, a continuous text set is formed for subsequent term statistics and topic representation construction. Based on the class-level TF-IDF method, each topic text is treated as a whole document. The term frequency of each term in the topic and its inverse document frequency in all topics are calculated to obtain the weight value of each term in the corresponding topic, which is used to characterize the importance of topic keywords. Topic vectors are generated based on topic keywords, and the probability distribution of text under each topic is calculated based on the similarity between topic vectors and text semantic embeddings, thus obtaining user interest topic vectors and corresponding topic probability distributions. The similarity between each text semantic embedding vector and each topic vector is calculated, and the similarity results corresponding to all topics are normalized so that the sum of the similarities of each topic is 1, thereby obtaining the probability distribution representation of the text under each topic.

[0023] In this embodiment, the topic constraint generation module includes: Obtain the topic probability distribution and construct the user topic distribution vector according to the user identifier; Based on the user topic distribution vector, the probabilities of each topic are sorted to obtain a topic sequence arranged in descending order of weight, and topics that meet the preset threshold conditions are selected from the topic sequence to form an effective topic set. The probability of each topic in the user topic distribution vector is compared with a preset threshold. Topics with a probability greater than or equal to the threshold are selected and retained as valid topics, while the remaining topics are removed. The topic probabilities in the effective topic set are normalized to generate normalized topic weights corresponding to each topic vector, and the normalized topic weights are bound to the corresponding topic vectors to form a topic weight mapping set. A user interest topic constraint function is constructed based on the topic weight mapping set. The normalized topic weight is used as a modulation coefficient to act on the corresponding topic vector, forming a multi-topic weighted constraint structure. The normalized topic weights are weighted and combined with the corresponding topic vectors to establish a mapping relationship with topic weights as control coefficients and topic vectors as constraint primitives. Based on this, a constraint function is defined for the input semantic vector in the semantic space to be jointly affected by various topics. The multi-topic weighted constraint structure is applied to the input semantic vector. The semantic correlation between the input semantic vector and each topic vector is calculated, and the semantic correlation is weighted and integrated with the corresponding normalized topic weights to obtain the constraint result of the input semantic vector under the user interest topic constraint. The similarity between the input semantic vector and each topic vector is calculated to obtain the semantic relevance value. Then, each semantic relevance value is multiplied by the corresponding normalized topic weight and summed to form a comprehensive semantic constraint value. This allows the input semantic vector to be jointly influenced by multiple topics according to their weight ratios during the calculation process. Based on the constraint results, the semantic expression of the input semantic vector is constrained and adjusted so that the input semantic vector converges in the semantic space towards the high-weight topic direction, generating a semantic constraint representation that conforms to the user topic distribution characteristics, and the semantic constraint representation is used as the output of the user interest topic constraint function; The input semantic vector is weighted and modulated based on the constraint results. By using the constraint results as adjustment factors on each dimension component of the input semantic vector, the vector is scaled or oriented, causing its semantic representation to shift towards the high-weight topic.

[0024] In this embodiment, the advertising slogan generation module includes: Obtain the user interest topic constraint function, and input the user interest topic constraint function as the generation constraint condition into the advertising slogan generation model; Based on the user interest topic constraint function, the conditional generation process is performed. During the generation process, the user interest topic constraint function is applied to the semantic representation of the generation model, so that the semantic expression in the generation process is modulated and constrained by the topic weight, and an advertising text sequence that satisfies the distribution of user interest topics is generated. The advertising text sequence is generated multiple times. Under the condition that the user interest topic constraint function remains unchanged, multiple advertising texts are generated by adjusting the decoding path of the generation model to obtain an initial advertising candidate set. While keeping the user interest topic constraint function unchanged, multiple decoding operations are performed on the same generated input. By changing the random sampling seed or decoding path selection strategy, the generation model can output multiple advertising texts with different semantic expressions under different decoding paths. The initial set of candidate advertising slogans is input into the user interest topic constraint function one by one. The semantic correlation between the semantic representation of each advertising slogan text and the topic vector is calculated, and the normalized topic weights are combined for weighted integration to obtain the topic constraint results of each advertising slogan text. Each advertising text is input into the semantic encoding model to obtain the corresponding semantic vector. The semantic vector is then matched one by one with the topic vectors. By calculating the cosine similarity or the inner product of the vectors, the semantic correlation between the advertising text and each topic is obtained. Based on the topic constraint results, the initial candidate set of advertising slogans is filtered, and advertising slogan texts that satisfy the user interest topic distribution constraints are retained to generate a constrained candidate set of advertising slogans. Based on the topic constraint results, each advertising text is thresholded, and advertising texts that meet the topic constraint scores are selected. The advertising texts that meet the conditions are then used as the final set of candidate advertising texts to be retained. The constrained candidate set of advertising slogans is input into the pre-trained language model for semantic encoding. Each advertising slogan text is then semantically encoded in context to generate an advertising slogan semantic vector corresponding to the advertising slogan text. The advertising text is input into a pre-trained language model, and word segmentation is used to obtain a sequence of word units. The model's multi-layer self-attention mechanism is then used to model the contextual relationships between word units, and the output is a text vector representation containing contextual semantic information.

