A product promotion type copy intelligent generation method and system

Abstract

The application discloses a product promotion type script intelligent generation method and system, and relates to the technical field of script intelligent generation.The product promotion type script intelligent generation method and system comprise the following steps: S1, collecting product script generation basic data, audience preference data and script generation hard constraint data and preprocessing; S2, generating a style constraint parameter set according to a collaborative filtering similar neighborhood discriminant analysis result; S3, outputting a feasible structure configuration and script text data according to a structure feasibility discriminant analysis result; S4, jointly evaluating information coverage, format rules, repeated stacking and word number overlimit, and performing directional repair and finalization signing and archiving.The product promotion type script intelligent generation method and system solve the problem that existing product promotion type script intelligent generation methods usually generate first and then cut and repair to meet the constraints of word number, format and forbidden words, but lack structure feasibility determination and consistency repair when multiple hard constraints coexist, which easily leads to the problem of unusable script structure degradation.

Description

Technical Field

[0001] This invention relates to the field of intelligent copywriting generation technology, specifically to a method and system for intelligently generating product promotion copy. Background Technology

[0002] Against the backdrop of rapid development in internet content production and commercial information dissemination, product promotion copywriting is widely used in e-commerce detail pages, feed ads, search engine optimization, social media seeding, and short video e-commerce to effectively convey product information, emphasize selling points, and reach the audience within limited display space. With the increase in the number of platform channels and the frequency of releases, copywriting production has gradually shifted from purely manual writing to a technological paradigm that combines template-based editing, rule-constrained generation, and generative model assistance. On the one hand, the industry has long relied on automated generation methods based on fixed-structure templates, keyword libraries, and rule engines to meet format specifications such as word count, layout, sensitive words, and compliance prompts. On the other hand, the development of natural language processing and generative artificial intelligence in recent years has driven the application of language model-based automatic writing and rewriting technologies, enabling systems to generate multiple candidate texts and perform filtering and optimization given product information and target audience. Simultaneously, recommendation systems and user profiling technologies are also frequently used to assist in determining copywriting style and expression strategies. Overall, existing intelligent generation technology for product promotion copywriting shows a development trend that combines "generation model output - rule constraint adaptation - quality verification and iterative optimization", and continues to evolve in terms of multi-channel standard adaptation, large-scale content production and automated evaluation.

[0003] For example, the invention patent with announcement number CN120746646B discloses a method, device, and storage medium for automatically generating marketing copy based on reinforcement learning. The method includes performing semantic matching retrieval on publicly available copy data to obtain candidate copy related to product information; constructing slotted rewriting instructions containing candidate copy, product information, and marketing points and inputting them into a pre-generated language model; filling in the corresponding information in preset slot positions to generate marketing copy; further inputting the generated copy into a rule-based evaluation system to obtain evaluation data; constructing partial-order training samples based on the evaluation data as reward signals to perform reinforcement learning training on the model; and finally calling the trained model to output the target marketing copy. This achieves closed-loop optimization of retrieval, slotted generation, and evaluation reinforcement, thereby improving the adaptability and controllability of copy generation.

[0004] For example, the invention patent with publication number CN119782816A discloses a method and related apparatus for constructing training corpus for promotional copywriting. The method includes acquiring a first training corpus of promotional copywriting, extracting professional features and copywriting style features, and classifying the training corpus accordingly to determine multiple style tags; then generating indicator words based on the target style tags, corresponding training corpus, and professional knowledge, and inputting them into the generation model to generate training samples corresponding to the target style tags, which are used to train the promotional copywriting recognition model to identify the type of promoted product and the copywriting style of the copywriting. This realizes the structured construction and generation expansion of training corpus oriented towards style tags, thereby enhancing the corpus's ability to support the training of the recognition model.

[0005] Existing intelligent copy generation technologies for product promotion rely on "generating first and then trimming and repairing" to meet hard constraints such as word count, fixed structure, required words and prohibited words. However, they lack a closed-loop mechanism for pre-generation judgment of constraint conflicts and structural feasibility, as well as post-generation slot-level consistency verification and targeted repair. This leads to structural degradation problems such as missing coverage information, duplication and stacking, format destruction and logical breaks when multiple hard constraints coexist, making it difficult to stably output usable copy.

[0006] Therefore, in response to the above problems, there is an urgent need for a method and system for intelligently generating product promotion copy. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent generation method and system for product promotion copy. It solves the problem that existing intelligent generation methods for product promotion copy typically generate the copy first and then trim and repair it to meet constraints such as word count, format, and prohibited words. However, when multiple hard constraints conflict, the method lacks structural feasibility assessment and consistency repair, which can easily lead to the degradation and unusability of the copy structure.

[0008] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent generation of product promotion copy, comprising: S1, collecting basic data for product copy generation, audience preference data, and hard constraint data for copy generation and preprocessing them to generate a user group-copy preference intensity matrix and a structure slot table; S2, performing collaborative filtering similarity neighborhood discriminant analysis on the user group-copy preference intensity matrix, and generating a set of style constraint parameters based on the results of the collaborative filtering similarity neighborhood discriminant analysis; S3, performing structural feasibility discriminant analysis on the structure slot table and the hard constraint data for copy generation, and outputting feasible structural configurations and copy text data based on the results of the structural feasibility discriminant analysis; S4, jointly evaluating the information coverage sufficiency, format rule satisfaction sufficiency, repetition and stacking penalty, and word count exceeding penalty on the copy text data and the hard constraint data for copy generation, and performing targeted repair and finalization and archiving.

[0009] Furthermore, the specific process of collecting basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation is as follows: Collecting basic data for product copywriting generation, which includes: product name data, product category data, product selling point data, product evidence data, set of available product keywords, and set of prohibited product words; collecting audience preference data, which includes: target audience group identification data, target audience profile tag data, historical promotional copy text data, historical promotional copy structure type annotation data, and historical preference interaction behavior data; collecting hard constraint data for copywriting generation, which includes: maximum word count data, fixed structure type data, list of information covered by the copy, set of required words in the copy, set of prohibited words in the copy, and copy format rules data.

[0010] Furthermore, the specific process of generating the user group-copywriting preference intensity matrix and structure slot table through preprocessing is as follows: First, the product selling point data, product evidence description data, and historical promotional copy text data are cleaned and formatted using a unified character normalization and punctuation rule correction algorithm. Second, phrases are extracted from the product selling point data and product evidence description data using word segmentation and entity phrase retention algorithms, and the extracted phrases are used as the matching dictionary input for keyword matching and semantic similarity alignment algorithms. Third, the product prohibited word set is processed using an automatic prohibited word indexing and mandatory word alignment algorithm. A fast matching index is established based on the data, the set of prohibited words in the copywriting, and the set of required words in the copywriting. Extremum removal and time window aggregation algorithms are used to process historical preference interaction data and perform time window aggregation. A distribution standardization algorithm is used to unify the scale of historical preference interaction data. A linear normalization algorithm is used to map the distribution-standardized historical preference interaction data to a unified interval and construct a user group-copywriting preference intensity matrix. A structure slot table is generated from copywriting fixed structure type data and copywriting format rule data using structure parsing and slot mapping algorithms.

[0011] Furthermore, the specific process of collaborative filtering and similarity neighborhood discrimination analysis on the user group-copywriting preference intensity matrix is ​​as follows: Similar neighborhoods are obtained on the user group-copywriting preference intensity matrix using a cosine similarity nearest neighbor retrieval algorithm based on the target audience identification data. The candidate set of similar neighborhoods is filtered for consistency using the target audience profile tag data, and the number of neighborhoods meeting the similarity threshold is counted to obtain the similar neighborhood count. The cosine similarity of each neighborhood within a similar neighborhood is averaged to obtain the neighborhood similarity mean. The number of historical promotional copywriting structure types labeled within the similar neighborhood is also calculated. The probability distribution is used to calculate the neighborhood preference dispersion using the Shannon entropy algorithm; the sum of the similarity neighborhood count and constant 1 is calculated and the natural logarithm is taken to obtain the logarithmic similarity neighborhood count term; the product of the logarithmic similarity neighborhood count term and the mean of neighborhood similarity is calculated to obtain the neighborhood similarity modulation term; the sum of the neighborhood preference dispersion and constant 1 is calculated to obtain the dispersion smoothing term; the neighborhood similarity modulation term is divided by the dispersion smoothing term to obtain the exponential input term; the natural exponential function is calculated after taking the negative value of the exponential input term to obtain the exponential decay term; the audience style adaptation discriminant value is obtained by subtracting the exponential decay term from constant 1.

