An e-commerce content generation method and system based on large model inference

Abstract

The application discloses an e-commerce content generation method and system based on large model reasoning, relates to the technical field of e-commerce content generation, and comprises the following steps: obtaining the commodity description information of a target commodity, the behavior preference information of a target user and content placement scene information, and structurally analyzing the information to obtain a commodity element set, a user demand set and a scene limitation condition set; extracting candidate selling point elements, screening and sequencing to generate a selling point priority sequence, and combining user demand to determine a target selling point combination; analyzing scene information to generate an expression strategy comprising a tone type, an information organization method and a guiding method; constructing a staged content generation reasoning process, dividing the generation process into a content organization stage and a sentence generation stage, and introducing a large model to perform generation stage by stage; performing consistency checking and conflict correction on the basis of the output of each stage, and integrating the results after completing all stages to generate target e-commerce content.

Description

Technical Field

[0001] This invention relates to the field of e-commerce content generation technology, specifically to an e-commerce content generation method and system based on large model reasoning. Background Technology

[0002] Against the backdrop of rapid development in large-scale models and generative artificial intelligence, e-commerce content production is gradually shifting from traditional manual writing to data-driven and intelligent generation. Content generation efficiency and personalized expression capabilities have become crucial factors in improving conversion rates and user experience on e-commerce platforms. Simultaneously, with the trends of diversified product categories, more complex user behaviors, and more refined advertising scenarios, content generation methods based solely on rules or templates are no longer sufficient to meet the demands for collaborative expression of multi-dimensional information.

[0003] However, existing e-commerce content generation methods still have shortcomings in multi-source information collaborative processing and content generation control. They lack effective constraint mechanisms in the correlation analysis between product information and user needs, resulting in a low degree of matching between generated content and user requirements. Furthermore, their ability to adapt to scenario factors during content expression is insufficient, making it difficult to meet the requirements of consistency and adaptability under different deployment conditions. Moreover, content generation typically relies on a single process, lacking a phased control and coordination mechanism, which can easily lead to instability in the connection between content organization and output. Additionally, the lack of verification and correction mechanisms for the generated results causes fluctuations in semantic consistency and structural stability, thus affecting the overall quality of content generation. Therefore, there is a need to provide an e-commerce content generation method and system based on phased reasoning of a large model and multi-constraint collaborative control to solve these problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides an e-commerce content generation method and system based on large model inference. This technical solution solves the problems mentioned in the background, such as the lack of an effective constraint mechanism in the correlation analysis of product information and user needs, which leads to a low matching degree between generated content and user needs. Furthermore, the content expression process lacks adaptability to scenario factors, making it difficult to meet the requirements of consistency and adaptability under different deployment conditions. Moreover, the content generation process typically relies on a single workflow, lacking a phased control and coordination mechanism, which easily leads to unstable connections between content organization and expression output. Additionally, the lack of a verification and correction mechanism for the generated results results in fluctuations in semantic consistency and structural stability, thus affecting the overall quality of content generation.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for generating e-commerce content based on large model inference, comprising: Obtain product description information, target user behavior preference information, and content delivery scenario information for the target product, and perform structured analysis to obtain a set of product elements, a set of user needs, and a set of scenario constraints. Based on the product element set, multiple candidate selling point elements are extracted, and each candidate selling point element is screened and sorted according to preset rules to form a selling point priority sequence. Then, based on the user demand set, the matching relationship between each selling point element in the selling point priority sequence and user demand is analyzed, and the target selling point combination is determined. Based on the set of scenario constraints, expressive constraints are imposed on the target selling point combination to determine the corresponding expressive strategy, which includes tone type, information organization method and guidance method. Based on the combination of target selling points and expression strategies, the content generation process is divided into stages, and a content generation reasoning process is constructed. The reasoning process includes a multi-stage processing of content organization and sentence generation. Obtain a pre-trained e-commerce content generation model, execute inference step by step according to the inference process, and output the corresponding stage results at each stage. During the step-by-step inference process, consistency checks are performed on the results of each stage to identify and correct expression conflicts. After completing the entire inference process, the output results of each stage are integrated to generate the target e-commerce content.

[0006] In an optional embodiment, the step of extracting multiple candidate selling point elements based on a set of product elements, filtering and sorting each candidate selling point element according to preset rules to form a selling point priority sequence, and then analyzing the matching relationship between each selling point element in the selling point priority sequence and user needs based on a set of user needs, and determining the target selling point combination, specifically includes: The product description information is split into fields to extract attribute fields, function fields and parameter fields, and semantic annotation is performed on each field to obtain structured product description data. Based on structured product description data, attribute fields, function fields, and parameter fields are categorized and organized to form a set of product elements; Behavioral preference information is analyzed and processed to extract browsing history, click history and purchase history, and user attention features are extracted based on the product feature information corresponding to each record; Based on user attention features, label mapping is performed on user attention features to form a set of user needs; The content delivery scenario information is analyzed and processed to extract the display location, display format and interaction method, forming a set of scenario-specific conditions. Based on the product element set, attribute fields, function fields and parameter fields are combined and analyzed to generate multiple candidate selling point elements; Based on candidate selling point elements, the candidate selling point elements are filtered and sorted according to preset rules to obtain a selling point priority sequence. The preset rules include statistical analysis of attribute fields, function fields and parameter fields in the candidate selling point elements, retaining candidate selling point elements whose statistical count is greater than or equal to the minimum field count requirement, and sorting them from high to low according to the statistical count. Based on the selling point priority sequence and the user demand set, the selling point elements in the selling point priority sequence are matched with the user attention features in the user demand set to determine the number of fields in each selling point element that meet the user attention features. The selling point elements in the priority sequence are filtered based on the number of fields that meet the user's attention characteristics, and the selling point elements with the highest number of fields that meet the user's attention characteristics are combined to obtain the target selling point combination.