[0025] In this embodiment, the unified semantic space mapping module includes: Obtain user interest topic vectors and construct a topic vector set; simultaneously obtain advertising slogan semantic vectors and construct an advertising slogan vector set. The topic vector set and the advertising slogan vector set are taken as input and input into the semantic embedding layer of the same pre-trained language model. A unified mapping process is performed on the topic vector and the advertising slogan semantic vector to generate the mapped topic representation vector and advertising slogan representation vector. The semantic embedding layer in the same pre-trained language model is a network structure inside the model that maps input text or vectors into fixed-dimensional continuous vector representations. This structure includes word vector mapping units and their subsequent linear transformation layers, which are used to project semantic vectors from different sources onto a unified vector representation space, so that each vector is represented under the same dimension and the same parameter system. The topic vector set and the advertising slogan semantic vector set are respectively input into the same semantic embedding layer. The vectors are linearly transformed and feature projected using the same mapping parameters, so that the vectors from different sources are converted into semantic vectors with consistent dimensions and representation. A semantic space is constructed based on the mapped topic representation vector and the advertising slogan representation vector. The semantic space is a vector space defined by the semantic embedding layer. The topic representation vector and the advertising slogan representation vector have the same vector dimension and a unified semantic measurement method in the semantic space. Based on the topic representation vector and the advertising slogan representation vector in the semantic space, the semantic similarity between each advertising slogan representation vector and each topic representation vector is calculated to obtain the semantic relevance between the advertising slogan and each topic. The semantic relevance results are arranged according to the correspondence between the advertising slogan representation vector and the topic representation vector to construct a semantic relevance matrix. The rows of the semantic relevance matrix correspond to the advertising slogan representation vector, and the columns correspond to the topic representation vector. Each element in the semantic relevance matrix represents the semantic relevance between the corresponding advertising slogan and the topic.

[0026] In this embodiment, the topic coupling kernel matrix construction module includes: Obtain the semantic relevance matrix, extract the topic relevance vector corresponding to each advertising slogan from the semantic relevance matrix, and obtain the set of semantic vectors for each advertising slogan; Based on the set of semantic vectors of advertising slogans, calculate the semantic similarity between any two semantic vectors of advertising slogans, and generate a set of semantic similarity relationships of advertising slogans; Match any two advertising slogan semantic vectors one by one, and obtain the semantic similarity value between the corresponding advertising slogan semantic vectors by calculating the cosine similarity or inner product value between the vectors. Associate and bind the topic relevance vector with the corresponding advertising slogan semantic vector, so that each advertising slogan semantic vector corresponds to a unique topic relevance vector; Based on the semantic similarity relationship set of advertising slogans and the topic relevance vector, the semantic similarity between any two advertising slogans is modulated so that the semantic similarity is adjusted according to the difference in the topic relevance vectors of the corresponding two advertising slogans, thereby generating advertising slogan similarity relationship after topic constraint; Based on the topic relevance vectors of the two corresponding advertising slogans, the degree of topic difference is calculated, and this degree of difference is used as a modulation factor to apply to the original semantic similarity. The semantic similarity is scaled or adjusted to obtain the similarity value after topic constraint. Based on the similarity relationship of advertising slogans constrained by the theme, all advertising slogans are combined in pairs and arranged to construct a theme coupling kernel matrix. Each element in the theme coupling kernel matrix is ​​determined by the modulated similarity relationship between the corresponding advertising slogans.