[0012] Furthermore, the specific process of generating a set of style constraint parameters based on the collaborative filtering similarity neighborhood discriminant analysis results is as follows: Real-time comparison of audience style fit discriminant values ​​and audience style fit discriminant thresholds: When the audience style fit discriminant value is greater than or equal to the audience style fit discriminant threshold, the target interval data for hot word usage intensity is obtained from historical promotional copy text data through hot word density statistics and quantile statistics mapping algorithms; the target interval data for sentence length is obtained from historical promotional copy text data through sentence segmentation and sentence length statistics and quantile statistics mapping algorithms; and the target interval data for information density is obtained from historical promotional copy text data combined with product selling point data and product evidence description data through coverage information matching density statistics and quantile statistics mapping algorithms. The target interval data for hot word usage intensity, the target interval data for sentence length, and the information density are then combined. The density target interval data is mapped to the slot-level constraint tables of the title slot and the body paragraph slot according to the structure slot table, and the style constraint parameter set is output. When the audience style fit discrimination value is less than the audience style fit discrimination threshold, the similar neighborhood expansion search is triggered and the audience style fit discrimination value is calculated a second time. If the result of the second calculation is still less than the audience style fit discrimination threshold, the general style constraint parameter set is obtained by combining the historical promotion copy text data with the historical promotion copy structure type annotation data through the quantile statistical mapping algorithm. The hot word usage intensity target interval data in the general style constraint parameter set is converged to the median interval of the historical distribution. The hot word usage intensity target interval data is obtained from the historical promotion copy text data through hot word density statistics and quantile statistical mapping algorithms, and then enters the structural feasibility discrimination analysis.

[0013] Furthermore, the specific process of performing structural feasibility discriminant analysis on the structural slot table and the hard constraint data for copywriting generation is as follows: Based on the structural slot table, the structural word count capacity value is obtained by using a slot-level word count budget allocation algorithm for the upper limit of copywriting word count data and the fixed structural type data; for each piece of information in the copywriting coverage information list data, the minimum expression length is obtained by using a shortest expressible sentence length estimation algorithm, and the minimum word count requirement value for the coverage information is obtained by summing these values; the set of prohibited words in copywriting, the set of required words in copywriting, the set of prohibited words for products, and the set of available keywords for products are analyzed by... The set conflict detection algorithm obtains the number of conflicting items, and combines this with the number of rules that cannot be simultaneously satisfied in the copywriting format rule data to obtain the constraint conflict degree; the minimum word count requirement value for covered information is calculated and summed with a constant to obtain the requirement smoothing term; the structural word count capacity value is divided by the requirement smoothing term to obtain the capacity requirement ratio term; the capacity requirement ratio term is calculated and summed with a constant to obtain the capacity requirement logarithm term; the constraint conflict degree is calculated and summed with a constant to obtain the conflict logarithm term; the capacity requirement logarithm term is subtracted from the conflict logarithm term and input into a logical function mapping to obtain the structural feasibility judgment value.

[0014] Furthermore, the specific process of outputting feasible structural configurations and copywriting text data based on the structural feasibility discrimination analysis results is as follows: Real-time comparison of structural feasibility discrimination values ​​and structural feasibility discrimination thresholds: When the structural feasibility discrimination value is less than the structural feasibility discrimination threshold, conflict items are decomposed and located based on the constraint conflict degree. Then, a structural reduction strategy and a coverage information compression and rearrangement strategy are executed. The structural reduction strategy involves switching the fixed structural type data of the copywriting to the fixed structural types of the title slot and key point slot. The fixed structural types of the title slot and key point slot are limited by the structural slot table generated by the structural parsing and slot mapping algorithm based on the fixed structural type data of the copywriting and the copywriting format rule data. The coverage information compression and rearrangement strategy involves compressing the coverage information expression budget based on the shortest expressible sentence length estimation algorithm while keeping the number of items in the copywriting coverage information list data unchanged. The number of items is equal to the number of items in the copywriting. The count of coverage information items in the coverage information list data is used to recalculate the structural feasibility judgment value until the structural feasibility judgment threshold is met. When the structural feasibility judgment value is greater than or equal to the structural feasibility judgment threshold, the upper limit of the word count data of the copywriting is allocated at the slot level based on the structural slot table, and the copywriting coverage information list data is bound to the corresponding slot to form a coverage information-slot binding table. At the same time, the style constraint parameter set is injected into the slot-level constraint table, and a feasible structural configuration is output. Based on the feasible structural configuration and the style constraint parameter set, the copywriting text data is generated. The generation is based on the structural slot table to construct a structural skeleton composed of title slots and body paragraph slots, and the copywriting coverage information list data is filled slot by slot according to the slot-level word count budget table and the coverage information-slot binding table. In the title slot, the product name data is inserted as a phrase with priority.

[0015] Furthermore, the specific process for jointly evaluating the information coverage sufficiency, format rule fulfillment sufficiency, repetition / stack penalty, and word count exceedance penalty of the generated copy text data and the hard constraint data is as follows: For the generated copy text data and the copy coverage information list data, the information coverage sufficiency is obtained by calculating the proportion of covered items using keyword matching and semantic similarity alignment algorithms; for the generated copy text data and the copy format rule data, the format rule fulfillment sufficiency is obtained by calculating the proportion of satisfied items using regular expression structure parsing and structure slot table alignment verification algorithms; for the generated copy text data, the longest common substring repetition / stacking / completion ... The duplication detection algorithm obtains the duplication ratio and maps it to obtain the duplication stacking penalty; the number of words in the generated text data and the upper limit of the text number are calculated to obtain the word count exceeding penalty; the product of the information coverage sufficiency and the format rule satisfaction sufficiency is calculated to obtain the coverage format joint term; the sum of the duplication stacking penalty and the word count exceeding penalty is calculated and a constant is added to obtain the penalty smoothing term; the coverage format joint term is divided by the penalty smoothing term to obtain the exponential input term; the exponential input term is negativeized and the natural exponential function is calculated to obtain the exponential decay term; the constant is calculated and the exponential decay term is subtracted to obtain the final draft consistency pass value.

[0016] Furthermore, the specific process of performing targeted repair and finalization signing and archiving is as follows: Real-time comparison of the finalization consistency pass value and the finalization consistency pass threshold: When the finalization consistency pass value is less than the finalization consistency pass threshold, targeted repair is performed: When at least one piece of copywriting coverage information list data in the generated copywriting text data is not matched by the keyword matching and semantic similarity alignment algorithm, the corresponding missing slot is filled in according to the coverage information-slot binding table; When the number of format rule violations obtained by the regular expression structure parsing and structure slot table alignment verification algorithm of the generated copywriting text data is not zero, the paragraph boundaries and line break rules are rearranged according to the structure slot table; When the number of words in the generated copywriting text data is greater than the upper limit of the number of words in the copywriting data or the longest common substring duplication detection algorithm identifies... When duplicate segments exist, compress the non-covered information expression according to the slot-level word count budget table and merge and rewrite the duplicate segments; after repair, recalculate the final draft consistency pass value. If it is still less than the final draft consistency pass threshold, fall back to the structural feasibility judgment analysis and switch the fixed structure type data of the copy according to the structural de-leveling strategy and regenerate the copy text data; when the final draft consistency pass value is greater than or equal to the final draft consistency pass threshold, perform full-text search verification of the product prohibited word set data and the copy prohibited word set data, and perform structural slot table alignment verification. Output the final promotion copy text data, the structure type identifier corresponding to the fixed structure type data of the copy, the coverage information coverage result data, the prohibited word hit verification result data, and the final draft issuance result data.

[0017] A second aspect of this invention provides an intelligent product promotion copywriting generation system, comprising: Beneficial effects The present invention has the following beneficial effects: (1) This invention, by introducing a structural feasibility judgment and feasible structural configuration output mechanism before generation, achieves the effect that the structure can be established a priori under multiple hard constraints, effectively solving the problem of structural degradation and unusability caused by prior generation and subsequent trimming in the prior art.

