E-commerce platform marketing content matching method based on semantic reasoning

Abstract

The invention discloses an e-commerce platform marketing content matching method based on semantic reasoning, and the method comprises the following steps: S1, generating a multi-hop interaction chain according to a user historical behavior sequence; s2, gradually performing mask prompt and inputting a language model, performing interactive chain thinking semantic reasoning, and generating a recommended candidate content set; s3, calculating the matching degree of the candidate content and the context semantics, and determining an initial sequence; s4, iteratively updating the sorting weight by adopting a secretary bird optimization algorithm; s5, adjusting the sorting of the candidate contents according to the sorting weight; and S6, updating a language model reasoning strategy and optimization algorithm parameters according to user feedback. According to the invention, the accuracy of the recommended content and the user experience are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce marketing content recommendation, and in particular to a method for matching marketing content on an e-commerce platform based on semantic reasoning. Background Art

[0002] With the rapid development of e-commerce, accurately matching marketing content with user interests has become a key issue that e-commerce platforms urgently need to address. Currently, methods for matching marketing content based on user historical behavior data mainly include collaborative filtering, content-based recommendation, and hybrid recommendation methods. Collaborative filtering utilizes the similarity of user historical behavior to make recommendations, but suffers from insufficient recommendation effectiveness for new users or in sparse data scenarios. Content-based recommendation methods primarily rely on static product attributes for feature matching, ignoring the dynamic changes in user behavior and making it difficult to fully capture users' real-time interests. Hybrid recommendation methods, while taking into account both user behavior and content features, still have limitations in capturing complex user intentions and dynamic changes in interests due to insufficient understanding of the semantics behind user interactions.

[0003] In recent years, with the development of pre-trained language models and deep learning technologies, marketing content recommendation methods based on semantic reasoning have gradually emerged. Existing semantic reasoning methods often rely on single-step interaction information or simple semantic matching rules, failing to effectively mine and utilize the multi-hop interaction links implicit in user historical behavior, resulting in an inability to accurately understand the chained semantic logic behind user behavior. Furthermore, existing ranking optimization methods based on intelligent optimization algorithms, such as particle swarm optimization and ant colony optimization, while somewhat effective in recommendation ranking applications, are often prone to local optimality and struggle to dynamically adapt to user feedback, making them unable to effectively achieve real-time optimization and accurate recommendation of marketing content ranking.

[0004] In summary, the existing technologies still have obvious shortcomings in achieving marketing content recommendation, such as insufficient depth of semantic understanding, insufficient mining of multi-hop interactive behaviors, and poor adaptability of sorting optimization methods.

[0005] Therefore, how to provide an e-commerce platform marketing content matching method based on semantic reasoning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a marketing content matching method for an e-commerce platform based on semantic reasoning, which optimizes the content sorting weight by combining interactive chain semantic reasoning with the Secretary Bird optimization algorithm, thereby achieving the effects of improving recommendation accuracy, enhancing sorting stability and optimizing user experience.

[0007] The e-commerce platform marketing content matching method based on semantic reasoning according to an embodiment of the present invention includes the following steps:

[0008] S1. Based on the user's historical behavior sequence on the e-commerce platform, the continuous multi-hop interaction behavior path between the user and the product is traced step by step to generate a multi-hop interaction chain for chain reasoning;

[0009] S2. For each multi-hop interaction chain, the interactive behaviors in the chain are gradually masked and prompted, and the pre-trained language model that has undergone supervised fine-tuning and reinforcement training is input to perform interactive chain thinking semantic reasoning to generate a set of recommended candidate content associated with the current target marketing content;

[0010] S3. For each candidate content in the recommended candidate content set, respectively calculate the matching degree score between the candidate content and the context semantics obtained by the interactive chain thinking semantic reasoning, and determine the initial ranking order of the candidate content;

[0011] S4. Build the initial search space for the Secretary Bird optimization algorithm based on the initial ranking order of candidate content, historical conversion rate prediction indicators, and content timeliness indicators. Simulating the Secretary Bird's exploration and jumping mechanism and local adjustment mechanism to avoid danger during hunting, it iteratively updates the ranking weight of each candidate content generation by generation until the preset convergence condition is reached.

[0012] S5. Re-adjust the ranking of candidate content using the optimized candidate content ranking weights to obtain a final ranking list for user marketing content recommendation;

[0013] S6. Collect user interactive feedback data on recommended marketing content in real time, and continuously update the chain semantic reasoning strategy of the pre-trained language model and the search variable parameters of the Secretary Bird optimization algorithm.

[0014] Optionally, the S1 specifically includes:

[0015] S11. Collecting the user's historical behavior sequence on the e-commerce platform, the historical behavior sequence includes the user's stay time on each product page, the interval between consecutive clicks, the page browsing depth, and the timestamps corresponding to the add-to-cart and favorites actions;

[0016] S12. Based on the collected historical behavior sequences, analyze the cross-category migration patterns of users between different product categories, identify the cross-category behavior association paths exhibited by users across multiple product categories, and record the cross-category association strength of each path;

[0017] S13. Calculate the comprehensive interaction strength of each interaction sequence in the historical behavior sequence using the cross-category association strength, and based on the comprehensive interaction strength, filter out interaction sequences with comprehensive interaction strengths below a preset threshold from the historical behavior sequence, and retain interaction sequences with comprehensive interaction strengths greater than or equal to the preset threshold;

[0018] S14. Based on the filtered interaction sequence, a reverse, step-by-step backtracking method based on the causal chain of user behavior is used. Starting from the target product, the interaction sequence between the user and historical products is reversed step by step to verify the logical rationality of the causal interactions between the interaction sequences step by step, and invalid interaction sequences introduced by non-active user behavior due to random or accidental touches are eliminated;

[0019] S15. Based on the interaction sequences retained after reverse retrospective verification, optimize and adjust the interaction sequences using cross-category association strength and user behavior stability indicators. Select the interaction sequence with the highest semantic relevance and optimal behavior logic for the target product recommendation in the historical behavior sequence to form an optimized multi-hop interaction behavior path. The user behavior stability indicator is a statistical indicator of the number of repeated interactions, the duration of a single interaction, and the duration between multiple interactions during the user's interaction with the same product or products of the same category.

[0020] S16. Based on the optimized multi-hop interaction behavior path, the interaction sequence is reorganized by combining the long-term preference characteristics and short-term interest fluctuation trends in the historical behavior sequence. The multi-hop interaction chain is constructed with the time continuity, behavior category alternation and semantic theme consistency of the interaction sequence as constraints.

[0021] Optionally, the S2 specifically includes:

[0022] S21. Selecting the interaction sequences in each multi-hop interaction chain in sequence, and calculating the historical interest reduction weight of each node according to the time decay coefficient of the interaction sequence node in the historical behavior sequence in the order from the historical interaction node to the target node, and gradually implementing mask prompts for each product node in the sequence based on the historical interest reduction weights to form a progressive mask prompt sequence;

[0023] S21. Select the interaction sequences in each multi-hop interaction chain in sequence, and calculate the corresponding historical interest reduction weight of each interaction sequence according to the time decay coefficient of each interaction sequence in the historical behavior sequence in the order from the historical behavior sequence to the target product. Then, mask prompts are gradually implemented on the interaction sequences based on the historical interest reduction weights to form a progressive mask prompt sequence.