[0027] In this embodiment, the advertising slogan screening and delivery module includes: Obtain the topic coupling kernel matrix and the set of advertising slogan candidates corresponding to each index position in the topic coupling kernel matrix; Based on the index correspondence between the topic coupling kernel matrix and the advertising slogan candidate set, a deterministic point process model is constructed. The topic coupling kernel matrix is ​​used as the kernel function matrix, and a mapping relationship is established between any subset of the advertising slogan candidate set and the corresponding kernel submatrix in the topic coupling kernel matrix. A candidate subset set is generated based on the candidate set of advertising slogans, and the kernel submatrix corresponding to each candidate subset is extracted from the topic coupling kernel matrix according to the mapping relationship; The determinant of the nucleus matrix is ​​calculated to obtain the determinant value of the corresponding candidate subset, and the determinant value is used as the selection metric for the candidate subset; Perform matrix determinant operation on the kernel matrix corresponding to each candidate subset, and calculate its determinant value by decomposing or recursively expanding the matrix to obtain the numerical result used to characterize the structural features of the subset. Based on the selection metric, a subset selection process is performed. Candidate advertising slogans are selected from the candidate set of advertising slogans and added to the current subset. After each addition, the kernel matrix is ​​updated according to the newly added advertising slogans until the preset subset size condition is met, and the optimal set of advertising slogans is generated. In the candidate set of advertising slogans, candidate advertising slogans are selected one by one in a preset order and added to the current subset. After each addition, the determinant value of the corresponding kernel matrix is ​​recalculated. The result of the change in the determinant value determines whether to retain the candidate advertising slogan until the subset size condition is met. Each time a new candidate slogan is selected and added to the current subset, its corresponding row and column are added to the existing kernel submatrix according to the index position of the slogan in the topic coupling kernel matrix, expanding to form a new kernel submatrix. The expanded kernel matrix is ​​then updated and calculated. This process continues until the number of slogans in the subset reaches the preset scale condition. Obtain the probability distribution of user interest topics, and construct a user topic distribution vector based on the topic probability distribution; The advertising slogans in the optimal set are associated with the topic relevance vectors corresponding to the semantic relevance matrix, and the topic relevance vectors of each advertising slogan are matched and calculated based on the user topic distribution vector to obtain the matching relationship between the advertising slogans and the target user group. Based on the matching relationship, the advertising slogans in the optimal set are assigned to the corresponding target user groups, and the advertising slogans are accurately delivered.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to an internet e-commerce advertising scenario. The platform needs to generate attractive advertising slogans that match user interests for different user groups and accurately deliver them to the target users. Traditional systems typically rely on keyword matching or fixed templates to generate advertising slogans. For example, for users searching for "sports shoes," the system directly generates generic slogans such as "high-quality sports shoes, comfortable and durable." While these slogans have a certain relevance, they lack expression of users' deeper interests. Furthermore, multiple slogans suffer from content repetition and monotonous expression, resulting in low click-through rates and unsatisfactory advertising conversion rates. In addition, during the advertising delivery phase, traditional methods typically match based on simple similarity or click history, lacking fine-grained modeling of semantic structure, making it difficult to achieve high-precision matching.

[0029] In this embodiment, the present invention is applied to an advertising slogan generation and delivery system for an e-commerce platform. The system first collects user behavior data on the platform, including user search keyword sequences, titles and descriptions of product pages viewed, and interactive text records generated when users click or skip advertisements. Taking a specific user as an example, over a period of time, this user repeatedly searched for "running shoes," "shock-absorbing sports shoes," and "lightweight and breathable shoes," browsed multiple sports brand product pages, and clicked on some advertisements. After uniformly structuring this text data, the system forms a set of user behavior texts, and through text cleaning, word segmentation, and semantic encoding, generates a user behavior semantic vector.

[0030] The system inputs the user's behavioral semantic vector into the topic semantic modeling module, and extracts the user's interest topics through an improved topic modeling method. For example, the user is identified as primarily interested in three topics: "running performance," "lightweight design," and "breathability," with corresponding topic probability distributions of 0.42, 0.33, and 0.25, respectively. Based on this topic distribution, the system constructs a user interest topic constraint function, mapping the aforementioned topic weights to semantic constraints for the generation of subsequent advertising copy.

[0031] During the ad copy generation process, the system uses a topic constraint function as input to the model, constraining semantic expression to ensure the generated content revolves around topics of interest to users. The system generates an initial set of candidate ad copy through multiple iterations, such as: "Lightweight and cushioned running experience, helping you move forward with every step," "Breathable design combined with comfortable feel, making running easier," and "Professional shock absorption technology, bringing stable running performance." Compared to traditional template generation, these ad copy slogans are semantically closer to user interests.