[0007] In an optional embodiment, the step of applying expression constraints to the target selling point combination based on the set of scenario-limiting conditions and determining the corresponding expression strategy specifically includes: The set of scene constraints is parsed and processed to extract features of display location, display form, and interaction method, resulting in a set of scene features. Based on the display location characteristics, the display location matching process is performed on each selling point element in the target selling point combination to determine the display location element corresponding to each selling point element, and the display location comparison process is performed on each selling point element to determine the display order relationship between each selling point element. Based on the order of presentation of each selling point element, the selling point elements in the target selling point combination are arranged sequentially to obtain the selling point presentation sequence. Based on the features of the selling point display sequence and display format, the display format type is extracted from the display format features. According to the content presentation requirements corresponding to each display format type, each selling point element in the selling point display sequence is classified and processed, and each selling point element is assigned to the corresponding display format type. Based on the display format type of each selling point element, the presentation structure of each selling point element is divided, the corresponding presentation structure type of each selling point element is determined, and the structure of each selling point element is adjusted to obtain the selling point structure sequence. Based on the characteristics of the interaction methods, obtain the expression requirements under different interaction methods; Based on the expression requirements under different interaction methods, the expression content of each selling point element in the selling point structure sequence is adjusted, the expression method corresponding to each selling point element is determined, and each selling point element is expressed according to the corresponding expression method to obtain the selling point expression sequence. Based on the selling point structure sequence, the selling point elements in the selling point structure sequence are grouped and processed. According to the arrangement order of each selling point element in the selling point structure sequence, the combination relationship and the order relationship between each selling point element are determined to obtain the selling point organization sequence. Based on the selling point organization sequence, the grouping results and arrangement relationships of each selling point element in the selling point organization sequence are analyzed to determine the information organization method. Based on the results of the selling point expression sequence and information organization method, the tone features of the expression content and organizational structure of each selling point element in the selling point expression sequence are analyzed to determine the tone type corresponding to each selling point element and obtain the tone type results. Based on the tone type results, the expression content of each selling point element in the target selling point combination is analyzed to guide the demand. According to the guidance method rules corresponding to different tone types, the guidance method of each selling point element is configured to obtain the guidance method results. The expression strategy is obtained by combining the results of information organization method, tone type and guidance method.

[0008] In an optional embodiment, the content generation process is divided into stages based on the combination of target selling points and expression strategies, and a content generation reasoning process is constructed. This reasoning process includes a multi-stage processing procedure for content organization and sentence generation, specifically including: The elements of each selling point in the target selling point combination are organized and processed according to the information organization method. The elements of each selling point are combined according to the arrangement structure corresponding to the information organization method to obtain the organized selling point content. This process is regarded as the first processing stage. Based on the tone type and guidance method results, the corresponding tone and guidance expression form are matched to each selling point element in the organized selling point content, and the sentence generation rules corresponding to each selling point element are determined. According to the statement generation rules, each selling point element in the organized selling point content is processed to generate corresponding content statements, and the statement generation process is regarded as the second processing stage. The first processing stage is defined as the content organization stage, and the second processing stage is defined as the statement generation stage. The organized selling points obtained in the content organization stage are used as input data for the statement generation stage, and a content generation reasoning process is constructed based on the data transfer relationship between the content organization stage and the statement generation stage.

[0009] In an optional embodiment, the process of acquiring a pre-trained e-commerce content generation model, performing inference stage by stage according to the inference process, and outputting the corresponding stage result at each stage specifically includes: Obtain a pre-trained e-commerce content generation model and perform inference-driven initialization of the e-commerce content generation model according to the content generation inference process; The content generation reasoning process is analyzed according to the processing logic of content organization and sentence generation, and it is determined that it consists of a content organization stage and a sentence generation stage. The processing order and dependencies between the content organization stage and the sentence generation stage are analyzed to determine the sequential dependencies between each stage. Based on sequential dependencies, each inference execution stage is bound to the execution order of the model to form a stage control relationship of the model; The model execution process is scheduled and controlled in stages based on stage control relationships, so as to drive the model to perform inference in the order of stages; Generate the corresponding stage output results in each inference execution stage.

[0010] In an optional embodiment, during the stage-by-stage inference process, consistency checks are performed on the results of each stage to identify and correct expression conflicts. After completing the entire inference process, the output results of each stage are integrated to generate the target e-commerce content. This specifically includes: During the step-by-step reasoning process, consistency verification is performed on the output results of the content organization stage and the statement generation stage. The results of each stage are identified from three dimensions: content semantics, expression structure, and expression mode to identify whether there are expression conflicts between the results of each stage. Based on the expression conflict, the results of the conflict stage are divided into different conflict types, namely, content semantic conflict, expression structure conflict, and expression mode conflict. The statement generation process is adjusted according to the conflict type to obtain the corrected statement generation process, and consistency checks are continuously performed during the adjusted statement generation process. Based on the revised statement generation process, the reasoning process is re-executed in the statement generation stage. After completing the entire reasoning process, obtain the final set of results for each stage; The final stage results are structured and integrated to form a unified content expression structure and generate target e-commerce content.