[0018] (2) This invention achieves the effect of calculable alignment between copywriting style and target audience expression preferences by generating a set of style constraint parameters based on audience preferences, effectively solving the problem that style selection in the prior art depends on experience and is difficult to reproduce consistently.

[0019] (3) This invention binds the coverage information to the structural slots and implements slot-level budget allocation, thereby achieving the effect of controllable bearing and orderly placement of the coverage information under a fixed number of characters and a fixed structure, effectively solving the problems of missing coverage information and chaotic paragraph content stacking in the prior art.

[0020] (4) This invention achieves the effect that the compliance constraints of the text can be stably met before the final draft is issued by performing full-text search verification on the set of prohibited words and triggering targeted repair or rollback and regeneration when the verification fails. This effectively solves the problem of semantic inconsistency and structural rule destruction caused by relying solely on generation and replacement in the prior art.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of a product promotion copy intelligent generation method according to the present invention; Figure 2 This is a structural diagram of a product promotion copywriting intelligent generation system according to the present invention; Figure 3 Flowchart for audience style adaptation judgment and constraint parameter generation in this invention; Figure 4 This is a scatter plot showing the relationship between the capacity requirement ratio and the structural feasibility discrimination value of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figures 1-4This invention provides a technical solution: an intelligent method for generating product promotion copy, comprising the following steps: S1, collecting basic data for product copy generation, audience preference data, and hard constraint data for copy generation, and preprocessing them to generate a user group-copy preference intensity matrix and a structure slot table; S2, performing collaborative filtering similarity neighborhood discriminant analysis on the user group-copy preference intensity matrix, and generating a set of style constraint parameters based on the results of the collaborative filtering similarity neighborhood discriminant analysis; S3, performing structural feasibility discriminant analysis on the structure slot table and the hard constraint data for copy generation, and outputting feasible structural configurations and copy text data based on the results of the structural feasibility discriminant analysis; S4, jointly evaluating the information coverage sufficiency, format rule satisfaction sufficiency, repetition and stacking penalty, and word count exceeding penalty on the copy text data and the hard constraint data for copy generation, and performing targeted repair and finalization and archiving.

[0025] Specifically, the process of collecting basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation is as follows: Basic data for product copywriting generation is collected, including: product name data, product category data, product selling point data, product evidence data, set of available product keywords, and set of prohibited product words. This data is obtained directly from product databases, product information tables, or compliance configuration tables. The set of prohibited product words is a set of product-side compliance restriction words, representing prohibited expressions related to product ingredients, efficacy, target audience, and qualification boundaries. It serves as the product-side constraint input for full-text prohibition verification and in-slot equivalent rewriting.

[0026] Audience preference data is collected, including: target audience identification data, target audience profile tag data, historical promotional copy text data, historical promotional copy structure type annotation data, and historical preference interaction behavior data. Among them, historical preference interaction behavior data includes: click count data, dwell time data, copy count data, favorite count data, and forward count data. Target audience profile tag data is obtained directly through the user profiling system or audience configuration table. Historical promotional copy text data and historical promotional copy structure type annotation data are obtained directly through the copy content library. Historical preference interaction behavior data is collected directly through exposure click logs and behavior tracking logs.

[0027] The data collected includes hard constraint data for copywriting generation, such as: maximum word count, fixed structure types, information coverage list, mandatory word set, prohibited word set, and format rules. This data is directly collected from the copywriting generation task configuration table, channel format specification table, or compliance rule table. The information coverage list, generated from the copywriting generation task configuration table, originates from the list of mandatory selling points and supporting evidence items configured by operations, serving as a unified benchmark input for information coverage sufficiency and the coverage information-slot binding table. The format rules data, jointly determined by the channel format specification table and compliance rule table, represents channel-side layout structure constraints and compliant expression structure constraints, serving as rule input for generating and aligning the structure slot table. The product prohibited word set data is a product-side compliant restricted word set used to constrain prohibited expressions related to the product itself; the copywriting prohibited word set data is a channel-side or platform-side compliant restricted word set used to constrain prohibited expressions related to the distribution channel rules. These data come from different sources but jointly participate in prohibited word hit verification and conflict detection.

[0028] This implementation plan establishes a unified input baseline covering the product's main semantics, audience preference data, and hard constraint data for copywriting generation by centrally collecting basic data, audience preference evidence, and generation boundary conditions. This ensures consistent data definitions, traceable sources, parsable field structures, and machine-readable constraint rules, solidifying task configurations. It provides input references that can be compared within the same batch and reproduced across batches, reducing the risk of generation instability caused by missing data, data drift, or inconsistent constraint versions. It also enhances the stable execution capabilities of subsequent audience style adaptation judgment, hard constraint conflict structure feasibility judgment, and final draft consistency verification and repair under a unified data version and unified constraint system, improving the auditability, traceability, and consistency of the promotional copywriting generation process.

[0029] Specifically, the preprocessing process for generating the user group-copywriting preference intensity matrix and structure slot table is as follows: First, using a unified character normalization and punctuation rule correction algorithm, the product selling point data, product evidence description data, and historical promotional copy text data are cleaned and formatted to reduce structural ambiguity. Second, using word segmentation and entity phrase retention algorithms, phrases are extracted from the product selling point data and product evidence description data. The phrase extraction results are used as the matching dictionary input for keyword matching and semantic similarity alignment algorithms, used for subsequent calculation of the coverage item ratio in the copywriting coverage information list data and slot mapping in the coverage information-slot binding table. Third, using automatic indexing of prohibited words and forced alignment of required words algorithms, the product prohibited word set data and the copywriting prohibited word set data are processed. A fast matching index is built from the data and the set of required words in the copywriting to avoid prohibited words and verify the constraints of required words during the generation period. Anomaly interaction removal and sliding time window aggregation algorithms are used to process historical preference interaction data by removing extreme values ​​and aggregating time windows to reduce the impact of noisy interactions. A distribution standardization algorithm is used to unify the scale of historical preference interaction data. A linear normalization algorithm is used to map the distribution-standardized historical preference interaction data to a unified interval (0 to 1), and a user group-copywriting preference intensity matrix is ​​constructed for collaborative filtering recommendation calculations. A structure slot table is generated from the fixed structure type data and copywriting format rule data of the copywriting data through structure parsing and slot mapping algorithms, used for structural feasibility determination and final draft consistency verification.

[0030] This implementation plan involves unified preprocessing of product selling point data, product evidence data, historical promotional copy text data, historical preference interaction behavior data, product prohibited word set data, copy prohibited word set data, copy essential word set data, copy fixed structure type data, and copy format rule data. This results in a standardized data state with consistent text expression, reusable phrase dictionaries, fast retrieval of constraint indexes, robust aggregation of interaction signals, and comparable and aligned numerical standards. The plan outputs machine-readable structured intermediate products such as user group-copy preference intensity matrix and structure slot table. This supports the stable execution of subsequent audience style adaptation judgment, copy coverage information list data coverage measurement, coverage information-slot binding table construction, structural feasibility judgment, and final draft consistency verification and repair under the same constraint standards, improving the reproducibility, auditability, and cross-batch consistency of the generation process.

[0031] Specifically, the collaborative filtering similarity neighborhood discrimination analysis of the user group-copywriting preference intensity matrix is ​​as follows: Similar neighborhoods are obtained from the target audience group identifier data on the user group-copywriting preference intensity matrix using a cosine similarity nearest neighbor retrieval algorithm. The candidate set of similar neighborhoods is filtered for consistency using the target audience profile label data, retaining the neighborhood items corresponding to user group identifier data that are consistent with the target audience profile label data. The number of neighborhoods that meet the similarity threshold is counted to obtain the similarity neighborhood count. The similarity threshold is obtained by statistically analyzing the quantiles of the pairwise cosine similarity of all user groups in the user group-copywriting preference intensity matrix, and the value corresponding to the nth quantile is taken as the similarity threshold. , n is obtained by calculating the dispersion of pairwise cosine similarity distributions using the Shannon entropy algorithm, with a value ranging from 0.70 to 0.80; the mean of cosine similarity of each neighborhood within a similar neighborhood is obtained by averaging; the dispersion of neighborhood preference is obtained by calculating the probability distribution of historical promotional copy structure type labeled data within a similar neighborhood using the Shannon entropy algorithm. The probability distribution is obtained by averaging the frequency of structure types in historical promotional copy structure type labeled data within a similar neighborhood. The probability value of the corresponding structure type is obtained by dividing the occurrence frequency of each structure type by the total number of labeled samples within a similar neighborhood. All structure type probability values ​​constitute the structure type probability distribution and are used as input to the Shannon entropy algorithm.