[0024] S22. Gradually input the progressive mask prompt sequence into the pre-trained language model that has undergone supervised fine-tuning and reinforcement training, dynamically weight the attention mechanism of the pre-trained language model using the historical interest decreasing weight, and predict the semantic information of the masked product node by node;

[0025] S23, real-time monitoring of the cosine similarity change trend of the semantic vectors of the predicted semantics of the masked products in each progressive mask prompt sequence during the pre-trained language model prediction process, using the fluctuation amplitude of the cosine similarity change trend as a semantic fluctuation index, and dynamically increasing the number of mask prompt nodes when the semantic fluctuation index does not reach a preset stability threshold until the semantic fluctuation index reaches or exceeds the stability threshold;

[0026] S24. Based on the stable progressive mask prompt sequence, analyze the cosine similarity between the predicted semantic vector of the masked product and the semantic vector of the target product by backtracking in reverse order, and quantitatively calculate the logical consistency score between each masked product and the target product;

[0027] S25. Based on the logical consistency score, a logical consistency threshold is set, and recommendation candidate content having a logical consistency score exceeding the logical consistency threshold is matched and extracted from the historical marketing content library to generate a set of matching recommendation candidate content;

[0028] S26, and store the recommended candidate content set and the predicted semantic information of all masked products in the corresponding chain thinking semantic reasoning process in association with each other.

[0029] Optionally, the pre-trained language model specifically includes a multi-granularity product embedding module, an interactive behavior temporal perception module, a bidirectional chain attention mechanism module, a mask prompt adaptive control module, a local semantic difference amplification module, a semantic path memory and retrieval module, and a self-enhanced feedback fine-tuning module:

[0030] The multi-granularity product embedding module receives a progressive mask prompt sequence and generates a multi-granularity product semantic vector by integrating a contextual dynamic weighting mechanism;

[0031] The interaction behavior temporal sequence perception module uses a continuous value encoded time decay function to calculate the temporal decay coefficient of each interaction sequence based on the timestamp corresponding to the interaction behavior to characterize the temporal decay relationship of the user's historical behavior;

[0032] The bidirectional chain attention mechanism module calculates a bidirectional chain semantic association representation of the interaction sequence based on the bidirectional path from the historical behavior sequence to the target product and from the target product to the historical behavior sequence based on the decreasing weight of the historical interest;

[0033] The mask prompt adaptive control module calculates the fluctuation amplitude of the cosine similarity of the predicted semantic vectors of the masked products in the progressive mask prompt sequence in real time, and dynamically adjusts the number of masked products based on the fluctuation amplitude until the semantic fluctuation amplitude reaches a stability threshold;

[0034] The local semantic difference amplification module amplifies the local differences between the predicted semantic vectors through nonlinear attention expansion based on the stable predicted semantic vectors output by the mask prompt adaptive control module;

[0035] The semantic path memory and retrieval module stores the predicted semantic vectors in the chain thinking semantic reasoning process in real time and retrieves the predicted semantic vectors based on the attention addressing method;

[0036] The self-enhanced feedback fine-tuning module receives user interactive feedback data on recommended marketing content in real time, uses the logical consistency score of the predicted semantic vector as a reward function, dynamically integrates supervised learning and reinforcement learning strategies, and continuously updates the network parameters of the pre-trained language model.

[0037] Optionally, the S23 specifically includes:

[0038] S231, real-time monitoring of the semantic vector cosine similarity fluctuation trend of the predicted semantic vector of the masked product in each progressive mask prompt sequence, and calculating the semantic entropy fluctuation, semantic direction deviation speed, and semantic aggregation change rate to form a multi-dimensional semantic fluctuation indicator set;

[0039] S232. Constructing a multidimensional semantic fluctuation tensor based on the multidimensional semantic fluctuation index set, and mapping the fluctuation index of each dimension to a unified high-dimensional space;

[0040] S233. For the multidimensional semantic fluctuation tensor, a dynamic attention weight fusion mechanism is adopted to update the fusion weight of each semantic fluctuation dimension in real time with the prediction stability index as the objective function, and generate a real-time updated fusion fluctuation trend vector;

[0041] S234, based on the real-time updated fusion fluctuation trend vector, a nonlinear adaptive threshold control strategy is used to determine the number of mask prompt nodes and the mask strength of the current mask prompt sequence;

[0042] S235 , repeating steps S231 to S234 until the fused fluctuation trend vector is lower than or equal to a preset stability threshold, and outputting a stable progressive mask prompt sequence after adaptive adjustment.

[0043] Optionally, the S3 specifically includes:

[0044] S31, extracting the product title text, product attribute feature data, and user comment text corresponding to each candidate content in the recommended candidate content set, and generating a multi-granularity semantic vector for each candidate content based on the multi-granularity product embedding module;

[0045] S32. Using the target product prediction semantic vector generated during the chain thinking semantic reasoning process as the central reference vector, a nonlinear radial basis mapping function is used to map the multi-granularity semantic vector of each candidate content into a two-dimensional polar coordinate semantic space centered on the target product prediction semantic vector.

[0046] S33. Calculate the radial distance parameter R of each candidate content in the two-dimensional polar coordinate semantic space:

[0047]

[0048] λ i =1+β·|X i -Y i | γ ;

[0049] Among them, R is the radial distance parameter, which represents the semantic distance between the candidate content and the target product, X i is the i-th dimension feature value of the candidate content multi-granularity semantic vector, Y i Predict the i-th dimension eigenvalue of the semantic vector for the target product, λ i is the nonlinear distance weighting factor of the i-th dimension feature, β is the distance amplification adjustment coefficient, which is used to control the prominence of the key dimension, γ is the nonlinear amplification index, and n is the dimension of the semantic vector;

[0050] S34. Calculate the semantic offset angle parameter θ of each candidate content:

[0051]

[0052] Among them, θ is the semantic offset angle parameter, ω i is the direction-sensitive weight of the i-th dimension feature, δ is the semantic direction sensitivity adjustment coefficient, and ∈ is a very small positive number to prevent the denominator from being zero;

[0053] S35. Perform nonlinear fusion of the radial distance parameter R and the semantic offset angle parameter θ to calculate the comprehensive semantic matching score P for each candidate content:

[0054]

[0055] Among them, P is the comprehensive semantic matching score, which integrates the characteristics of semantic distance and semantic offset direction, α is the distance attenuation adjustment coefficient, μ is the angle sensitivity enhancement index, R is the radial distance parameter, and θ is the semantic offset angle parameter;

[0056] S36. Determine an initial sequence of comprehensive semantic matching scores P for each candidate content in the recommended candidate content set, and sort them in descending order to form an initial sort order for the candidate content set.

[0057] S37. The position coordinate data of the candidate content in the two-dimensional polar coordinate semantic space is associated with the corresponding comprehensive semantic matching degree score P and stored to provide a visual explanation basis for interactive chain thinking semantic reasoning.

[0058] Optionally, the S4 specifically includes:

[0059] S41. Convert the initial ranking order of the candidate content, the historical conversion rate prediction index, and the content timeliness index into corresponding ranking dimensions, conversion rate dimensions, and timeliness dimensions, respectively, and combine them in a unified space to construct an initial search space with the candidate content weight combination as the spatial position. The historical conversion rate prediction index is based on the user's historical behavior data on the e-commerce platform, and is obtained by statistically analyzing the user's historical conversion behavior for similar candidate content or products of the same category to obtain the predicted conversion probability for each candidate content. The content timeliness index is a comprehensive value obtained by quantitatively analyzing the interval length from the candidate content's publication to the current time, the popularity decay rate of the content topic, and the real-time attention change trend of the product or marketing topic involved in the candidate content;

[0060] S42. In the initial search space, based on the density of the candidate content, a Gaussian kernel smoothing method is used to calculate the density value of each position in the space, and the area with higher density in the space is determined as the priority exploration area for the search;

[0061] S43 simulates the density gradient-guided jump exploration mechanism used by secretary birds during hunting. The probability distribution of the next exploration location is determined based on the density gradient value in space. Locations with larger density gradients have higher exploration probabilities. At each iteration, a new search location is determined non-uniformly based on the probability, and random perturbations are introduced in each jump.