[0011] Furthermore, an e-commerce content generation system based on large model reasoning is proposed to implement any of the e-commerce content generation methods mentioned above, including: The selling point analysis module is used to perform structured analysis on the product description information of the target product, the behavioral preference information of the target user, and the content placement scenario information to generate a set of product elements, a set of user needs, and a set of scenario limiting conditions. The selling point filtering and combination module is used to extract candidate selling point elements based on the product element set, generate a selling point priority sequence according to preset rules, and perform matching analysis on the selling point priority sequence according to the user demand set to generate a target selling point combination. An expression strategy construction module is used to perform expression constraint analysis on the target selling point combination based on a set of scenario-limiting conditions, and generate a corresponding expression strategy. The expression strategy includes tone type, information organization method and guidance method. The reasoning process construction module is used to divide the content generation process into stages based on the combination of target selling points and expression strategies, and to construct a content generation reasoning process that includes a content organization stage and a sentence generation stage. The phased execution module is used to acquire the pre-trained e-commerce content generation model, and execute the inference step by step according to the content generation inference process, and output the corresponding stage results at each stage. The consistency verification and correction module is used to verify the consistency of the output results of each stage during the stage-by-stage execution process, identify expression conflicts and make corrections, and integrate the results of each stage after the reasoning is completed to generate the target e-commerce content.

[0012] In an optional embodiment, the selling point filtering and combination module includes: The field parsing unit is used to split the product description information into attribute fields, function fields and parameter fields, and perform semantic annotation processing to generate structured product description data. A commodity element construction unit is used to classify attribute fields, functional fields and parameter fields based on structured commodity description data to generate a commodity element set. The user behavior analysis unit is used to analyze browsing history, click history and purchase history in behavior preference information to extract user attention features; The user demand mapping unit is used to perform tag mapping processing based on user attention features to generate a set of user demands. The candidate selling point generation unit is used to combine and analyze attribute fields, functional fields and parameter fields based on the product element set to generate candidate selling point elements. The selling point sorting and filtering unit is used to filter and sort candidate selling point elements based on preset rules, generate a selling point priority sequence, and perform matching analysis on the selling point priority sequence based on the user demand set to filter and obtain the target selling point combination.

[0013] In an optional embodiment, the expression strategy construction module includes: The scene analysis unit is used to analyze the content delivery scene information, extract display location features, display format features and interaction method features, and generate a scene feature set. The position and order analysis unit is used to perform position matching and sequential relationship analysis on the target selling point combination based on the display position characteristics, and generate a selling point display sequence; A form and structure mapping unit is used to classify the selling point display sequence based on the display form characteristics, determine the presentation structure type corresponding to each selling point element, and generate a selling point structure sequence. An expression mode configuration unit is used to adjust the expression content of the selling point structure sequence based on the interaction mode features to generate a selling point expression sequence. The organization method analysis unit is used to analyze the grouping and arrangement relationships of selling point elements based on the selling point structure sequence, and generate information organization method results; The strategy generation unit is used to perform tone type analysis and guidance method configuration processing based on the information organization method results and the selling point expression sequence to generate expression strategies.

[0014] In an optional embodiment, the verification and correction include: The stage result verification unit is used to verify the consistency between the output results of the content organization stage and the sentence generation stage, and to identify expression conflicts from three dimensions: content semantics, expression structure and expression mode. A conflict classification unit is used to classify expression conflicts based on consistency verification results, generating content semantic conflicts, expression structure conflicts, and expression mode conflicts. A correction generation unit is used to adjust the statement generation process based on the conflict type and regenerate the corrected stage output results. An iterative verification unit is used to continuously verify the consistency of the stage output results after correction, and to continue to correct and verify when there are expression conflicts until the consistency requirements are met. The result integration unit is used to perform structured integration processing on the final stage output result set, generate a unified content expression structure, and output the target e-commerce content.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This proposal presents an e-commerce content generation method and system based on large-scale model reasoning. It achieves unified expression and collaborative analysis of multi-source information by structuring product information, user behavior preference information, and content delivery scenario information. By introducing a rule-constrained selling point selection and user demand matching mechanism, it improves the accuracy of the correlation between generated content and user needs. Furthermore, by constructing a scenario-constrained expression strategy generation mechanism, it enables adaptive adjustment of content expression methods under different delivery conditions. By establishing a phased content generation reasoning process, it decouples and collaboratively controls the content organization and sentence generation processes, improving the structural stability of the generation process. Finally, by introducing inter-stage consistency verification and conflict correction mechanisms, it reduces deviations in semantic consistency and expression structure of generated content, thereby enhancing the accuracy, stability, and adaptability of e-commerce content generation. Attached Figure Description