[0032] The logarithmic similarity neighborhood count term is obtained by summing the similarity neighborhood counts with a constant and taking the natural logarithm. A natural logarithmic transformation is used to sublinearly compress the similarity neighborhood counts, making their marginal contribution decrease as the neighborhood size expands. Logarithmic compression is used to transform the size of the nearest neighbor into a stable contribution characterization, reducing the dominant effect of extreme neighborhood sizes on the discrimination results. The neighborhood similarity modulation term is obtained by multiplying the logarithmic similarity neighborhood count term with the mean of neighborhood similarity. Product coupling is used to jointly characterize quantitative and qualitative reliability, avoiding statistical instability caused by insufficient similarity neighborhood counts when only the mean of neighborhood similarity is used. Product coupling is used to simultaneously constrain the sufficiency of nearest neighbor coverage and the consistency of nearest neighbors, suppressing the risk of style misjudgment caused by accidental matching in small samples. The dispersion smoothing term is obtained by summing the neighborhood preference dispersion with a constant. The smoothing of the constant is introduced to handle the degenerate distribution when the neighborhood preference dispersion is close to zero. The structure ensures that the lower bound of the dispersion smoothing term is controlled. The dispersion smoothing term is used to characterize the degree of control over the uncertainty of the structure type preference, serving as a penalty constraint on the extrapolability of the style constraint parameter set. The exponential input term is obtained by dividing the neighborhood similarity modulation term by the dispersion smoothing term. The ratio structure is used to normalize the neighborhood similarity modulation term under the neighborhood preference dispersion constraint. Ratio normalization is used to form a comparable fit strength characterization under uncertainty constraints. The natural exponential function is calculated after taking the negative value of the exponential input term to obtain the exponential decay term. The negative exponential mapping of the natural exponential function is used to construct a monotonically increasing and asymptotically saturated response curve. The negative exponential saturation mapping is used to suppress sensitive fluctuations caused by extreme inputs. The audience style fit discriminant value is obtained by calculating the constant - exponential decay term. The form of the constant - exponential decay term is used to constrain the audience style fit discriminant value to a bounded interval from zero to constant one, while ensuring that the audience style fit discriminant value maintains a strictly monotonically responsive attitude towards the exponential input term. The audience style fit discriminant value is used to measure the reliability of the style constraint parameter set being stably extrapolated from the similar neighborhood. The specific calculation formula is as follows: ; In the formula, This represents the audience style fit discrimination value, which is used to form a stable set of style constraint parameters as a selection criterion and to support cross-batch comparison; This represents the similarity neighborhood count, used to characterize the degree of neighborhood coverage that satisfies the similarity threshold constraint and reflect the size of available nearest neighbor samples; It represents the mean of neighborhood similarity, characterizing the consistency level of nearest neighbors within similar neighborhoods and reflecting the concentration of nearest neighbor quality; It represents the dispersion of neighborhood preferences, used to characterize the level of uncertainty in the preference for structural types within similar neighborhoods.

[0033] This implementation plan completes the construction of similar neighborhoods, consistency filtering, and statistical robust characterization of the target audience identification data, forming a structured discriminative basis that can be used to characterize the coverage scale of neighbors, the level of neighbor consistency, and the degree of control over preference uncertainty. It outputs audience style adaptation discriminant values ​​as a unified quantitative basis for the selection and distribution of style constraint parameter sets, realizing the transformation of style constraint configuration from subjective experience to verifiable adjudication. It suppresses the risk of style drift and structural type selection fluctuations caused by insufficient neighbor sample size, unstable similarity quality, or divergent distribution of structural type preferences, and improves the reproducibility of style constraint parameter sets under the same data version, cross-batch comparison consistency, and auditability of strategy triggering criteria.

[0034] Specifically, the process of generating a set of style constraint parameters based on the results of collaborative filtering similarity neighborhood discriminant analysis is as follows: Figure 3 The diagram shows the flowchart for audience style adaptation judgment and constraint parameter generation, which compares the audience style adaptation judgment value and the audience style adaptation judgment threshold in real time. When the audience style fit discrimination value is greater than or equal to the audience style fit discrimination threshold, the statistical sample selection of historical promotional copy text data uses product category data as the filtering condition, and filters historical promotional copy text data corresponding to the same product category data to form a statistical subset. The historical promotional copy text data is then processed using a hot word density statistics and quantile statistics mapping algorithm to obtain the target interval data for hot word usage intensity. Word segmentation is performed on the historical promotional copy text data, and the proportion of hot word units in each text to the total number of units in that text is counted as the hot word density sequence. Hot word units are defined as units that rank in the top 10% by frequency of occurrence within the historical promotional copy text data. The quantiles corresponding to the 30th and 70th percentiles of the hot word density sequence are calculated, and the interval formed by these two values ​​is taken as the target interval data for hot word usage intensity. The historical promotional copy text data is then processed using a sentence segmentation and sentence length statistics and quantile statistics mapping algorithm to obtain the target interval data for sentence length. The historical promotional copy text data is then segmented by period, question mark, etc. Sentences are segmented using commas, exclamation marks, and line breaks. The character length of each sentence is counted to form a sentence length sequence. The 30th and 70th percentiles of the sentence length sequence are calculated, and the interval formed by these two values ​​is taken as the target interval data for sentence length. The information density target interval data is obtained by combining historical promotional copy text data with product selling point data and product evidence data through coverage information matching density statistics and quantile statistics mapping algorithms. Keyword matching and semantic similarity alignment are performed on product selling point data and product evidence data in historical promotional copy text data, and the number of matching items within every 100 characters is counted to form an information density sequence. The 30th and 70th percentiles of the information density sequence are calculated, and the interval formed by these two values ​​is taken as the target interval data for information density. The target interval data for hot word usage intensity, sentence length, and information density are mapped to the slot-level constraint tables of title slots and body paragraph slots according to the structure slot table, and the set of style constraint parameters is output. The set of style constraint parameters includes: target interval data for hot word usage intensity, target interval data for sentence length, target interval data for information density, and priority sequence data for structural types. Among them, the target interval data for hot word usage intensity, target interval data for sentence length, and target interval data for information density are obtained by combining historical promotional copy text data with historical promotional copy structural type annotation data through a quantile statistical mapping algorithm; the priority sequence data for structural types is obtained by statistically analyzing and sorting the frequency of occurrence of each structural type in historical promotional copy structural type annotation data within similar neighborhoods.

[0035] When the audience style fit discrimination value is less than the audience style fit discrimination threshold, a similar neighborhood expansion search is triggered, and the audience style fit discrimination value is recalculated. If the result of the recalculation is still less than the audience style fit discrimination threshold, a set of general style constraint parameters is obtained by combining historical promotion copy text data with historical promotion copy structure type annotation data through a quantile statistical mapping algorithm. The target interval data of hot word usage intensity in the set of general style constraint parameters is converged to the median interval of the historical distribution. The target interval data of hot word usage intensity is obtained from historical promotion copy text data through hot word density statistics and quantile statistical mapping algorithms, and then enters the structural feasibility discrimination analysis.

[0036] This implementation plan establishes an adaptive generation and distribution mechanism for style constraint parameter sets driven by audience style adaptation discriminant values. It achieves the solidification and output of statistical subsets of historical promotional copy text data under the filtering caliber of product category data. It completes the quantile interval extraction of target interval data for hot word usage intensity, sentence length, and information density, and distributes them to the slot-level constraint table according to the structure slot table. It completes the statistical sorting output of structure type priority sequence data, and constructs a dedicated style constraint parameter set generation path for compliant scenarios and a similar neighborhood expansion retrieval and general style constraint parameter set fallback path for non-compliant scenarios. This reduces the risk of style constraint parameter set drift caused by sample sparsity and unstable preferences, and improves the input stability, cross-batch comparison consistency, and full-process traceability and archiving capabilities of subsequent structural feasibility discriminant analysis.