[0062] S44. For each determined search position, calculate the ranking performance fitness score of the candidate content combination, and based on the ranking performance fitness score, construct a fitness field that describes the ranking performance fitness change trend. Use the fitness gradient information in the fitness field to determine the search direction with the greatest potential for improving ranking performance fitness.

[0063] S45. Simulate the local adjustment mechanism used by secretary birds to avoid danger during hunting. Identify locations in the fitness field where the sorting performance fitness gradient fluctuates dramatically as high-risk areas. Determine the safe distance from the high-risk areas based on the degree of fitness gradient fluctuation at each search location. Adjust the spatial position of each search location in real time to actively avoid high-risk areas and move toward areas where the sorting performance fitness gradient fluctuates smoothly.

[0064] S46. Loop execution based on the density gradient-guided jump exploration mechanism and the local adjustment mechanism for avoiding high-risk areas, and update the candidate content sorting weight combination in real time until the sorting performance fitness gradient change of two consecutive iterations is less than or equal to the preset gradient convergence threshold, and determine the optimal sorting weight when the Secretary Bird optimization algorithm finally converges.

[0065] Optionally, the S5 specifically includes:

[0066] S51, mapping the optimized candidate content ranking weights and the initial ranking order of the candidate content to the same ranking weight adjustment space, and determining the re-ranking position of the candidate content through a nonlinear dynamic mapping strategy;

[0067] S52. Establish a ranking stability monitoring mechanism for the re-ranked candidate content. By continuously tracking the position change trajectory of the candidate content in the ranking weight adjustment space, the consistency of the ranking position change magnitude and ranking change direction of each candidate content over multiple consecutive adjustment cycles is calculated to obtain a ranking stability index.

[0068] S53. For candidate content whose ranking stability index is lower than a preset stability threshold, the multi-hop interaction chain associated with the target product in the chain thinking semantic reasoning process is traced back step by step, and the continuous semantic state change sequence of the multi-hop interaction chain is gradually extracted. The difference change trend between the continuous semantic state change sequence and the multi-granularity semantic vector of the corresponding candidate content is analyzed to locate the fundamental semantic cause of the insufficient ranking stability of the candidate content;

[0069] S54. Based on the fundamental semantic reason, a dynamic semantic difference compensation mechanism is used to calculate the semantic difference compensation value between the candidate content and the target product in real time. The semantic difference compensation value is used to locally fine-tune the ranking weight of the candidate content to eliminate the inconsistent semantics between the multi-hop interaction chain and the candidate content.

[0070] S55. Recalculate the ranking comprehensive score of each candidate content based on the ranking weight after implementing the semantic difference compensation mechanism, and re-rank the candidate content based on the ranking comprehensive score to form a ranking list of user marketing content recommendations.

[0071] The beneficial effects of the present invention are:

[0072] (1) The present invention generates a multi-hop interactive chain based on the user's historical behavior and combines it with the interactive chain thinking semantic reasoning process to accurately infer the user's real interests and intentions, effectively improving the semantic matching accuracy between recommended content and user needs, and enhancing the accuracy of marketing content recommendations on e-commerce platforms and user experience.

[0073] (2) The present invention introduces the Secretary Bird optimization algorithm to jointly construct the initial search space with the initial sorting order of candidate content, historical conversion rate prediction and content timeliness index, which significantly improves the optimization efficiency and accuracy of the recommended content sorting weight, and shows better adaptability and sorting stability in the real-time dynamic change of recommendation scenarios on e-commerce platforms.

[0074] (3) In terms of the stability of candidate content ranking, the present invention effectively solves the problem of poor ranking stability in the existing technology through a nonlinear dynamic mapping strategy and a ranking stability index monitoring mechanism, breaks through the bottleneck of the traditional ranking method's insufficient response to user interest fluctuations, and realizes real-time and accurate adjustment of the recommended ranking position, thereby effectively improving user satisfaction with marketing content recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0076] Figure 1 This is a flowchart of the method for matching marketing content on an e-commerce platform based on semantic reasoning proposed by the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0078] refer to Figure 1 ,The e-commerce platform marketing content matching method based on semantic reasoning includes the following steps:

[0079] S1. Based on the user's historical behavior sequence on the e-commerce platform, the continuous multi-hop interaction behavior path between the user and the product is traced step by step to generate a multi-hop interaction chain for chain reasoning;

[0080] S2. For each multi-hop interaction chain, the interactive behaviors in the chain are gradually masked and prompted, and the pre-trained language model that has undergone supervised fine-tuning and reinforcement training is input to perform interactive chain thinking semantic reasoning to generate a set of recommended candidate content associated with the current target marketing content;

[0081] S3. For each candidate content in the recommended candidate content set, respectively calculate the matching degree score between the candidate content and the context semantics obtained by the interactive chain thinking semantic reasoning, and determine the initial ranking order of the candidate content;

[0082] S4. Build the initial search space for the Secretary Bird optimization algorithm based on the initial ranking order of candidate content, historical conversion rate prediction indicators, and content timeliness indicators. Simulating the Secretary Bird's exploration and jumping mechanism and local adjustment mechanism to avoid danger during hunting, it iteratively updates the ranking weight of each candidate content generation by generation until the preset convergence condition is reached.

[0083] S5. Re-adjust the ranking of candidate content using the optimized candidate content ranking weights to obtain a final ranking list for user marketing content recommendation;

[0084] S6. Collect user interactive feedback data on recommended marketing content in real time, and continuously update the chain semantic reasoning strategy of the pre-trained language model and the search variable parameters of the Secretary Bird optimization algorithm.

[0085] By combining interactive chain semantic reasoning with the Secretary Bird optimization algorithm, we can deeply explore and accurately capture the multi-hop interaction chains in user historical behavior data, effectively improving the semantic matching accuracy between marketing content and user needs; at the same time, by dynamically iteratively updating the sorting weights and optimizing parameter settings in real time based on user feedback, we can achieve the accuracy and stability of marketing content sorting results, thereby improving the user experience and recommendation efficiency of marketing content recommendations on e-commerce platforms.