[0016] Figure 1 This is a flowchart of an e-commerce content generation method based on large model reasoning proposed in this invention; Figure 2 This is a flowchart of the selling point selection process in this invention; Figure 3 This is a flowchart illustrating the expression strategy construction and phased reasoning process in this invention. Figure 4 This is a system framework diagram of an e-commerce content generation system based on large model reasoning proposed in this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 - Figure 4 As shown in the figure, an e-commerce content generation method based on large model inference in an embodiment of the present invention includes: Obtain product description information, target user behavior preference information, and content delivery scenario information for the target product, and perform structured analysis to obtain a set of product elements, a set of user needs, and a set of scenario constraints. Based on the product element set, multiple candidate selling point elements are extracted, and each candidate selling point element is screened and sorted according to preset rules to form a selling point priority sequence. Then, based on the user demand set, the matching relationship between each selling point element in the selling point priority sequence and user demand is analyzed, and the target selling point combination is determined. Based on the set of scenario constraints, expressive constraints are imposed on the target selling point combination to determine the corresponding expressive strategies, including tone type, information organization method and guidance method. Based on the combination of target selling points and expression strategies, the content generation process is divided into stages, and a content generation reasoning process is constructed. The reasoning process includes a multi-stage process of content organization and sentence generation. Obtain a pre-trained e-commerce content generation model, execute inference step by step according to the inference process, and output the corresponding stage results at each stage. During the step-by-step inference process, consistency checks are performed on the results of each stage to identify and correct expression conflicts. After completing the entire inference process, the output results of each stage are integrated to generate the target e-commerce content.

[0019] Furthermore, multiple candidate selling point elements are extracted based on the product element set, and each candidate selling point element is screened and sorted according to preset rules to form a selling point priority sequence. Then, based on the user demand set, the matching relationship between each selling point element in the selling point priority sequence and user needs is analyzed, and the target selling point combination is determined, specifically including: The product description information is split into fields to extract attribute fields, function fields and parameter fields, and semantic annotation is performed on each field to obtain structured product description data. Based on structured product description data, attribute fields, function fields, and parameter fields are categorized and organized to form a set of product elements; Behavioral preference information is analyzed and processed to extract browsing history, click history and purchase history, and user attention features are extracted based on the product feature information corresponding to each record; Based on user attention features, label mapping is performed on user attention features to form a set of user needs; The content delivery scenario information is analyzed and processed to extract the display location, display format and interaction method, forming a set of scenario-specific conditions. Specifically, in the early stages of e-commerce content generation, it is necessary to complete the unified parsing and structured accumulation of three types of raw unstructured data: products, users, and scenarios. This involves transforming the scattered text, behavior logs, and scenario configuration data acquired by the platform into standardized sets of data that are matchable, computable, and callable. The overall process includes three parallel processing logics: product data parsing, user behavior data parsing, and scenario data parsing. For product-side data, the raw product description information is first subjected to field splitting. This field splitting process refers to dividing the raw text such as product descriptions, parameter cards, and product attribute bars into three basic fields: attribute fields, functional fields, and parameter fields, based on the common data dimensions of e-commerce products and the field division criteria formed by historical product field classification statistics. Among them, attribute fields correspond to the inherent appearance, material, and structural characteristics of the product; functional fields correspond to the achievable usage effects and operational capabilities of the product; and parameter fields correspond to the quantifiable performance indicators of the product. After field splitting, semantic annotation is performed on all resulting fields. Semantic annotation involves binding standardized e-commerce semantic tags to each independent field to eliminate semantic ambiguity, colloquial descriptions, and repetitive expressions, thus unifying the semantic meaning of the fields. These semantic tags are derived from the platform's historical product semantic tag system and the industry's general product semantic classification rules. A stable tag set is formed through long-term accumulation of product data and annotation statistics, thereby generating structured product description data with a unified format and clear dimensions. "Structured" refers to transforming the raw text into a data format categorized by field type and possessing clear semantic identifiers. Subsequently, the annotated attributes, functions, and parameter fields in the structured product description data are categorized and organized, unifying the scattered fields to form a complete set of product elements. For example, for a target product like a household floor fan, the original product description includes descriptions such as "five-blade fan design, height adjustable, seven-speed support, removable and washable mesh cover, and low noise operation." After being broken down, the attribute fields are five-blade fan and removable mesh cover, the function fields are height adjustment, wind speed adjustment, and water washing, and the parameter fields are seven-speed support and 28dB low noise operation. Then, through semantic annotation to unify the interpretation and remove semantic repetition or decorative expressions without actual reference, a complete set of product elements is finally formed.

[0020] While structuring product data, the system simultaneously performs structured analysis of user behavior preference information and content delivery scenario information. For user behavior preference information, the analysis process involves retrieving the target user's historical interaction logs on the e-commerce platform and extracting browsing history, product click history, and order purchase history. During this process, the effectiveness of the behavioral data is identified, retaining records that reflect the user's true interests. Based on the product feature information corresponding to each behavioral record, the system extracts product features that the user has long-term focus on and frequently prefers, forming user attention features. On this basis, tag mapping is performed, matching discrete user attention features with the platform's built-in e-commerce user demand standard tag library. This transforms behavioral features into standardized demand tags. This standard tag library is constructed based on statistical analysis and summarization of the correlation between historical user behavior data and product features. Through long-term statistical analysis of the correspondence between user behavior and transaction results, a stable tag system is formed, ultimately summarizing to form a set of user demands. For content delivery scenario information, scenario analysis and processing involves identifying the delivery channels and display rules corresponding to e-commerce content. This extracts three types of scenario feature data: display location, display format, and interaction method. These scenario features are then uniformly organized into a set of scenario constraints. This set of constraints limits the location, presentation format, and interaction method of content during content generation. For example, in a product details page's image and text display scenario, the corresponding display location is the first screen area of ​​the product details page, the display format is long text and images, and the interaction method is user browsing and page scrolling. After standardizing and organizing these features, a set of scenario constraints corresponding to the delivery scenario is formed, thus providing a constraint basis for subsequent selling point selection and expression strategy construction.