[0037] Specifically, the process of performing structural feasibility discrimination analysis on the structural slot table and the hard constraint data of copywriting generation is as follows: Based on the structural slot table, the upper limit of copywriting word count data and the fixed structural type data are used to obtain the structural word count capacity value through a slot-level word count budget allocation algorithm. This algorithm uses the structural overhead word count of each slot in the structural slot table as a benchmark and deducts from the remaining slots. The remaining allocable words are then proportionally allocated according to the slot weight to form the budget for each slot. The slot weight is jointly determined by the slot type and the priority of the coverage information-slot binding, with a value range between 0 and 1. The structural word count capacity value serves as the upper bound of the information capacity that the structural slot table can carry, providing a verifiable budget boundary and overload warning basis for the subsequent slot-level constraint table. For each piece of information in the copywriting coverage information list data, the minimum expressible sentence length is obtained through a minimum expressible sentence length estimation algorithm, and the minimum word count requirement for the coverage information is obtained by summing these values. The minimum expressible sentence length estimation algorithm generates a minimum character length estimate based on the shortest sentence template of the corresponding slot in the structural slot table. The value and sentence template are obtained from the slot statistics of fixed structure type data of copywriting and historical promotion copywriting text data. The minimum word count requirement value of the covered information is used as the lower bound constraint to characterize the coverage. It is used to identify the minimum resource consumption that cannot be omitted under the condition of compressed expression of the covered information list data and to support the rigid baseline of feasibility decision. The number of conflict items is obtained by set conflict detection algorithm for copywriting prohibited word set data, copywriting required word set data, product prohibited word set data and product available keyword set data. The number of rules that cannot be satisfied at the same time is obtained by rule dual conflict detection on copywriting format rule data. Rule dual conflict detection is used to identify and count mutually exclusive punctuation constraints, mutually exclusive paragraph boundary constraints and mutually exclusive required word position constraints in the same slot. Combined with the number of rules that cannot be satisfied at the same time in copywriting format rule data, the constraint conflict degree is obtained. The constraint conflict degree is used as a unified quantitative caliber of the strength of inability to be satisfied at the same time. It is used to characterize the degree of shrinkage of the solvable space of the constraint set under the same generation task configuration and to provide sorting input for subsequent conflict item decomposition and positioning.

[0038] The minimum word count requirement for coverage information is summed with a constant to obtain a demand smoothing term. This constant is introduced to provide a lower bound constraint on the denominator, preventing numerical divergence in the ratio calculation when the minimum word count requirement is close to zero. The capacity requirement ratio is obtained by dividing the structural word count capacity by the demand smoothing term. This ratio is used to construct a dimensionless relative capacity measure, allowing direct comparison between the structural word count capacity and the minimum word count requirement for coverage information under different magnitudes and task scales. Finally, the capacity requirement logarithm term is obtained by summing the capacity requirement ratio term with a constant and taking its natural logarithm. The method introduces a diminishing marginal gain mechanism, causing the improvement in the structural feasibility judgment value as the capacity demand ratio increases to gradually converge. The sum of the constraint conflict degree and the constant 1 is calculated, and the natural logarithm is taken to obtain the conflict logarithm term. A natural logarithm mapping is used to smooth the discrete cumulative effect of the constraint conflict degree into a continuous penalty scale, ensuring that the suppression of the structural feasibility judgment value as the constraint conflict degree increases exhibits a controllable nonlinear growth. The structural feasibility judgment value is obtained by subtracting the conflict logarithm term from the capacity demand logarithm term and inputting it into a logistic function mapping. The logistic function mapping is used to compress the net feasibility to a bounded interval from zero to the constant 1 while maintaining monotonicity. The specific calculation formula is as follows: ; In the formula, The structural feasibility judgment value is used to characterize the degree to which a feasible structural configuration can be formed under the constraints of the hard constraints of the document generation data and the constraints of the structural slot table. This represents the structure word count capacity value, used to characterize the total number of slot-level word resources that can be allocated under the constraints of the structure slot table; This represents the minimum word count requirement for the information to be covered, and describes the minimum word count requirement for the list of information to be covered in the copy under the constraint of the shortest possible expression. Indicates the degree of constraint conflict, used to characterize the strength of the inability to simultaneously satisfy the data of the set of required words in the copywriting, the set of prohibited words, the set of available keywords for the product, and the copywriting format rules. This represents a logistic function mapping used to compress the difference logarithm field result to a bounded interval while maintaining monotonicity.

[0039] In this embodiment, Table 1 is a structural feasibility judgment value comparison table, which records in detail the structural word count capacity value, minimum word count requirement value of covered information, constraint conflict degree and the final calculated structural feasibility judgment value under different constraint combination scenarios. It is used to quantify whether the copy structure has feasible configuration space under the combined effect of the upper limit constraint of copy word count, the constraint of copy coverage information list and the constraint of vocabulary and format conflict. Specifically: Scenario 1 has a structural character capacity of 220, a minimum character requirement for coverage information of 120, a constraint conflict degree of 0, and a structural feasibility discrimination value of 0.738, indicating that the structural feasibility is within the acceptable range under the condition that the coverage requirement is achievable and the conflict degree is zero; Scenario 2 has a structural character capacity of 160, a minimum character requirement for coverage information of 150, a constraint conflict degree of 0, and a structural feasibility discrimination value of 0.673, showing that the discrimination value remains usable but convergence occurs when the coverage requirement approaches the capacity; Scenario 3 has a structural character capacity of 150, a minimum character requirement for coverage information of 150, a constraint conflict degree of 2, and a structural feasibility discrimination value of 0.399, indicating that it is feasible when the capacity and requirement are equal and conflict items occur. The feasibility is significantly suppressed. In scenario 4, the structural word count capacity is 120, the minimum word count requirement for the covered information is 150, the constraint conflict degree is 2, and the structural feasibility discrimination value is 0.374, indicating that feasibility further decreases when insufficient capacity and conflict coexist. In scenario 5, the structural word count capacity is 170, the minimum word count requirement for the covered information is 150, the constraint conflict degree is 6, and the structural feasibility discrimination value is 0.233, indicating that the accumulation of conflict degree will strongly suppress feasibility. In scenario 6, the structural word count capacity is 260, the minimum word count requirement for the covered information is 150, the constraint conflict degree is 6, and the structural feasibility discrimination value is 0.280, indicating that even if the capacity is increased, structural feasibility is still mainly suppressed by the constraint conflict item when the conflict degree remains unchanged.

[0040] Table 1. Comparison of Structural Feasibility Judgment Values like Figure 4The scatter plot shows the relationship between the capacity-demand ratio and the structural feasibility discriminant value. Table 1 shows that the structural feasibility discriminant value exhibits a non-linear improvement trend with increasing capacity-demand ratio, but is significantly modulated by the degree of constraint conflict. Specifically, Scenario 1 has the highest capacity-demand ratio and a constraint conflict degree of 0, corresponding to a structural feasibility discriminant value of 0.738, indicating that capacity advantage can be effectively converted into structural feasibility. Scenario 2 has a near-equilibrium capacity-demand ratio and a constraint conflict degree of 0, with a structural feasibility discriminant value of 0.673, reflecting convergence of marginal gains. Scenario 3 and 4 have capacity-demand ratios in the equilibrium or insufficient range and a constraint conflict degree of 2, with structural feasibility discriminant values ​​of 0.399 and 0.374 respectively, showing that the discriminant value declines rapidly when capacity advantage cannot cover demand or when conflicts exist. Scenario 5 and 6 have a constraint conflict degree of 6. Although the capacity-demand ratio in Scenario 6 is significantly higher than in Scenario 5, the structural feasibility discriminant value only increases from 0.233 to 0.280, indicating that the inhibitory effect of conflict degree on structural feasibility is persistent. Overall, the relationship diagram intuitively reflects the constraint that "an increase in the capacity-demand ratio does not necessarily lead to a simultaneous increase in the discriminant value," and can serve as a quantitative basis for triggering strategies such as structural downgrading, coverage information compression and rearrangement, and conflict item decomposition and location.