[0086] In this embodiment, S1 specifically includes:

[0087] S11. Collecting the user's historical behavior sequence on the e-commerce platform, the historical behavior sequence includes the user's stay time on each product page, the interval between consecutive clicks, the page browsing depth, and the timestamps corresponding to the add-to-cart and favorites actions;

[0088] S12. Based on the collected historical behavior sequences, analyze the cross-category migration patterns of users between different product categories, identify the cross-category behavior association paths exhibited by users across multiple product categories, and record the cross-category association strength of each path;

[0089] S13. Calculate the comprehensive interaction strength of each interaction sequence in the historical behavior sequence using the cross-category association strength, and based on the comprehensive interaction strength, filter out interaction sequences with comprehensive interaction strengths below a preset threshold from the historical behavior sequence, and retain interaction sequences with comprehensive interaction strengths greater than or equal to the preset threshold;

[0090] S14. Based on the filtered interaction sequence, a reverse, step-by-step backtracking method based on the causal chain of user behavior is used. Starting from the target product, the interaction sequence between the user and historical products is reversed step by step to verify the logical rationality of the causal interactions between the interaction sequences step by step, and invalid interaction sequences introduced by non-active user behavior due to random or accidental touches are eliminated;

[0091] S15. Based on the interaction sequences retained after reverse retrospective verification, optimize and adjust the interaction sequences using cross-category association strength and user behavior stability indicators. Select the interaction sequence with the highest semantic relevance and optimal behavior logic for the target product recommendation in the historical behavior sequence to form an optimized multi-hop interaction behavior path. The user behavior stability indicator is a statistical indicator of the number of repeated interactions, the duration of a single interaction, and the duration between multiple interactions during the user's interaction with the same product or products of the same category.

[0092] S16. Based on the optimized multi-hop interaction behavior path, the interaction sequence is reorganized by combining the long-term preference characteristics and short-term interest fluctuation trends in the historical behavior sequence. The multi-hop interaction chain is constructed with the time continuity, behavior category alternation and semantic theme consistency of the interaction sequence as constraints.

[0093] By meticulously collecting and analyzing users' multi-dimensional historical behavior data on e-commerce platforms, deeply exploring the migration patterns of cross-category products, and using cross-category association strength and comprehensive interaction strength to accurately filter user behavior sequences, and combining reverse backtracking verification and stability indicators to implement interaction sequence optimization, the problems of insufficient semantic understanding and insufficient utilization of interaction data in existing technologies are effectively solved, thereby significantly improving the logical rationality and semantic association accuracy of multi-hop interaction chains, and enhancing the pertinence and effectiveness of marketing content recommendations.

[0094] In this embodiment, S2 specifically includes:

[0095] S21. Selecting the interaction sequences in each multi-hop interaction chain in sequence, and calculating the historical interest reduction weight of each node according to the time decay coefficient of the interaction sequence node in the historical behavior sequence in the order from the historical interaction node to the target node, and gradually implementing mask prompts for each product node in the sequence based on the historical interest reduction weights to form a progressive mask prompt sequence;

[0096] S21. Select the interaction sequences in each multi-hop interaction chain in sequence, and calculate the corresponding historical interest reduction weight of each interaction sequence according to the time decay coefficient of each interaction sequence in the historical behavior sequence in the order from the historical behavior sequence to the target product. Then, mask prompts are gradually implemented on the interaction sequences based on the historical interest reduction weights to form a progressive mask prompt sequence.

[0097] S22. Gradually input the progressive mask prompt sequence into the pre-trained language model that has undergone supervised fine-tuning and reinforcement training, dynamically weight the attention mechanism of the pre-trained language model using the historical interest decreasing weight, and predict the semantic information of the masked product node by node;

[0098] S23, real-time monitoring of the cosine similarity change trend of the semantic vectors of the predicted semantics of the masked products in each progressive mask prompt sequence during the pre-trained language model prediction process, using the fluctuation amplitude of the cosine similarity change trend as a semantic fluctuation index, and dynamically increasing the number of mask prompt nodes when the semantic fluctuation index does not reach a preset stability threshold until the semantic fluctuation index reaches or exceeds the stability threshold;

[0099] S24. Based on the stable progressive mask prompt sequence, analyze the cosine similarity between the predicted semantic vector of the masked product and the semantic vector of the target product by backtracking in reverse order, and quantitatively calculate the logical consistency score between each masked product and the target product;

[0100] S25. Based on the logical consistency score, a logical consistency threshold is set, and recommendation candidate content having a logical consistency score exceeding the logical consistency threshold is matched and extracted from the historical marketing content library to generate a set of matching recommendation candidate content;

[0101] S26, and store the recommended candidate content set and the predicted semantic information of all masked products in the corresponding chain thinking semantic reasoning process in association with each other.

[0102] By introducing a progressive mask prompt mechanism, combining the historical interest decreasing weight to dynamically adjust the predicted attention of the language model, and real-time monitoring and stable prediction of semantic fluctuation trends, the ability to accurately predict product semantic information in the chain semantic reasoning process is effectively improved; at the same time, by quantitatively calculating the logical consistency score to ensure the semantic logical consistency between the recommended candidate content and the target product, the matching accuracy of the recommended content and user needs and the rationality of the recommendation results are significantly improved.

[0103] In this embodiment, the pre-trained language model specifically includes a multi-granularity product embedding module, an interactive behavior temporal perception module, a bidirectional chain attention mechanism module, a mask prompt adaptive control module, a local semantic difference amplification module, a semantic path memory and retrieval module, and a self-enhanced feedback fine-tuning module:

[0104] The multi-granularity product embedding module receives a progressive mask prompt sequence and generates a multi-granularity product semantic vector by integrating a contextual dynamic weighting mechanism;

[0105] The interaction behavior temporal sequence perception module uses a continuous value encoded time decay function to calculate the temporal decay coefficient of each interaction sequence based on the timestamp corresponding to the interaction behavior to characterize the temporal decay relationship of the user's historical behavior;

[0106] The bidirectional chain attention mechanism module calculates a bidirectional chain semantic association representation of the interaction sequence based on the bidirectional path from the historical behavior sequence to the target product and from the target product to the historical behavior sequence based on the decreasing weight of the historical interest;

[0107] The mask prompt adaptive control module calculates the fluctuation amplitude of the cosine similarity of the predicted semantic vectors of the masked products in the progressive mask prompt sequence in real time, and dynamically adjusts the number of masked products based on the fluctuation amplitude until the semantic fluctuation amplitude reaches a stability threshold;

[0108] The local semantic difference amplification module amplifies the local differences between the predicted semantic vectors through nonlinear attention expansion based on the stable predicted semantic vectors output by the mask prompt adaptive control module;

[0109] The semantic path memory and retrieval module stores the predicted semantic vectors in the chain thinking semantic reasoning process in real time and retrieves the predicted semantic vectors based on the attention addressing method;

[0110] The self-enhanced feedback fine-tuning module receives user interactive feedback data on recommended marketing content in real time, uses the logical consistency score of the predicted semantic vector as a reward function, dynamically integrates supervised learning and reinforcement learning strategies, and continuously updates the network parameters of the pre-trained language model.

[0111] By constructing a pre-trained language model with multi-granularity product embedding, interactive behavior temporal perception and bidirectional chain attention mechanism, the mask prompt process and semantic prediction stability are dynamically adjusted in real time to effectively capture the changing trends of users' long-term preferences and short-term interests; at the same time, through semantic path memory and retrieval, and self-enhanced feedback fine-tuning mechanism, the language model parameters are dynamically updated using user feedback data, which significantly enhances the model's perception and response speed to the user's true intentions, thereby improving the real-time and accuracy of marketing content recommendations.