[0021] Based on the product element set, attribute fields, function fields and parameter fields are combined and analyzed to generate multiple candidate selling point elements; Based on candidate selling point elements, the candidate selling point elements are filtered and sorted according to preset rules to obtain a selling point priority sequence. The preset rules include statistical analysis of attribute fields, function fields and parameter fields in the candidate selling point elements, retaining candidate selling point elements whose statistical count is greater than or equal to the minimum field count requirement, and sorting them from high to low according to the statistical count. Based on the selling point priority sequence and the user demand set, the selling point elements in the selling point priority sequence are matched with the user attention features in the user demand set to determine the number of fields in each selling point element that meet the user attention features. The selling point elements in the priority sequence are filtered based on the number of fields that meet the user's attention characteristics, and the selling point elements with the highest number of fields that meet the user's attention characteristics are combined to obtain the target selling point combination.

[0022] Specifically, after constructing the structured set of product elements and the set of user needs, the first step is to conduct field combination analysis based on the product element set to generate multiple candidate selling point elements. This combination analysis involves processing attribute fields, functional fields, and parameter fields to combine different fields to collectively represent a specific characteristic of the product, thus forming meaningful selling point elements. In the specific processing, various fields in the product element set are combined so that the combined candidate selling point elements reflect product attributes, functions, or parameters, and express the selling point elements around the specific usage characteristics or performance of the product, thereby improving the relevance and comprehensibility of the selling point content. For example, for a household floor fan, the product element set includes attribute fields such as "five-blade fan, detachable mesh cover, floor-standing body", function fields such as "height adjustment, tilt adjustment, detachable washable, multi-speed adjustment", and parameter fields such as "28dB low noise, seven wind speeds, three height adjustments". By combining these fields, multiple candidate selling point elements such as "all-area air supply", "easy to clean", "quiet and energy-saving", and "multi-angle adaptation" can be generated. These multiple candidate selling point elements represent different product characteristics. At the same time, different candidate selling point elements correspond to different product focuses, enabling the subsequent screening process to distinguish between multiple selling points.

[0023] After generating candidate selling point elements, these elements are filtered and sorted according to preset rules to obtain a priority sequence of selling points. The minimum number of fields required is determined based on statistical analysis of historical e-commerce copywriting data from the platform. By statistically analyzing the distribution of fields contained in a single selling point in high-conversion product copywriting, a lower limit for the number of fields used to represent effective selling points is determined, and this lower limit is used as the minimum number of fields required; in this embodiment, it is set to 2. During the filtering process, the number of attribute fields, functional fields, and parameter fields contained within each candidate selling point element is counted, and candidate selling point elements whose field count meets the minimum field count requirement are retained. Furthermore, while meeting the minimum field count requirement, different candidate selling point elements may differ in the completeness of information expression due to differences in the number of fields. Therefore, candidate selling point elements are sorted by the total number of fields, arranged from highest to lowest, to obtain a priority sequence of selling points. After obtaining the priority sequence of selling points, each selling point element is matched against the user needs set. Based on the user attention characteristics in the user needs set, each selling point element in the priority sequence is analyzed to determine the number of fields in each selling point element that satisfy the user attention characteristics. These user attention characteristics characterize the user's focus on product attributes or usage effects. The correspondence between fields and user attention characteristics reflects the coverage of user needs by the selling point elements. For example, when user attention characteristics include "low noise, easy to clean, and adaptable to multiple scenarios," they can be mapped to relevant selling point elements. After matching, each selling point element is filtered based on the number of fields satisfying user attention characteristics. The selling point elements with the highest number of fields satisfying user attention characteristics are combined to obtain the target selling point combination. In this combination, each selling point element collectively covers the main content of the user attention characteristics, thus forming the input basis for subsequent content generation.

[0024] Furthermore, based on the set of scenario-specific constraints, expression constraints are imposed on the target selling point combination to determine the corresponding expression strategy, specifically including: The set of scene constraints is parsed and processed to extract features of display location, display form, and interaction method, resulting in a set of scene features. Based on the display location characteristics, the display location matching process is performed on each selling point element in the target selling point combination to determine the display location element corresponding to each selling point element, and the display location comparison process is performed on each selling point element to determine the display order relationship between each selling point element. Based on the order of presentation of each selling point element, the selling point elements in the target selling point combination are arranged sequentially to obtain the selling point presentation sequence. Based on the features of the selling point display sequence and display format, the display format type is extracted from the display format features. According to the content presentation requirements corresponding to each display format type, each selling point element in the selling point display sequence is classified and processed, and each selling point element is assigned to the corresponding display format type. Based on the display format type of each selling point element, the presentation structure of each selling point element is divided, the corresponding presentation structure type of each selling point element is determined, and the structure of each selling point element is adjusted to obtain the selling point structure sequence. Understandably, after determining the target selling point combination, it's necessary to analyze and process the previously constructed set of scenario constraints to adapt the selling point elements to different display positions and formats. The set of scenario constraints contains scenario information related to e-commerce content placement. Through analysis and processing, three types of scenario features are extracted: display position features, display format features, and interaction method features. These are then organized to form a set of scenario features that can be used for matching and adjusting selling point elements. Specifically, display position features characterize the area where content is displayed on the page, display format features characterize the way content is presented, and interaction method features characterize how users interact with the content. For example, when the placement scenario is an e-commerce product details page displaying images and text, the corresponding display position features are the different display areas on the product details page, the display format features are the image and text content, and the interaction method features are the way users interact with the page. By organizing these features, a corresponding set of scenario features can be formed.