[0041] This implementation plan establishes a structural feasibility judgment mechanism for hard constraint data and structural slot tables in copywriting generation. It outputs structural feasibility judgment values ​​as a unified basis for adjudicating feasible structural configurations. This achieves comparable alignment and stable compressed expression of slot-level word count resource allocation capabilities, minimum carrying capacity requirements of copywriting coverage information list data, and the inability to simultaneously satisfy constraint conflict strengths within the same dimensionless judgment domain. It suppresses the risk of capacity requirement ratio distortion caused by differences in task scale, suppresses the risk of judgment scale jumps caused by the discrete accumulation of conflict items, and provides a net feasibility quantification with monotonicity and boundedness. This supports the subsequent feasible structural configuration generation process of slot-level word count budget tables and coverage information-slot binding tables, reduces the risk of structural degradation caused by post-generation trimming and patching, and improves the reproducibility, auditability, and cross-batch consistency of the structural configuration generation process.

[0042] Specifically, the process of outputting feasible structural configurations and textual data based on the structural feasibility discriminant analysis results is as follows: Real-time comparison of structural feasibility discriminant values ​​and structural feasibility thresholds: When the structural feasibility discrimination value is less than the structural feasibility discrimination threshold, the conflict items of the constraint conflict degree are decomposed and located, and then the structural reduction strategy and the coverage information compression and rearrangement strategy are executed. The structural reduction strategy is to switch the fixed structure type data of the copy to the fixed structure type of the title slot and the key point slot. The fixed structure type of the title slot and the key point slot is limited by the structure slot table generated by the structural parsing and slot mapping algorithm of the copy fixed structure type data and the copy format rule data. The coverage information compression and rearrangement strategy is to compress the coverage information expression budget based on the shortest expressible sentence length estimation algorithm while keeping the number of items in the copy coverage information list data unchanged. The number of items is the result of the coverage information item count in the copy coverage information list data. The structural feasibility discrimination value is recalculated until the structural feasibility discrimination threshold is met.

[0043] When the structural feasibility judgment value is greater than or equal to the structural feasibility judgment threshold, slot-level budget allocation is performed on the upper limit data of the word count of the copywriting based on the structural slot table. The copywriting coverage information list data is then bound to the corresponding slot to form a coverage information-slot binding table. Simultaneously, a set of style constraint parameters is injected into the slot-level constraint table to limit the target range data of hot word usage intensity and sentence length for each slot. A feasible structural configuration is output, including: a slot-level word count budget table and a coverage information-slot binding table. The coverage information-slot binding table is obtained by mapping the copywriting coverage information list data line by line according to the structural slot table. Based on the feasible structural configuration and the set of style constraint parameters, copywriting text data is generated. Furthermore, the product name data is preferentially inserted as a phrase in the title slot to ensure consistent reference to the product subject in the title slot. Specifically, a structural skeleton consisting of title slots and body paragraph slots is constructed based on the structural slot table. The copywriting coverage information list data is filled into each slot according to the slot-level word count budget table and the coverage information-slot binding table. During the generation process, slot-level interval consistency constraint checks are performed on the target interval data of hot word usage intensity, sentence length, and information density. Real-time hit interception and in-slot equivalent rewriting are performed on the copywriting prohibited word set data and the product prohibited word set data. Finally, the text of each slot is concatenated according to the copywriting format rules and output as the copywriting text data.

[0044] This implementation plan forms a closed loop for the adjudication and generation of feasible structural configurations, driven by structural feasibility judgment values. It completes the iterative convergence handling of conflict items in unsatisfactory situations, including location and decomposition of structural order reduction and compression and rearrangement of coverage information. It completes the structured and solidified output of slot-level word count budget tables and coverage information-slot binding tables for satisfactory situations. It enables the distribution of executable slot-level configurations of copywriting word count limits, fixed structure type data, cover information list data, and style constraint parameter sets under the structural slot table caliber. It establishes budget boundaries, coverage responsibility boundaries, and style constraint boundaries for the generation process, reduces the risk of structural damage caused by post-generation trimming and repair, and improves the structural stability, constraint satisfaction, cross-batch consistency, and full-process traceability and archiving capabilities of the generated copywriting text data.

[0045] Specifically, the joint evaluation process for the information coverage sufficiency, format rule fulfillment sufficiency, repetition / stack penalty, and word count exceedance penalty of the generated copy text data and the hard constraint data of the copy generation is as follows: For the generated copy text data and the copy coverage information list data, the information coverage sufficiency is obtained by calculating the proportion of covered items using keyword matching and semantic similarity alignment algorithms; for the generated copy text data and the copy format rule data, the format rule fulfillment sufficiency is obtained by calculating the proportion of satisfied items using regular expression structure parsing and structure slot table alignment verification algorithms; for the generated copy text data, the repetition ratio is obtained by using the longest common substring repetition detection algorithm and mapped to obtain the repetition / stack penalty, which is one minus the natural exponential function value with a negative repetition ratio as the exponent; for the generated copy text data, the word count exceeds the limit and is mapped to obtain the word count exceedance penalty.

[0046] The product of information coverage sufficiency and format rule fulfillment sufficiency is calculated to obtain the coverage-format joint term. Product coupling is used to emphasize the synergistic relationship between information coverage sufficiency and format rule fulfillment sufficiency, restricting the joint term when either dimension is not met. This prevents a single dimension's achievement from masking a deficiency in another, thus avoiding an artificially high final draft consistency pass value. The sum of the repetition / stack penalty and the word count excess penalty, plus a constant, is calculated to obtain the penalty smoothing term. The constant-smoothing provides a lower bound constraint on the denominator and achieves continuous and controllable accumulation of the penalty term. The coverage-format joint term is divided by the penalty smoothing term to obtain the exponential input term, which is then used in a ratio structure to... Positive consistency contribution and negative degradation penalty are unified within the dimensionless discriminant domain; after taking the negative value of the exponential input term, the natural exponential function is calculated to obtain the exponential decay term. A monotonically increasing and asymptotically convergent response curve is constructed using a negative exponential saturation mapping, which reduces the sensitivity of local fluctuations in the exponential input term to the discriminant output and enhances the stability of the judgment; the constant minus the exponential decay term is calculated to obtain the final draft consistency pass value. The constant minus structure maintains the bounded output interval and makes the pass degree exhibit interpretable gain convergence characteristics as the exponential input term increases, which facilitates the formation of a consistency judgment caliber in the targeted repair and rollback regeneration process. The specific calculation formula is as follows: ; In the formula, This indicates the final draft consistency pass value, which serves as a quantitative basis for determining whether targeted repairs are triggered and whether the final draft is approved. It indicates the sufficiency of information coverage, which is used to measure the degree to which the information list data of the copywriting coverage is valid in the generated copywriting text data and to constrain the risk of information omission; It indicates the sufficiency of format rule satisfaction, measures the degree to which the generated copy text data structurally satisfies the copy format rule data, and constrains the risk of layout structure disruption; This represents the penalty degree for repetition and stacking, which imposes a penalty constraint on information redundancy and expression degradation caused by repeated segments; This indicates the penalty for exceeding the word limit, used to penalize and constrain the risk of pruning and structural damage caused by the generated text data exceeding the word limit.

[0047] This implementation plan establishes a quantitative verification standard for the consistency of the final draft of generated text data. It outputs a joint characterization result of the sufficiency of information coverage and the sufficiency of format rule compliance on the validity of coverage and structural compliance. It also outputs the penalty results of repetition and stacking penalties and word count exceeding penalties on the risk of expression degradation and pruning damage. Through the final draft consistency pass value, the positive consistency contribution and negative degradation penalty are compressed into a bounded discriminant domain of zero to a constant one. This solidifies the unified judgment basis for targeted repair triggering and final draft issuance decisions, reduces the risk of missed detection due to single-dimensional compliance masking coverage omissions or format violations, reduces the risk of false triggering due to local fluctuations causing instability in judgment, and improves the interpretability of the final draft stage, cross-batch comparison consistency, and the executableness and auditability of the repair rollback process.