[0112] In this embodiment, the S23 specifically includes:

[0113] S231, real-time monitoring of the semantic vector cosine similarity fluctuation trend of the predicted semantic vector of the masked product in each progressive mask prompt sequence, and calculating the semantic entropy fluctuation, semantic direction deviation speed, and semantic aggregation change rate to form a multi-dimensional semantic fluctuation indicator set;

[0114] S232. Constructing a multidimensional semantic fluctuation tensor based on the multidimensional semantic fluctuation index set, and mapping the fluctuation index of each dimension to a unified high-dimensional space;

[0115] S233. For the multidimensional semantic fluctuation tensor, a dynamic attention weight fusion mechanism is adopted to update the fusion weight of each semantic fluctuation dimension in real time with the prediction stability index as the objective function, and generate a real-time updated fusion fluctuation trend vector;

[0116] S234, based on the real-time updated fusion fluctuation trend vector, a nonlinear adaptive threshold control strategy is used to determine the number of mask prompt nodes and the mask strength of the current mask prompt sequence;

[0117] S235 , repeating steps S231 to S234 until the fused fluctuation trend vector is lower than or equal to a preset stability threshold, and outputting a stable progressive mask prompt sequence after adaptive adjustment.

[0118] By real-time monitoring and quantifying the multidimensional fluctuation trend of semantic vectors, a unified multidimensional semantic fluctuation high-dimensional space is established, and the dynamic attention mechanism is used to fuse the multidimensional fluctuation trend information, effectively and dynamically adjusting the number of nodes and the prompt strength of the mask prompt, thereby ensuring the semantic prediction stability of the progressive mask prompt process and improving the reliability and consistency of semantic prediction in the chain semantic reasoning process.

[0119] In this embodiment, S3 specifically includes:

[0120] S31, extracting the product title text, product attribute feature data, and user comment text corresponding to each candidate content in the recommended candidate content set, and generating a multi-granularity semantic vector for each candidate content based on the multi-granularity product embedding module;

[0121] S32. Using the target product prediction semantic vector generated during the chain thinking semantic reasoning process as the central reference vector, a nonlinear radial basis mapping function is used to map the multi-granularity semantic vector of each candidate content into a two-dimensional polar coordinate semantic space centered on the target product prediction semantic vector.

[0122] S33. Calculate the radial distance parameter R of each candidate content in the two-dimensional polar coordinate semantic space:

[0123]

[0124] λ i =1+β·|X i -Y i | γ ;

[0125] Among them, R is the radial distance parameter, which represents the semantic distance between the candidate content and the target product, X i is the i-th dimension feature value of the candidate content multi-granularity semantic vector, Y i Predict the i-th dimension eigenvalue of the semantic vector for the target product, λ iis the nonlinear distance weighting factor of the i-th dimension feature, β is the distance amplification adjustment coefficient, which is used to control the prominence of the key dimension, γ is the nonlinear amplification index, and n is the dimension of the semantic vector;

[0126] The formula is obtained by introducing a nonlinear distance weighting factor λ i , so that the differences in each dimension between the semantic vector of the candidate content and the predicted semantic vector of the target product are nonlinearly amplified, thereby highlighting the key difference dimensions in semantic matching and achieving accurate quantification of the semantic distance between the candidate content and the target product; based on the Euclidean distance, with the help of the adjustment coefficient β and the amplification exponent γ, the weight of the important dimension differences is nonlinearly enhanced, strengthening the accurate expression of the degree of matching between the recommended content and the user's interests.

[0127] S34. Calculate the semantic offset angle parameter θ of each candidate content:

[0128]

[0129] Among them, θ is the semantic offset angle parameter, ω i is the direction-sensitive weight of the i-th dimension feature, δ is the semantic direction sensitivity adjustment coefficient, and ∈ is a very small positive number to prevent the denominator from being zero;

[0130] The formula accurately quantifies the degree of semantic direction deviation of candidate content relative to the target product by calculating the weighted cosine similarity between the multi-granularity semantic vector of the candidate content and the predicted semantic vector of the target product; it introduces a direction-sensitive weight factor to dynamically adjust the sensitivity to semantic differences on key dimensions, thereby effectively highlighting the semantic direction differences between the recommended content and the user's target intention, and improving the precision and accuracy of semantic matching evaluation.

[0131] S35. Perform nonlinear fusion of the radial distance parameter R and the semantic offset angle parameter θ to calculate the comprehensive semantic matching score P of each candidate content:

[0132]

[0133] Among them, P is the comprehensive semantic matching score, which integrates the characteristics of semantic distance and semantic offset direction, α is the distance attenuation adjustment coefficient, μ is the angle sensitivity enhancement index, R is the radial distance parameter, and θ is the semantic offset angle parameter;

[0134] The formula comprehensively quantifies the comprehensive matching degree between candidate content and target products in terms of semantic distance and semantic direction through nonlinear fusion of radial distance parameters and semantic offset angle parameters; uses an exponential function to attenuate the semantic distance, and uses the angle sensitivity enhancement index to amplify the impact of semantic direction differences, thereby more carefully measuring the overall semantic matching effect between candidate content and target products, and improving the accuracy and rationality of the recommendation result ranking.

[0135] S36. Determine an initial sequence of comprehensive semantic matching scores P for each candidate content in the recommended candidate content set, and sort them in descending order to form an initial sort order for the candidate content set.

[0136] S37. The position coordinate data of the candidate content in the two-dimensional polar coordinate semantic space is associated with the corresponding comprehensive semantic matching degree score P and stored to provide a visual explanation basis for interactive chain thinking semantic reasoning.

[0137] By constructing a two-dimensional polar coordinate semantic space centered on the predicted semantic vector of the target product, the radial distance parameter and semantic offset angle parameter are used to accurately quantify the semantic distance and directional deviation between the recommended candidate content and the target product, and further nonlinear fusion is used to calculate the comprehensive semantic matching degree score, a more detailed and accurate semantic matching evaluation is achieved, which effectively improves the accuracy and rationality of the recommended content ranking, thereby significantly enhancing users' acceptance of the recommended content and the recommendation effect.

[0138] In this embodiment, the S4 specifically includes:

[0139] S41. Convert the initial ranking order of the candidate content, the historical conversion rate prediction index, and the content timeliness index into corresponding ranking dimensions, conversion rate dimensions, and timeliness dimensions, respectively, and combine them in a unified space to construct an initial search space with the candidate content weight combination as the spatial position. The historical conversion rate prediction index is based on the user's historical behavior data on the e-commerce platform, and is obtained by statistically analyzing the user's historical conversion behavior for similar candidate content or products of the same category to obtain the predicted conversion probability for each candidate content. The content timeliness index is a comprehensive value obtained by quantitatively analyzing the interval length from the candidate content's publication to the current time, the popularity decay rate of the content topic, and the real-time attention change trend of the product or marketing topic involved in the candidate content;

[0140] S42. In the initial search space, based on the density of the candidate content, a Gaussian kernel smoothing method is used to calculate the density value of each position in the space, and the area with higher density in the space is determined as the priority exploration area for the search;

[0141] S43 simulates the density gradient-guided jump exploration mechanism used by secretary birds during hunting. The probability distribution of the next exploration location is determined based on the density gradient value in space. Locations with larger density gradients have higher exploration probabilities. At each iteration, a new search location is determined non-uniformly based on the probability, and random perturbations are introduced in each jump.

[0142] S44. For each determined search position, calculate the ranking performance fitness score of the candidate content combination, and based on the ranking performance fitness score, construct a fitness field that describes the ranking performance fitness change trend. Use the fitness gradient information in the fitness field to determine the search direction with the greatest potential for improving ranking performance fitness.

[0143] S45. Simulate the local adjustment mechanism used by secretary birds to avoid danger during hunting. Identify locations in the fitness field where the sorting performance fitness gradient fluctuates dramatically as high-risk areas. Determine the safe distance from the high-risk areas based on the degree of fitness gradient fluctuation at each search location. Adjust the spatial position of each search location in real time to actively avoid high-risk areas and move toward areas where the sorting performance fitness gradient fluctuates smoothly.