[0032] After generating candidate ad copy, the system semantically encodes each ad copy and maps it to a unified semantic space, aligning it with user interest topic vectors. By calculating the semantic relevance between ad copy and topics, the system obtains the degree of matching between ad copy and each topic. Subsequently, the system constructs a semantic relevance matrix and further builds a topic coupling kernel matrix based on the semantic similarity and topic relevance relationships between ad copy. This kernel matrix not only reflects the semantic similarity between ad copy but also introduces topic distribution differences, enabling ad copy that is semantically similar but has significant topic deviations to be separated in subsequent filtering.

[0033] During the slogan selection phase, the system constructs a determinant-based process model based on a topic-coupled kernel matrix to select a subset of the candidate slogans. The system calculates the determinant of the kernel matrix corresponding to different slogan combinations and selects the subset with the largest determinant value as the optimal set of slogans. This process ensures that the selected slogans possess high semantic diversity while maintaining relevance to user interests. For example, from 15 candidate slogans, the system ultimately selects 5 slogans that highlight different aspects such as "shock absorption performance," "lightweight design," "breathable structure," "long-term comfort," and "movement stability," avoiding content duplication.

[0034] During the campaign delivery phase, the system matches the optimal set of selected ad copy with the distribution of user interest themes, assigning different ad copy to the corresponding user groups. For example, for users whose primary interest is "shock absorption performance," ad copy emphasizing cushioning effects is prioritized, while for users who are more concerned about "breathability," ad copy with related expressions is prioritized, thus achieving precise ad copy distribution.

[0035] Verification using actual operational data on the platform shows that the method of this invention outperforms traditional methods in multiple metrics. In the ad copy generation stage, compared to traditional template generation methods, the semantic relevance score is improved by an average of approximately 23%, and the consistency between the generated text and user interest topics is significantly enhanced. In the ad copy selection stage, by introducing a topic coupling kernel matrix and a matrix-based point process model, the repetition rate of the ad copy set decreases from 32% to 11%, and the semantic diversity index improves by approximately 41%. Regarding ad performance, after the system went live, the user click-through rate increased from 2.8% to 4.1%, an increase of approximately 46%, the conversion rate increased from 1.2% to 1.9%, and the return on investment for advertising improved by approximately 38%.

[0036] In tests with different user scales, when the number of candidate ad copy was 20, the traditional method still left about 7 semantically similar ad copy after screening, while the method of this invention reduced the number of semantically highly similar ad copy to less than 2. Furthermore, in large-scale user tests, the system processed ad delivery tasks for over 100,000 users, maintaining an average ad copy matching time of less than 120 milliseconds, meeting real-time delivery requirements.