[0048] Specifically, the process of performing targeted repairs and finalizing the manuscript for signing and archiving involves: real-time comparison of the final manuscript conformity pass value and the final manuscript conformity pass threshold. When the final draft consistency pass value is less than the final draft consistency pass threshold, targeted repair is performed: when at least one piece of copywriting coverage information list data in the generated copywriting text data is not matched by the keyword matching and semantic similarity alignment algorithm, the corresponding missing slot is filled according to the coverage information-slot binding table; when the number of format rule violations obtained by the regular expression structure parsing and structure slot table alignment verification algorithm of the generated copywriting text data is not zero, the paragraph boundaries and line break rules are rearranged according to the structure slot table; when the text number of the generated copywriting text data is greater than the upper limit of the copywriting word count or the longest common substring duplication detection algorithm identifies the existence of duplicate segments, the expression of non-covered information is compressed according to the slot-level word count budget table and the duplicate segments are merged and rewritten; after repair, the final draft consistency pass value is recalculated. If it is still less than the final draft consistency pass threshold, it reverts to the structural feasibility discrimination analysis and switches the copywriting fixed structure type data according to the structural reduction strategy and regenerates the copywriting text data.

[0049] When the final draft consistency pass value is greater than or equal to the final draft consistency pass threshold, a full-text search verification is performed on the product prohibited word set data and the copy prohibited word set data, and a structural slot table alignment review is performed. The final promotion copy text data, the structural type identifier corresponding to the copy fixed structure type data, the coverage information coverage result data, the prohibited word hit verification result data, and the final draft issuance result data are output. The product copy generation basic data, audience preference data, copy generation hard constraint data, user group-copy preference intensity matrix, structural slot table, audience style adaptation judgment value, structural feasibility judgment value, final draft consistency pass value, targeted repair action record data, and final copy version identifier data are created and archived into the final draft copy database.

[0050] This implementation plan establishes a closed-loop finalization process and archiving mechanism driven by the finalization consistency pass value. It solidifies the convergence path for targeted repair in scenarios where standards are not met, including filling missing coverage slots, rearranging format rules and structures, and compressing and rewriting word counts and repetitions. It also establishes an executable regeneration path for post-repair recalculation and rollback to structural feasibility analysis. Furthermore, it completes pre-issuance compliance review of the product and copywriting prohibited word sets data for compliant scenarios, including full-text search verification and alignment of the structural slot table. The plan outputs final promotional copy text data, structural type identifiers, coverage information results, prohibited word hit verification results, and finalization issuance results. It establishes a full-link archiving closed loop encompassing basic product copy generation data, audience preference data, hard constraint data for copy generation, user group-copywriting preference intensity matrix, structural slot table, audience style adaptation discrimination value, structural feasibility discrimination value, finalization consistency pass value, targeted repair action record data, and final copy version identifier data. This improves the compliance stability, reproducibility, traceability, and audit consistency of the finalization results.

[0051] like Figure 2 As shown, a second aspect of the present invention provides an intelligent generation system for product promotion copy, comprising: This implementation plan forms an end-to-end verifiable constraint generation closed loop for intelligent generation tasks of product promotion copywriting. It achieves structured and solidified input of basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation under the same data version. It outputs a user group-copywriting preference intensity matrix and a structural slot table as a unified basis for style discrimination and structural adjudication. It completes the adaptive generation of style constraint parameter sets and the distribution of slot-level constraints, completes the generation of feasible structural configuration driven by structural feasibility discrimination analysis and stable output of copywriting text data, and completes the joint adjudication and targeted repair convergence handling of information coverage sufficiency, format rule compliance sufficiency, repetition and stacking penalty, and word count exceedance penalty in the final draft stage. It outputs final promotion copywriting text data that can be signed and establishes a traceable archiving link for the entire process of strategy events, discrimination results, and version identifiers. This reduces the risk of structural degradation and unusability caused by hard constraint conflicts and improves the constraint satisfaction, structural stability, cross-batch comparison consistency, and audit reproducibility of the generated results.

[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0053] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligently generating product promotion copy, characterized in that, Includes the following steps: S1: Collect and preprocess basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation to generate a user group-copywriting preference intensity matrix and structural slot table. S2, perform collaborative filtering similarity neighborhood discriminant analysis on the user group-copywriting preference intensity matrix, and generate a set of style constraint parameters based on the results of the collaborative filtering similarity neighborhood discriminant analysis; S3, perform structural feasibility discrimination analysis on the structural slot table and the hard constraint data generated by the copywriting, and output feasible structural configuration and copywriting text data based on the results of the structural feasibility discrimination analysis; S4 performs a joint evaluation of the copy text data and the hard constraint data generated by the copy, assessing the sufficiency of information coverage, the sufficiency of format rule compliance, the penalty for repetition and stacking, and the penalty for exceeding the word limit, and then performs targeted repairs and finalizes the draft for signing and archiving.

2. The intelligent generation method for product promotion copywriting according to claim 1, characterized in that: The specific process for collecting basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation is as follows: Collect basic data for product copywriting generation, which includes: product name data, product category data, product selling point data, product evidence data, set of available keywords for the product, and set of prohibited words for the product. Collect audience preference data, which includes: target audience identification data, target audience profile tag data, historical promotional copy text data, historical promotional copy structure type annotation data, and historical preference interaction behavior data; Collect hard constraint data for copywriting generation, which includes: upper limit of word count for copywriting, fixed structure type of copywriting, list of information covered by copywriting, set of required words for copywriting, set of prohibited words for copywriting, and copywriting format rules data.

3. The intelligent generation method for product promotion copy according to claim 1, characterized in that: The specific process of generating the user group-copywriting preference intensity matrix and structural slot table through preprocessing is as follows: Through unified character standardization and punctuation rule correction algorithms, character cleaning and format standardization are performed on product selling point data, product evidence description data, and historical promotional copy text data. Through word segmentation and entity phrase retention algorithms, phrase extraction is performed on product selling point data and product evidence description data, and the phrase extraction results are used as input to the matching dictionary for keyword matching and semantic similarity alignment algorithms. Through automatic indexing of prohibited words and forced alignment of required words, a fast matching index is established for product prohibited word sets, copy prohibited word sets, and copy required word sets. Through abnormal interaction removal and sliding time window aggregation algorithms, extreme value removal and time window aggregation are performed on historical preference interaction behavior data. The historical preference interaction data is scaled uniformly using a distribution standardization algorithm; the distributed standardization historical preference interaction data is mapped to a unified interval using a linear normalization algorithm, and a user group-copy preference intensity matrix is ​​constructed; and a structure slot table is generated from copy fixed structure type data and copy format rule data using a structure parsing and slot mapping algorithm.

4. The intelligent generation method for product promotion copywriting according to claim 1, characterized in that: The specific process of performing collaborative filtering similarity neighborhood discrimination analysis on the user group-copywriting preference intensity matrix is ​​as follows: The target audience identification data is used to obtain similar neighborhoods on the user group-copywriting preference intensity matrix through the cosine similarity nearest neighbor retrieval algorithm. The candidate set of similar neighborhoods is filtered for consistency by the target audience profile tag data. The number of neighborhoods that meet the similarity threshold is counted to obtain the similar neighborhood count value. The cosine similarity of each neighborhood within the similar neighborhood is calculated by the mean to obtain the neighborhood similarity mean. The probability distribution of historical promotional copy structure type labeled data within similar neighborhoods is used to obtain the neighborhood preference dispersion by Shannon entropy calculation algorithm; The logarithmic similarity neighborhood count term is obtained by summing the similarity neighborhood count with a constant 1 and taking the natural logarithm. The neighborhood similarity modulation term is obtained by multiplying the logarithmic similarity neighborhood count term with the mean neighborhood similarity; the dispersion smoothing term is obtained by summing the neighborhood preference dispersion with a constant; the exponential input term is obtained by dividing the neighborhood similarity modulation term by the dispersion smoothing term; and the exponential decay term is obtained by taking the negative value of the exponential input term and then calculating the natural exponential function. The audience style fit discriminant value is obtained by calculating the constant minus the exponential decay term.