[0144] S46. Loop execution based on the density gradient-guided jump exploration mechanism and the local adjustment mechanism for avoiding high-risk areas, and update the candidate content sorting weight combination in real time until the sorting performance fitness gradient change of two consecutive iterations is less than or equal to the preset gradient convergence threshold, and determine the optimal sorting weight when the Secretary Bird optimization algorithm finally converges.

[0145] By constructing an initial search space with the initial ranking of candidate content, historical conversion rate prediction and content timeliness as dimensions, and using the Secretary Bird optimization algorithm with density gradient-guided exploration and fitness gradient risk avoidance, the candidate content ranking weights are dynamically and accurately optimized, which significantly improves the problems of existing recommendation methods that are prone to falling into local optimality and lack of adaptability, thereby effectively improving the ranking rationality, optimization accuracy and adaptability of marketing content recommendations to dynamic user needs.

[0146] In this embodiment, the S5 specifically includes:

[0147] S51, mapping the optimized candidate content ranking weights and the initial ranking order of the candidate content to the same ranking weight adjustment space, and determining the re-ranking position of the candidate content through a nonlinear dynamic mapping strategy;

[0148] S52. Establish a ranking stability monitoring mechanism for the re-ranked candidate content. By continuously tracking the position change trajectory of the candidate content in the ranking weight adjustment space, the consistency of the ranking position change magnitude and ranking change direction of each candidate content over multiple consecutive adjustment cycles is calculated to obtain a ranking stability index.

[0149] S53. For candidate content whose ranking stability index is lower than a preset stability threshold, the multi-hop interaction chain associated with the target product in the chain thinking semantic reasoning process is traced back step by step, and the continuous semantic state change sequence of the multi-hop interaction chain is gradually extracted. The difference change trend between the continuous semantic state change sequence and the multi-granularity semantic vector of the corresponding candidate content is analyzed to locate the fundamental semantic cause of the insufficient ranking stability of the candidate content;

[0150] S54. Based on the fundamental semantic reason, a dynamic semantic difference compensation mechanism is used to calculate the semantic difference compensation value between the candidate content and the target product in real time. The semantic difference compensation value is used to locally fine-tune the ranking weight of the candidate content to eliminate the inconsistent semantics between the multi-hop interaction chain and the candidate content.

[0151] S55. Recalculate the ranking comprehensive score of each candidate content based on the ranking weight after implementing the semantic difference compensation mechanism, and re-rank the candidate content based on the ranking comprehensive score to form a ranking list of user marketing content recommendations.

[0152] By introducing a nonlinear dynamic mapping strategy to accurately adjust the ranking position of candidate content, and combining the ranking stability monitoring mechanism with chain thinking semantic backtracking analysis, the fundamental semantic reasons for the insufficient stability of candidate content ranking are identified and compensated in real time, which significantly improves the stability and logical consistency of candidate content ranking, effectively solves the problem that existing recommendation ranking methods are insufficiently responsive to changes in user interests, and improves the stability of marketing content recommendation results and user satisfaction.

[0153] Example 1:

[0154] To verify the feasibility of the present invention in practice, the present invention was applied to the personalized marketing content recommendation task of a well-known e-commerce platform. Based on the actual historical user behavior data of the platform, accurate matching recommendations for marketing content were carried out. In this application scenario, traditional recommendation methods mainly include collaborative filtering methods, content-based methods, and simple hybrid methods. These methods generally have shortcomings in understanding user interests, manifested in low matching accuracy between recommended content and users' actual needs, lower-than-expected user click-through rates, and especially a lack of effective capture of dynamic changes in user interests. For example, in traditional recommendation algorithms, historical user behavior sequences are generally treated as simple click or purchase sequences, without a deep understanding of the chain logic behind the behavior, resulting in the recommended marketing content failing to effectively match user interests. According to platform data statistics, the recommendation click-through rate using traditional collaborative filtering methods is generally less than 8.5%, while the actual user conversion rate of content-based recommendation methods, despite incorporating product description features, is generally no higher than 2.7%.

[0155] To address this issue, the present invention employs the following specific implementation methods: First, based on the historical behavior sequences of 200,000 randomly selected active users from the platform over the past 90 days, including the duration of users' stay on each product page, the interval between consecutive clicks, the depth of page views, and the timestamps corresponding to add-to-cart and save-to-favorite actions, a continuous multi-hop interaction behavior path was constructed. These paths reflect the shifting patterns of user interests, such as a user starting out by searching for electronic products, gradually shifting to digital accessories, and finally becoming interested in a specific brand of headphones.

[0156] Next, an interactive chain-thinking semantic reasoning framework is constructed based on the multi-hop interaction path. The product sequence in the user's historical behavior path is gradually masked. Semantic reasoning is performed using a pre-trained language model to determine the user's current semantic interests. This is then used to match and recommend a set of candidate content. This embodiment then comprehensively evaluates the degree of match between each candidate content and the user's interests using parameters such as radial distance and semantic offset angle, resulting in a preliminary ranked list of recommendations.

[0157] Afterwards, the Secretary Bird optimization algorithm was further adopted to establish a search space with initial sorting, historical conversion rate indicators and timeliness indicators as dimensions, simulating the Secretary Bird's density gradient-guided jump exploration and danger avoidance mechanism. Through 22 iterations, the optimized combination scheme of sorting weights was automatically determined, and the fitness index of sorting performance was gradually improved from the initial value of 0.73 to above 0.91, achieving significant optimization of the candidate content sorting.

[0158] Subsequently, the system performs nonlinear dynamic mapping on the optimized sorting weights to form the final sorting adjustment strategy, and further implements dynamic semantic difference compensation for candidate content with low sorting stability, and finally outputs a user marketing recommendation list with stable sorting and accurate semantics.

[0159] The results of the embodiment show that the present invention has achieved significant improvement in the actual test of the above platform. As shown in the following table, the user recommendation click rate and conversion rate of 5 typical product categories were selected for comparative analysis:

[0160] Table 1 Comparison of recommended effects between the method of the present invention and the traditional method

[0161]

[0162]

[0163] Table 1 clearly shows that the proposed method significantly improves the click-through rate (CTR) of recommendations. For digital products, for example, the CTR increased from 8.32% using the traditional method to 17.85%, and the conversion rate increased from 2.14% to 7.96%, representing improvements of 114.54% and 271.96%, respectively. Furthermore, other categories, such as beauty and personal care, saw CTR and conversion rates increase by 124.41% and 243.15%, respectively, demonstrating the proposed method's superiority in accurately matching marketing content with user interests.

[0164] Further analysis revealed that the advantages of this invention primarily stem from its ability to accurately capture and understand the dynamic changes in user interests. For example, a user who previously focused on photography equipment has recently shown a significant increase in interest in live streaming-related equipment. Traditional methods would still recommend general photography products to this user, but this method accurately captures this shift in user interests and promptly adjusts recommendations to related products such as mobile phone live streaming accessories and sound cards, thereby increasing the user's willingness to click and purchase.

[0165] At the same time, a satisfaction survey of 1,000 randomly selected users showed that users generally reflected that the recommended content of the present invention was more in line with their actual interests and needs. The average user satisfaction score was 4.62 points (out of 5 points), which was significantly higher than the average score of 3.14 points of the traditional method. This also reflects the significant improvement of the present invention in user experience.