[0037] As can be seen from the above embodiments, the present invention forms a complete semantic processing chain in all stages of advertising slogan generation, screening and placement. From user behavior semantic modeling to topic constraint generation, and then to advertising slogan generation and screening, it finally achieves accurate placement and effectively solves the problems of insufficient semantic matching accuracy, advertising slogan repetition and poor placement effect in the prior art.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent advertisement sentence generation and accurate delivery system fused with natural language processing, characterized in that, Includes the following modules: The user behavior corpus construction module collects user search keyword data, user browsing page text data, and user historical advertising interaction text data, and performs text cleaning, word segmentation, and semantic embedding encoding to generate user behavior semantic vector data. The topic semantic modeling module inputs the user behavior semantic vector data into the BERTopic model and outputs user interest topic vectors and corresponding topic probability distributions. The topic constraint generation module constructs user interest topic constraint functions based on topic probability distribution. The ad slogan generation module generates a candidate set of ad slogans under the constraints of the user interest topic constraint function, and performs semantic encoding on the candidate set of ad slogans to obtain the semantic vector of the ad slogans; The unified semantic space mapping module maps user interest topic vectors and advertising slogan semantic vectors to the same semantic space, and calculates the semantic relevance matrix between advertising slogans and user interest topics based on the mapping results; The topic coupling kernel matrix construction module constructs a topic coupling kernel matrix based on the similarity relationship between the semantic relevance matrix and the semantic vector of the advertising slogan; The ad slogan selection and delivery module constructs a matrix-based process model using the theme coupling kernel matrix as input. It performs subset selection on the candidate ad slogan set to obtain the optimal ad slogan set, and matches the optimal ad slogan set to the target user group based on the theme probability distribution, outputting accurate ad slogan delivery results. 2.The system of claim 1, wherein, The user behavior corpus construction module includes: Collect user search keyword data, which includes the sequence of query terms entered by the user, query time information, and query frequency information; Collect user browsing page text data, which includes webpage title text, body text text, and page tag information corresponding to keyword sequences; Collect user's historical advertising interaction text data, which includes the advertising title text, advertising description text, and interaction behavior markers triggered by the user during browsing. User search keyword data, user page browsing text data, and user historical advertising interaction text data are processed in a unified structure and then concatenated in chronological order to generate the original user behavior text set. The keyword sequence set, browsing text sequence and advertising interaction text set are uniformly converted into structured records containing user identifier, timestamp, text content and behavior type fields, and sorted and aligned according to timestamp to form time-series text data in a unified format. The original user behavior text set is subjected to text cleaning processing, which includes removing stop words, invalid symbols, HTML tags and unified character encoding, to generate a standardized text set; Perform word segmentation on the standardized text set, dividing the standardized text into word sequence and constructing a word list; Semantic embedding encoding is performed on the lexical sequence based on the lexical vocabulary, mapping the lexical sequence to a semantic vector representation, and the lexical vectors are aggregated at the text granularity to generate user behavior semantic vector data corresponding to user behavior. 3.The system of claim 1, wherein, The topic semantic modeling module includes: User behavior semantic vector data is used as text corpus. The text corpus is then semantically encoded based on a pre-trained language model to generate corresponding high-dimensional text semantic embedding representations and to construct a set of text semantic embeddings. The topic information in historical corpora is extracted based on the set of text semantic embeddings, a set of topic prior vectors is constructed, and the similarity between each text semantic embedding and each topic prior vector is calculated to obtain the text topic association score. Based on the text topic association score, the text semantic embedding is reconstructed directionally to generate a topic-constrained semantic embedding representation; The semantic embedding representation guided by the topic constraint is used as input to perform nonlinear dimensionality reduction processing to obtain a set of low-dimensional semantic representations; The semantic similarity between texts is calculated based on a low-dimensional semantic representation set, a semantic similarity matrix is ​​generated, and a text topic relevance matrix is ​​constructed based on the text topic association score. A topic coupling kernel matrix is ​​constructed based on the semantic similarity matrix and the text topic relevance matrix. The topic coupling kernel matrix is ​​then used as clustering input to perform density clustering based on the kernel matrix on the low-dimensional semantic representation set, generating a set of topic clusters. A class-level text set is constructed based on the topic cluster set. The texts within each topic cluster are merged to generate topic texts. The term weights are calculated based on the topic texts to obtain the topic keyword representation. Topic vectors are generated based on topic keywords, and the probability distribution of text under each topic is calculated based on the similarity between topic vectors and text semantic embeddings, thus obtaining user interest topic vectors and corresponding topic probability distributions.

4. The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to claim 1, characterized in that, The topic constraint generation module includes: Obtain the topic probability distribution and construct the user topic distribution vector according to the user identifier; Based on the user topic distribution vector, the probabilities of each topic are sorted to obtain a topic sequence arranged in descending order of weight, and topics that meet the preset threshold conditions are selected from the topic sequence to form an effective topic set. The topic probabilities in the effective topic set are normalized to generate normalized topic weights corresponding to each topic vector, and the normalized topic weights are bound to the corresponding topic vectors to form a topic weight mapping set. A user interest topic constraint function is constructed based on the topic weight mapping set. The normalized topic weight is used as a modulation coefficient to act on the corresponding topic vector, forming a multi-topic weighted constraint structure. The multi-topic weighted constraint structure is applied to the input semantic vector. The semantic correlation between the input semantic vector and each topic vector is calculated, and the semantic correlation is weighted and integrated with the corresponding normalized topic weights to obtain the constraint result of the input semantic vector under the user interest topic constraint. Based on the constraint results, the semantic expression of the input semantic vector is constrained and adjusted to generate a semantic constraint representation that conforms to the user topic distribution characteristics. The semantic constraint representation is then used as the output of the user interest topic constraint function.