[0025] After constructing the scene feature set, based on the display position features, the display position matching process is performed on each selling point element in the target selling point combination, so that each selling point element corresponds to a corresponding display position element. The display order relationship between each selling point element is determined based on the page area order relationship corresponding to different display position features. The display position features themselves contain positional information within the page, such as the front display area, middle display area, and subsequent display areas. Different display areas correspond to different content presentation orders. Therefore, by matching each selling point element with its corresponding display position feature, the display order relationship between selling points can be formed according to the natural arrangement of display positions on the page. For example, when a selling point element is matched to the first screen display position, it has priority in the display sequence; when matched to a middle or later display position, it is placed later accordingly, thus determining the display order relationship. Based on this, according to the determined display order relationship, the selling point elements in the target selling point combination are sequentially arranged to obtain an ordered selling point display sequence.

[0026] Subsequently, based on the selling point display sequence and display format characteristics, the display format types within the display format characteristics are extracted. Then, according to the content presentation requirements corresponding to different display format types, each selling point element in the selling point display sequence is classified, so that each selling point element corresponds to its respective display format type. Specifically, the display format type represents the expression form of content on the page. For example, the graphic format corresponds to an expression method combining text and images. This type of expression method usually requires the content to have hierarchy and structure. Therefore, during classification, different selling point elements are divided into corresponding display expression categories based on their display position and display format characteristics. After completing the display format classification, the presentation structure of each selling point element is divided according to its display format type. Combined with the expression requirements corresponding to the display format, the content expression method of the selling point elements is adapted and adjusted to ensure consistency between its expression structure and the display format, thereby forming a selling point structure sequence that matches the display scenario. For example, in a graphic and text-based display format, the key selling points can be presented in a concise or segmented manner, ensuring that the structure of different selling points remains consistent with the page display format.

[0027] Based on the characteristics of the interaction methods, obtain the expression requirements under different interaction methods; Based on the expression requirements under different interaction methods, the expression content of each selling point element in the selling point structure sequence is adjusted, the expression method corresponding to each selling point element is determined, and each selling point element is expressed according to the corresponding expression method to obtain the selling point expression sequence. Based on the selling point structure sequence, the selling point elements in the selling point structure sequence are grouped and processed. According to the arrangement order of each selling point element in the selling point structure sequence, the combination relationship and the order relationship between each selling point element are determined to obtain the selling point organization sequence. Based on the selling point organization sequence, the grouping results and arrangement relationships of each selling point element in the selling point organization sequence are analyzed to determine the information organization method. Based on the results of the selling point expression sequence and information organization method, the tone features of the expression content and organizational structure of each selling point element in the selling point expression sequence are analyzed to determine the tone type corresponding to each selling point element and obtain the tone type results. Based on the tone type results, the expression content of each selling point element in the target selling point combination is analyzed to guide the demand. According to the guidance method rules corresponding to different tone types, the guidance method of each selling point element is configured to obtain the guidance method results. The expression strategy is obtained by combining the results of information organization method, tone type and guidance method.

[0028] Specifically, after generating a selling point structure sequence adapted to the scenario specifications, the expression of selling point content can be adapted based on the parsed interaction method features, thereby generating a standardized selling point expression sequence. Interaction method features are a component of scenario features, referring to the user's browsing and interactive e-commerce copywriting behavior patterns, covering different user operation forms such as static browsing, click-based interaction, and comment / consultation. Different user interaction behaviors correspond to different content reception habits, thus corresponding to differentiated copywriting expression needs. In this solution, the expression requirements corresponding to various interaction methods are summarized based on user behavior data from historical e-commerce scenarios, used to characterize the different requirements of users for the level of detail, wording style, and content focus of the copywriting under different interaction behaviors. Based on these expression requirements, the expression content of all selling point elements within the selling point structure sequence can be adjusted, including optimizing sentence wording, adjusting detailed descriptions, and changing the emphasis of advantage expression, so that the output form of each selling point element is adapted to the current user interaction mode, ultimately determining the expression method corresponding to each selling point element and forming the selling point expression sequence. For example, in the scenario of displaying images and text on a product details page, the corresponding interaction method is static browsing by the user. This is characterized by users passively reading and spending a relatively long time on the page. Therefore, the corresponding expression requirements are that the content information needs to be relatively complete, the expression method needs to be clear and intuitive, and the overall wording needs to remain plain and stable. Under this expression requirement, the noise reduction effect in a home use scenario is added to the fan's quiet and energy-saving selling point; the repetitive structure of the multi-angle adaptability selling point is simplified; and the disassembly, assembly, and washing operation instructions are added to the easy-to-clean selling point, thus completing the adaptation of the selling point expression method.