5. The intelligent generation method for product promotion copy according to claim 1, characterized in that: The specific process for generating the style constraint parameter set based on the collaborative filtering similarity neighborhood discriminant analysis results is as follows: Real-time comparison of audience style fit discrimination value and audience style fit discrimination threshold: When the audience style fit discrimination value is greater than or equal to the audience style fit discrimination threshold, the target interval data of hot word usage intensity is obtained from the historical promotion copy text data through the hot word density statistics and quantile statistics mapping algorithm. The target interval data of sentence length is obtained from the historical promotion copy text data through the sentence segmentation and sentence length statistics and quantile statistics mapping algorithm. The target interval data of information density is obtained from the historical promotion copy text data combined with product selling point item data and product evidence description data through the coverage information matching density statistics and quantile statistics mapping algorithm. The target interval data of hot word usage intensity, sentence length, and information density are mapped to the slot-level constraint table of title slot and body paragraph slot according to the structure slot table, and the style constraint parameter set is output. When the audience style fit discrimination value is less than the audience style fit discrimination threshold, a similar neighborhood expansion search is triggered, and the audience style fit discrimination value is recalculated. If the result of the recalculation is still less than the audience style fit discrimination threshold, a set of general style constraint parameters is obtained by combining historical promotion copy text data with historical promotion copy structure type annotation data through a quantile statistical mapping algorithm. The target interval data of hot word usage intensity in the set of general style constraint parameters is converged to the median interval of the historical distribution. The target interval data of hot word usage intensity is obtained from historical promotion copy text data through hot word density statistics and quantile statistical mapping algorithms, and then enters the structural feasibility discrimination analysis.

6. The intelligent generation method for product promotion copy according to claim 1, characterized in that: The specific process of performing structural feasibility discrimination analysis on the structural slot table and the hard constraint data generated from the text is as follows: Based on the structure slot table, the upper limit of the word count in the copywriting and the fixed structure type of the copywriting are used to obtain the structure word count capacity value through the slot-level word count budget allocation algorithm; for each piece of information in the copywriting coverage information list data, the minimum expression length is obtained through the shortest expressible sentence length estimation algorithm and the minimum word count requirement value of the coverage information is obtained by summing them. The number of conflicting items is obtained by using a set conflict detection algorithm on the data of prohibited words in copywriting, the data of required words in copywriting, the data of prohibited words for products, and the data of available keywords for products. The degree of constraint conflict is obtained by combining the number of rules that cannot be satisfied at the same time in the copywriting format rule data. The minimum word count requirement for the covered information is summed with a constant to obtain the demand smoothing term; the capacity requirement ratio is obtained by dividing the structural word count capacity by the demand smoothing term; the capacity requirement ratio is summed with a constant and the natural logarithm is obtained to obtain the capacity requirement logarithm term; the constraint conflict degree is summed with a constant and the natural logarithm is obtained to obtain the conflict logarithm term; the conflict logarithm term is subtracted from the capacity requirement logarithm term and input into a logic function mapping to obtain the structural feasibility judgment value.

7. The intelligent generation method for product promotion copywriting according to claim 1, characterized in that: The specific process of outputting feasible structural configurations and text data based on the structural feasibility analysis results is as follows: Real-time comparison of structural feasibility judgment values ​​and structural feasibility judgment thresholds: When the structural feasibility discrimination value is less than the structural feasibility discrimination threshold, the conflict items of the constraint conflict degree are decomposed and located, and then the structural reduction strategy and the coverage information compression and rearrangement strategy are executed. The structural reduction strategy is to switch the fixed structure type data of the copy to the fixed structure type of the title slot and the key point slot. The fixed structure type of the title slot and the key point slot is limited by the structure slot table generated by the structural parsing and slot mapping algorithm of the copy fixed structure type data and the copy format rule data. The coverage information compression and rearrangement strategy is to compress the coverage information expression budget based on the shortest expressible sentence length estimation algorithm while keeping the number of items in the copy coverage information list data unchanged. The number of items is the result of the coverage information item count in the copy coverage information list data. The structural feasibility discrimination value is recalculated until the structural feasibility discrimination threshold is met. When the structural feasibility judgment value is greater than or equal to the structural feasibility judgment threshold, the upper limit of the word count of the copywriting is allocated at the slot level based on the structural slot table, and the copywriting coverage information list data is bound to the corresponding slot to form a coverage information-slot binding table. At the same time, the style constraint parameter set is injected into the slot-level constraint table, and a feasible structural configuration is output. Based on the feasible structural configuration and the style constraint parameter set, the copywriting text data is generated. The generation is based on the structural slot table to construct a structural skeleton composed of title slots and body paragraph slots, and the copywriting coverage information list data is filled in slot by slot according to the slot-level word count budget table and the coverage information-slot binding table. In the title slot, the product name data is inserted as a phrase with priority.

8. The intelligent generation method for product promotion copy according to claim 1, characterized in that: The specific process for jointly evaluating the information coverage sufficiency, format rule fulfillment sufficiency, repetition penalty, and word count limit penalty of the copy text data and the hard constraint data of copy generation is as follows: The information coverage sufficiency is obtained by calculating the proportion of covered items using keyword matching and semantic similarity alignment algorithms to compare the generated copy text data and the copy coverage information list data. The format rule satisfaction sufficiency is obtained by calculating the proportion of satisfied items using regular expression structure parsing and structure slot table alignment verification algorithms to compare the generated copy text data and the copy format rule data. The repetition ratio is obtained by using the longest common substring repetition detection algorithm to compare the generated copy text data and the repetition stacking penalty. The word count of the generated copy text data and the upper limit of the copy word count data are used to calculate the over-limit and map the over-limit penalty. The product of information coverage sufficiency and format rule satisfaction is used to obtain the coverage format union term; The sum of the duplicate stacking penalty and the word count excess penalty is calculated and a constant is added to obtain the penalty smoothing term; the coverage format joint term is divided by the penalty smoothing term to obtain the exponential input term; the natural exponential function is calculated after taking the negative value of the exponential input term to obtain the exponential decay term; the constant is calculated and the exponential decay term is subtracted to obtain the final draft consistency pass value.

9. The intelligent generation method for product promotion copy according to claim 1, characterized in that: The specific process of performing targeted repair and finalization, signing off, and archiving is as follows: Real-time comparison of the final draft conformity pass value and the final draft conformity pass threshold: When the final draft consistency pass value is less than the final draft consistency pass threshold, targeted repair is performed: when at least one piece of copywriting coverage information list data in the generated copywriting text data is not matched by the keyword matching and semantic similarity alignment algorithm, the corresponding missing slot is filled according to the coverage information-slot binding table; when the number of format rule violations obtained by the regular expression structure parsing and structure slot table alignment verification algorithm of the generated copywriting text data is not zero, the paragraph boundaries and line break rules are rearranged according to the structure slot table; when the text number of the generated copywriting text data is greater than the upper limit of the copywriting word count or the longest common substring duplication detection algorithm identifies the existence of duplicate segments, the expression of non-covered information is compressed according to the slot-level word count budget table and the duplicate segments are merged and rewritten; after repair, the final draft consistency pass value is recalculated. If it is still less than the final draft consistency pass threshold, it reverts to the structural feasibility discrimination analysis and switches the copywriting fixed structure type data according to the structural reduction strategy and regenerates the copywriting text data; When the final draft consistency pass value is greater than or equal to the final draft consistency pass threshold, perform full-text search verification of the product prohibited word set data and the copy prohibited word set data, and perform structural slot table alignment review. Output the final promotion copy text data, the structural type identifier corresponding to the copy fixed structure type data, the coverage information coverage result data, the prohibited word hit verification result data, and the final draft issuance result data.

10. A product promotion copy intelligent generation system, employing the product promotion copy intelligent generation method according to any one of claims 1-9, comprising: The data collection and preprocessing module is used to collect basic data for product copywriting generation, audience preference data, and hard constraint data for copywriting generation, and to preprocess them to generate a user group-copywriting preference intensity matrix and a structural slot table. The audience style adaptation constraint discrimination module is used to perform collaborative filtering similarity neighborhood discrimination analysis on the user group-copy preference intensity matrix, and generate a set of style constraint parameters based on the results of the collaborative filtering similarity neighborhood discrimination analysis. The hard constraint conflict structure feasibility judgment module is used to perform structural feasibility judgment analysis on the structural slot table and the hard constraint data generated by the text, and output feasible structural configuration and text data based on the structural feasibility judgment analysis results. The final draft consistency verification, repair, and issuance module is used to jointly evaluate the information coverage sufficiency, format rule compliance sufficiency, repetition and stacking penalty, and word count exceedance penalty of the text data and the hard constraint data generated by the text, and to perform targeted repair and final draft issuance and archiving.

Images