5. The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to claim 1, characterized in that, The advertising slogan generation module includes: Obtain the user interest topic constraint function, and input the user interest topic constraint function as the generation constraint condition into the advertising slogan generation model; Based on the user interest topic constraint function, the execution condition generation process generates an advertising text sequence that satisfies the user interest topic distribution; The advertising text sequence is generated multiple times. Under the condition that the user interest topic constraint function remains unchanged, multiple advertising texts are generated by adjusting the decoding path of the generation model to obtain an initial advertising candidate set. The initial set of candidate advertising slogans is input into the user interest topic constraint function one by one. The semantic correlation between the semantic representation of each advertising slogan text and the topic vector is calculated, and the normalized topic weights are combined for weighted integration to obtain the topic constraint results of each advertising slogan text. Based on the topic constraint results, the initial candidate set of advertising slogans is filtered, and advertising slogan texts that satisfy the user interest topic distribution constraints are retained to generate a constrained candidate set of advertising slogans. The constrained candidate set of advertising slogans is input into a pre-trained language model for semantic encoding. Each advertising slogan text is then subjected to contextual semantic encoding to generate an advertising slogan semantic vector corresponding to the advertising slogan text.

6. The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to claim 1, characterized in that, The unified semantic space mapping module includes: Obtain user interest topic vectors and construct a topic vector set; simultaneously obtain advertising slogan semantic vectors and construct an advertising slogan vector set. The topic vector set and the advertising slogan vector set are taken as input and input into the semantic embedding layer of the same pre-trained language model. A unified mapping process is performed on the topic vector and the advertising slogan semantic vector to generate the mapped topic representation vector and advertising slogan representation vector. A semantic space is constructed based on the mapped topic representation vector and the advertising slogan representation vector, wherein the semantic space is a vector space defined by the semantic embedding layer; Based on the topic representation vector and the advertising slogan representation vector in the semantic space, the semantic similarity between each advertising slogan representation vector and each topic representation vector is calculated to obtain the semantic relevance between the advertising slogan and each topic. The semantic relevance results are arranged according to the correspondence between the advertising slogan representation vector and the topic representation vector to construct a semantic relevance matrix. The rows of the semantic relevance matrix correspond to the advertising slogan representation vector, and the columns correspond to the topic representation vector. Each element in the semantic relevance matrix represents the semantic relevance between the corresponding advertising slogan and the topic.

7. The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to claim 1, characterized in that, The topic coupling kernel matrix construction module includes: Obtain the semantic relevance matrix, extract the topic relevance vector corresponding to each advertising slogan from the semantic relevance matrix, and obtain the set of semantic vectors for each advertising slogan; Based on the set of semantic vectors of advertising slogans, calculate the semantic similarity between any two semantic vectors of advertising slogans, and generate a set of semantic similarity relationships of advertising slogans; Associate and bind the topic relevance vector with the corresponding advertising slogan semantic vector; Based on the semantic similarity relationship set of advertising slogans and the topic relevance vector, the semantic similarity between any two advertising slogans is modulated to generate advertising slogan similarity relationship after topic constraint; Based on the similarity relationship of advertising slogans constrained by themes, the results of pairwise combinations of all advertising slogans are arranged to construct a theme coupling kernel matrix.

8. The intelligent advertising slogan generation and precise delivery system integrating natural language processing according to claim 1, characterized in that, The advertising slogan selection and placement module includes: Obtain the topic coupling kernel matrix and the set of advertising slogan candidates corresponding to each index position in the topic coupling kernel matrix; Based on the index correspondence between the topic coupling kernel matrix and the advertising slogan candidate set, a deterministic point process model is constructed. The topic coupling kernel matrix is ​​used as the kernel function matrix, and a mapping relationship is established between any subset of the advertising slogan candidate set and the corresponding kernel submatrix in the topic coupling kernel matrix. A candidate subset set is generated based on the candidate set of advertising slogans, and the kernel submatrix corresponding to each candidate subset is extracted from the topic coupling kernel matrix according to the mapping relationship; The determinant of the nucleus matrix is ​​calculated to obtain the determinant value of the corresponding candidate subset, and the determinant value is used as the selection metric for the candidate subset; Based on the selection metric, a subset selection process is performed. Candidate advertising slogans are selected from the candidate set of advertising slogans and added to the current subset. After each addition, the kernel matrix is ​​updated according to the newly added advertising slogans until the preset subset size condition is met, and the optimal set of advertising slogans is generated. Obtain the probability distribution of user interest topics, and construct a user topic distribution vector based on the topic probability distribution; The advertising slogans in the optimal set are associated with the topic relevance vectors corresponding to the semantic relevance matrix, and the topic relevance vectors of each advertising slogan are matched and calculated based on the user topic distribution vector to obtain the matching relationship between the advertising slogans and the target user group. Based on the matching relationship, the advertising slogans in the optimal set are assigned to the corresponding target user groups, and the advertising slogans are accurately delivered.

Images