[0029] After adapting the expression of selling points, the process involves grouping and logically organizing selling points based on the structural sequence, generating a hierarchical organizational sequence, and further determining the overall information organization of the copy. Grouping involves categorizing and integrating selling point elements with similar themes or complementary relationships based on their product and functional attributes, forming a unified expression unit for similar selling points. After grouping, the order of each selling point element in the structural sequence is analyzed to determine the combination and sequence relationships between different selling points and groups, thus forming the selling point organizational sequence. For example, the quiet and energy-saving selling point and the multi-angle adaptation selling point are grouped into the same experience optimization unit, and the convenient and clean selling point is grouped into the maintenance convenience unit, forming a sequential connection with the original arrangement, thereby maintaining structural coherence in the overall selling point expression. Based on this, we conduct organizational pattern analysis. The organizational pattern is used to characterize the way selling points are arranged in the overall content. Its essence is the overall summary of the grouping structure and arrangement relationship of selling points, which is used to determine the information organization method.

[0030] Simultaneously, after determining the information organization method and the sequence of selling points, a tone feature analysis is performed on each selling point element in the sequence. This analysis comprehensively judges the word choice, sentence structure, and information emphasis of each selling point element in its specific expression to characterize its overall language expression tendency and determine the corresponding tone type, thus forming a tone type result. The tone type is used to represent the overall expression style of e-commerce copywriting. Its determination process is based on the correspondence between scenario attributes, user characteristics, and content structure. By matching and analyzing the consistency between user reading preferences and content presentation methods in different scenarios, the overall expression style is adapted to the reading habits of the target users. For example, in the scenario of displaying images and text on a product details page, since users primarily acquire information, the corresponding tone type tends to be a plain, clear, and highly explanatory expression style. Based on the tone type results, the guidance method configuration analysis of each selling point element is conducted. The guidance method is used to characterize the connecting expression form and content transition method between selling points. The configuration process adapts and adjusts the introduction, connection, and succession methods between selling points according to the requirements of expression coherence and rhythm control under different tone types, so that different selling points form a natural transition relationship, such as using direct succession, explanatory transition, or scene introduction. Finally, the information organization method, tone type results, and guidance method results are combined and processed. The information organization method is used to determine the overall content arrangement framework, the tone type results are used to determine the language expression style, and the guidance method results are used to determine the content connection form. The three work together to form a unified expression strategy to achieve consistent expression of e-commerce content in terms of structure, style, and connection.

[0031] Furthermore, based on the combination of target selling points and expression strategies, the content generation process is divided into stages, and a content generation reasoning process is constructed. This reasoning process includes a multi-stage process of content organization and sentence generation, specifically including: The elements of each selling point in the target selling point combination are organized and processed according to the information organization method. The elements of each selling point are combined according to the arrangement structure corresponding to the information organization method to obtain the organized selling point content. This process is regarded as the first processing stage. Based on the tone type and guidance method results, the corresponding tone and guidance expression form are matched to each selling point element in the organized selling point content, and the sentence generation rules corresponding to each selling point element are determined. According to the statement generation rules, each selling point element in the organized selling point content is processed to generate corresponding content statements, and the statement generation process is regarded as the second processing stage. The first processing stage is defined as the content organization stage, and the second processing stage is defined as the statement generation stage. The organized selling points obtained in the content organization stage are used as input data for the statement generation stage, and a content generation reasoning process is constructed based on the data transfer relationship between the content organization stage and the statement generation stage.

[0032] Understandably, after completing the construction of a comprehensive expression strategy, the overall e-commerce content generation process can be modularized and broken down into stages based on the established information organization methods, tone types, and guidance methods, thereby constructing a logically closed-loop content generation reasoning process. The first step is to build the first processing stage, which is essentially a content structuring and reorganization process. This stage is mainly used to logically organize and structurally connect all selling point elements within the target selling point combination. The information organization method corresponds to a standardized content layout framework obtained from the preliminary scenario analysis. This framework constrains the order, hierarchy, and combination logic of selling point elements within the overall content, ensuring that core selling points are presented first and auxiliary selling points are expressed progressively. Based on this layout structure, the scattered selling point elements are systematically pieced together and integrated, eliminating problems of logical disarray, hierarchical chaos, and disordered arrangement among selling points. The final output is a structurally complete and logically progressively organized selling point content. This process of organizing selling points based on established rules is uniformly defined as the content organization stage. For example, based on the progressive information organization method established above, the quiet and energy-saving selling points, multi-angle adaptability selling points, and convenient cleaning selling points can be arranged in sequence according to the arrangement structure of "home experience optimization group first, equipment maintenance group second", forming a structured selling point content with clear hierarchy and logical coherence, and completing the first processing stage output.

[0033] Building upon this foundation, and leveraging the structured selling point data output from the content organization phase, the second processing phase—rule matching and sentence generation—is initiated. After completing the selling point structure organization, individual selling points are only neatly arranged and do not yet form directly displayable e-commerce text sentences. Therefore, it is necessary to combine the tone type and guidance method results from the expression strategy to match corresponding expression tones and guiding expressions for each selling point element. Specifically, sentence generation rules constrain the expression form and connection method of individual selling point texts. These rules are jointly determined by the tone type and guidance method, ensuring consistency and coherence in expression style and transition methods across different selling points. Through rule matching, the emphasis and connection methods of each selling point element in expression can be clarified, providing a unified basis for subsequent text generation. For example, for a gentle and professional tone type and guiding methods such as pain point introduction, scene transition, and benefit summary, the silent energy-saving selling point can emphasize pain point explanation, the multi-angle adaptation selling point can emphasize scene description, and the convenient and clean selling point can emphasize usage benefits, thus forming a sentence generation method that matches the characteristics of each selling point.