[0166] In summary, this embodiment verifies the technical advantages and effect improvement of the method of the present invention in practical applications with detailed data and specific application results, indicating that the present invention has good implementation feasibility and practical application value in the task of personalized recommendation of e-commerce marketing content.

[0167] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The e-commerce platform marketing content matching method based on semantic reasoning is characterized by: The following steps are involved: S1. Based on the user's historical behavior sequence on the e-commerce platform, the continuous multi-hop interaction behavior path between the user and the product is traced step by step to generate a multi-hop interaction chain for chain reasoning; S2. For each multi-hop interaction chain, the interactive behaviors in the chain are gradually masked and prompted, and the pre-trained language model that has undergone supervised fine-tuning and reinforcement training is input to perform interactive chain thinking semantic reasoning to generate a set of recommended candidate content associated with the current target marketing content; S3. For each candidate content in the recommended candidate content set, respectively calculate the matching degree score between the candidate content and the context semantics obtained by the interactive chain thinking semantic reasoning, and determine the initial ranking order of the candidate content; S4. Build the initial search space for the Secretary Bird optimization algorithm based on the initial ranking order of candidate content, historical conversion rate prediction indicators, and content timeliness indicators. Simulating the Secretary Bird's exploration and jumping mechanism and local adjustment mechanism to avoid danger during hunting, it iteratively updates the ranking weight of each candidate content generation by generation until the preset convergence condition is reached. S5. Re-adjust the ranking of the candidate content using the optimized candidate content ranking weights to obtain a final ranking list for user marketing content recommendation; S6. Collect user interactive feedback data on recommended marketing content in real time, and continuously update the chain semantic reasoning strategy of the pre-trained language model and the search variable parameters of the Secretary Bird optimization algorithm.

2. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 1 is characterized in that: Said S1 specifically includes: S11. Collecting the user's historical behavior sequence on the e-commerce platform, the historical behavior sequence includes the user's stay time on each product page, the interval between consecutive clicks, the page browsing depth, and the timestamps corresponding to the add-to-cart and favorites actions; S12. Based on the collected historical behavior sequences, analyze the cross-category migration patterns of users between different product categories, identify the cross-category behavior association paths exhibited by users across multiple product categories, and record the cross-category association strength of each path; S13. Calculate the comprehensive interaction strength of each interaction sequence in the historical behavior sequence using the cross-category association strength, and based on the comprehensive interaction strength, filter out interaction sequences with comprehensive interaction strengths below a preset threshold from the historical behavior sequence, and retain interaction sequences with comprehensive interaction strengths greater than or equal to the preset threshold; S14. Based on the filtered interaction sequence, a reverse, step-by-step backtracking method based on the causal chain of user behavior is used. Starting from the target product, the interaction sequence between the user and historical products is reversed step by step to verify the logical rationality of the causal interactions between the interaction sequences step by step, and invalid interaction sequences introduced by non-active user behavior due to random or accidental touches are eliminated; S15. Based on the interaction sequences retained after reverse retrospective verification, optimize and adjust the interaction sequences using cross-category association strength and user behavior stability indicators. Select the interaction sequence with the highest semantic relevance and optimal behavior logic for the target product recommendation in the historical behavior sequence to form an optimized multi-hop interaction behavior path. The user behavior stability indicator is a statistical indicator of the number of repeated interactions, the duration of a single interaction, and the duration between multiple interactions during the user's interaction with the same product or products of the same category. S16. Based on the optimized multi-hop interaction behavior path, the interaction sequence is reorganized by combining the long-term preference characteristics and short-term interest fluctuation trends in the historical behavior sequence. The multi-hop interaction chain is constructed with the time continuity, behavior category alternation and semantic theme consistency of the interaction sequence as constraints.

3. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 1 is characterized in that: The S2 specifically includes: S21. Selecting the interaction sequences in each multi-hop interaction chain in sequence, and calculating the historical interest reduction weight of each node according to the time decay coefficient of the interaction sequence node in the historical behavior sequence in the order from the historical interaction node to the target node, and gradually implementing mask prompts for each product node in the sequence based on the historical interest reduction weights to form a progressive mask prompt sequence; S21. Select the interaction sequences in each multi-hop interaction chain in sequence, and calculate the corresponding historical interest reduction weight of each interaction sequence according to the time decay coefficient of each interaction sequence in the historical behavior sequence in the order from the historical behavior sequence to the target product. Then, mask prompts are gradually implemented on the interaction sequences based on the historical interest reduction weights to form a progressive mask prompt sequence. S22. Gradually input the progressive mask prompt sequence into the pre-trained language model that has undergone supervised fine-tuning and reinforcement training, dynamically weight the attention mechanism of the pre-trained language model using the historical interest decreasing weight, and predict the semantic information of the masked product node by node; S23, real-time monitoring of the cosine similarity change trend of the semantic vectors of the predicted semantics of the masked products in each progressive mask prompt sequence during the pre-trained language model prediction process, using the fluctuation amplitude of the cosine similarity change trend as a semantic fluctuation index, and dynamically increasing the number of mask prompt nodes when the semantic fluctuation index does not reach a preset stability threshold until the semantic fluctuation index reaches or exceeds the stability threshold; S24. Based on the stable progressive mask prompt sequence, analyze the cosine similarity between the predicted semantic vector of the masked product and the semantic vector of the target product by backtracking in reverse order, and quantitatively calculate the logical consistency score between each masked product and the target product; S25. Based on the logical consistency score, a logical consistency threshold is set, and recommendation candidate content having a logical consistency score exceeding the logical consistency threshold is matched and extracted from the historical marketing content library to generate a set of matching recommendation candidate content; S26, and store the recommended candidate content set and the predicted semantic information of all masked products in the corresponding chain thinking semantic reasoning process in association with each other.

4. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 3 is characterized in that: The pre-trained language model specifically includes a multi-granularity product embedding module, an interactive behavior temporal perception module, a bidirectional chain attention mechanism module, a mask prompt adaptive control module, a local semantic difference amplification module, a semantic path memory and retrieval module, and a self-enhanced feedback fine-tuning module: The multi-granularity product embedding module receives a progressive mask prompt sequence and generates a multi-granularity product semantic vector by integrating a contextual dynamic weighting mechanism; The interaction behavior temporal sequence perception module uses a continuous value encoded time decay function to calculate the temporal decay coefficient of each interaction sequence based on the timestamp corresponding to the interaction behavior to characterize the temporal decay relationship of the user's historical behavior; The bidirectional chain attention mechanism module calculates a bidirectional chain semantic association representation of the interaction sequence based on the bidirectional path from the historical behavior sequence to the target product and from the target product to the historical behavior sequence based on the decreasing weight of the historical interest; The mask prompt adaptive control module calculates the fluctuation amplitude of the cosine similarity of the predicted semantic vectors of the masked products in the progressive mask prompt sequence in real time, and dynamically adjusts the number of masked products based on the fluctuation amplitude until the semantic fluctuation amplitude reaches a stability threshold; The local semantic difference amplification module amplifies the local differences between the predicted semantic vectors through nonlinear attention expansion based on the stable predicted semantic vectors output by the mask prompt adaptive control module; The semantic path memory and retrieval module stores the predicted semantic vectors in the chain thinking semantic reasoning process in real time and retrieves the predicted semantic vectors based on the attention addressing method; The self-enhanced feedback fine-tuning module receives user interactive feedback data on recommended marketing content in real time, uses the logical consistency score of the predicted semantic vector as a reward function, dynamically integrates supervised learning and reinforcement learning strategies, and continuously updates the network parameters of the pre-trained language model.

5. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 3 is characterized in that: The S23 specifically includes: S231, real-time monitoring of the semantic vector cosine similarity fluctuation trend of the predicted semantic vector of the masked product in each progressive mask prompt sequence, and calculating the semantic entropy fluctuation, semantic direction deviation speed, and semantic aggregation change rate to form a multi-dimensional semantic fluctuation indicator set; S232. Constructing a multidimensional semantic fluctuation tensor based on the multidimensional semantic fluctuation index set, and mapping the fluctuation index of each dimension to a unified high-dimensional space; S233. For the multidimensional semantic fluctuation tensor, a dynamic attention weight fusion mechanism is adopted to update the fusion weight of each semantic fluctuation dimension in real time with the prediction stability index as the objective function, and generate a real-time updated fusion fluctuation trend vector; S234, based on the real-time updated fusion fluctuation trend vector, a nonlinear adaptive threshold control strategy is used to determine the number of mask prompt nodes and the mask strength of the current mask prompt sequence; S235 , repeating steps S231 to S234 until the fused fluctuation trend vector is lower than or equal to a preset stability threshold, and outputting a stable progressive mask prompt sequence after adaptive adjustment.

6. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 1 is characterized in that: The S3 specifically includes: S31, extracting the product title text, product attribute feature data, and user comment text corresponding to each candidate content in the recommended candidate content set, and generating a multi-granularity semantic vector for each candidate content based on the multi-granularity product embedding module; S32. Using the target product prediction semantic vector generated during the chain thinking semantic reasoning process as the central reference vector, a nonlinear radial basis mapping function is used to map the multi-granularity semantic vector of each candidate content into a two-dimensional polar coordinate semantic space centered on the target product prediction semantic vector. S33. Calculate the radial distance parameter R of each candidate content in the two-dimensional polar coordinate semantic space: l i =1+β·|X i -Y i | γ ; Among them, R is the radial distance parameter, which represents the semantic distance between the candidate content and the target product, X i is the i-th dimension feature value of the candidate content multi-granularity semantic vector, Y i Predict the i-th dimension eigenvalue of the semantic vector for the target product, λ i is the nonlinear distance weighting factor of the i-th dimension feature, β is the distance amplification adjustment coefficient, which is used to control the prominence of the key dimension, γ is the nonlinear amplification index, and n is the dimension of the semantic vector; S34. Calculate the semantic offset angle parameter θ of each candidate content: Among them, θ is the semantic offset angle parameter, ω i is the direction-sensitive weight of the i-th dimension feature, δ is the semantic direction sensitivity adjustment coefficient, and ∈ is a very small positive number to prevent the denominator from being zero; S35. Perform nonlinear fusion of the radial distance parameter R and the semantic offset angle parameter θ to calculate the comprehensive semantic matching score P for each candidate content: Among them, P is the comprehensive semantic matching score, which integrates the characteristics of semantic distance and semantic offset direction, α is the distance attenuation adjustment coefficient, μ is the angle sensitivity enhancement index, R is the radial distance parameter, and θ is the semantic offset angle parameter; S36. Determine an initial sequence of comprehensive semantic matching scores P for each candidate content in the recommended candidate content set, and sort them in descending order to form an initial sort order for the candidate content set. S37. The position coordinate data of the candidate content in the two-dimensional polar coordinate semantic space is associated with the corresponding comprehensive semantic matching degree score P and stored to provide a visual explanation basis for interactive chain thinking semantic reasoning.

7. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 1 is characterized in that: The S4 specifically includes: S41. Convert the initial ranking order of the candidate content, the historical conversion rate prediction index, and the content timeliness index into corresponding ranking dimensions, conversion rate dimensions, and timeliness dimensions, respectively, and combine them in a unified space to construct an initial search space with the candidate content weight combination as the spatial position. The historical conversion rate prediction index is based on the user's historical behavior data on the e-commerce platform, and is obtained by statistically analyzing the user's historical conversion behavior for similar candidate content or products of the same category to obtain the predicted conversion probability for each candidate content. The content timeliness index is a comprehensive value obtained by quantitatively analyzing the interval length from the candidate content's publication to the current time, the popularity decay rate of the content topic, and the real-time attention change trend of the product or marketing topic involved in the candidate content; S42. In the initial search space, based on the density of the candidate content, a Gaussian kernel smoothing method is used to calculate the density value of each position in the space, and the area with higher density in the space is determined as the priority exploration area for the search; S43 simulates the density gradient-guided jump exploration mechanism used by secretary birds during hunting. The probability distribution of the next exploration location is determined based on the density gradient value in space. Locations with larger density gradients have higher exploration probabilities. At each iteration, a new search location is determined non-uniformly based on the probability, and random perturbations are introduced in each jump. S44. For each determined search position, calculate the ranking performance fitness score of the candidate content combination, and based on the ranking performance fitness score, construct a fitness field that describes the ranking performance fitness change trend. Use the fitness gradient information in the fitness field to determine the search direction with the greatest potential for improving ranking performance fitness. S45. Simulate the local adjustment mechanism used by secretary birds to avoid danger during hunting. Identify locations in the fitness field where the sorting performance fitness gradient fluctuates dramatically as high-risk areas. Determine the safe distance from the high-risk areas based on the degree of fitness gradient fluctuation at each search location. Adjust the spatial position of each search location in real time to actively avoid high-risk areas and move toward areas where the sorting performance fitness gradient fluctuates smoothly. S46. Loop execution based on the density gradient-guided jump exploration mechanism and the local adjustment mechanism for avoiding high-risk areas, and update the candidate content sorting weight combination in real time until the sorting performance fitness gradient change of two consecutive iterations is less than or equal to the preset gradient convergence threshold, and determine the optimal sorting weight when the Secretary Bird optimization algorithm finally converges.

8. The e-commerce platform marketing content matching method based on semantic reasoning according to claim 1 is characterized in that: The S5 specifically includes: S51, mapping the optimized candidate content ranking weights and the initial ranking order of the candidate content to the same ranking weight adjustment space, and determining the re-ranking position of the candidate content through a nonlinear dynamic mapping strategy; S52. Establish a ranking stability monitoring mechanism for the re-ranked candidate content. By continuously tracking the position change trajectory of the candidate content in the ranking weight adjustment space, the consistency of the ranking position change magnitude and ranking change direction of each candidate content over multiple consecutive adjustment cycles is calculated to obtain a ranking stability index. S53. For candidate content whose ranking stability index is lower than a preset stability threshold, the multi-hop interaction chain associated with the target product in the chain thinking semantic reasoning process is traced back step by step, and the continuous semantic state change sequence of the multi-hop interaction chain is gradually extracted. The difference change trend between the continuous semantic state change sequence and the multi-granularity semantic vector of the corresponding candidate content is analyzed to locate the fundamental semantic cause of the insufficient ranking stability of the candidate content; S54. Based on the fundamental semantic reason, a dynamic semantic difference compensation mechanism is used to calculate the semantic difference compensation value between the candidate content and the target product in real time. The semantic difference compensation value is used to locally fine-tune the ranking weight of the candidate content to eliminate the inconsistent semantics between the multi-hop interaction chain and the candidate content. S55. Recalculate the ranking comprehensive score of each candidate content based on the ranking weight after implementing the semantic difference compensation mechanism, and re-rank the candidate content based on the ranking comprehensive score to form a ranking list of user marketing content recommendations.

Images