[0034] Simultaneously, after configuring the sentence generation rules for all selling point elements, the organized selling point content is transcribed and transformed point by point based on the established rules. This transforms the structured static selling point data into fluent, readable, and scenario-appropriate standardized content sentences, and the text generation process driven by these rules is uniformly defined as the sentence generation stage. Having clarified the functions and output formats of the two independent processing stages, the data flow logic between the stages can be analyzed. The structured selling point content output by the content organization stage serves as the input basis for the sentence generation stage, which in turn generates the text based on this. The two stages are executed sequentially, thus forming a coherent content generation reasoning process, achieving a continuous processing flow from the structured organization of selling points to the final copywriting sentence generation.

[0035] Furthermore, a pre-trained e-commerce content generation model is acquired, and inference is performed stage by stage according to the inference process. At each stage, the corresponding stage result is output, specifically including: Obtain a pre-trained e-commerce content generation model and perform inference-driven initialization of the e-commerce content generation model according to the content generation inference process; The content generation reasoning process is analyzed according to the processing logic of content organization and sentence generation, and it is determined that it consists of a content organization stage and a sentence generation stage. The processing order and dependencies between the content organization stage and the sentence generation stage are analyzed to determine the sequential dependencies between each stage. Based on sequential dependencies, each inference execution stage is bound to the execution order of the model to form a stage control relationship of the model; The model execution process is scheduled and controlled in stages based on stage control relationships, so as to drive the model to perform inference in the order of stages; Generate the corresponding stage output results in each inference execution stage.

[0036] Specifically, after completing the content generation inference process setup, the pre-trained e-commerce content generation model can be invoked to perform phased inference tasks. First, the model undergoes a dedicated initialization process tailored to this inference flow. This e-commerce content generation model is pre-trained using samples of e-commerce product copywriting, scenario-based language, and user preference copywriting, possessing structured content organization and sentence generation capabilities. Inference-driven initialization involves loading the previously established phase division rules, information organization methods, tone rules, guidance rules, and phase dependency logic into the model's inference parameters. This switches the model from a general generation method to a specific generation method tailored to the product, user, and scenario, thus achieving adaptation initialization for the inference process. For example, for the task of generating copy for a home floor fan details page, progressive information organization rules, a gentle and professional tone, and guidance rules can be loaded into the model, allowing it to complete the initial configuration according to the predetermined expression strategy.

[0037] Based on this, the content generation reasoning process is logically decomposed and its relationships analyzed to clarify the execution hierarchy and constraint logic of the overall model reasoning. The content generation reasoning process can be broken down into two main processing stages: content organization and sentence generation. The content organization stage is used to group, sort, and structurally combine selling point elements, transforming raw selling point data into structured expressions. The sentence generation stage is used to generate and optimize text based on the structured selling point content and expression rules. Analysis of the content processed in these two stages reveals that the content organization stage is a prerequisite step, and the sentence generation stage depends on the results of the prerequisite step; the two form a fixed sequential execution relationship. Based on this sequential relationship, each reasoning execution stage is bound to the internal execution order of the model, forming a stage control relationship. This control relationship enables staged execution scheduling of the model. After stage binding is completed, the model executes the reasoning process in the order of "content organization first, sentence generation second," ensuring that the processing logic of each stage is not disordered. The staged scheduling process controls the model to execute only the corresponding tasks in different stages, avoiding the mixing of functions between stages. After each stage is completed, the corresponding stage results are output. The content organization stage outputs structured selling point content, and the sentence generation stage outputs e-commerce text content, thus providing a basis for subsequent content consistency verification and result integration.

[0038] Furthermore, during the phased inference process, consistency checks are performed on the results of each stage to identify and correct any expression conflicts. After completing the entire inference process, the outputs of each stage are integrated to generate the target e-commerce content, specifically including: During the step-by-step reasoning process, consistency verification is performed on the output results of the content organization stage and the statement generation stage. The results of each stage are identified from three dimensions: content semantics, expression structure, and expression mode to identify whether there are expression conflicts between the results of each stage. Based on the expression conflict, the results of the conflict stage are divided into different conflict types, namely content semantic conflict, expression structure conflict, and expression mode conflict. The statement generation process is adjusted according to the conflict type to obtain the corrected statement generation process, and consistency checks are continuously performed during the adjusted statement generation process. Based on the revised statement generation process, the reasoning process is re-executed in the statement generation stage. After completing the entire reasoning process, obtain the final set of results for each stage; The final stage results are structured and integrated to form a unified content expression structure and generate target e-commerce content.

[0039] Understandably, during the process of the large model completing content organization reasoning and sentence generation reasoning in stages, it is necessary to continuously perform consistency checks on the stage results output by the two stages to identify expression deviations and content conflicts across stages, ensuring that the final e-commerce copy maintains consistency in logic and expression. Consistency check, as a post-verification mechanism throughout the entire reasoning process, focuses on detecting the compatibility between the results of the preceding content organization stage and the subsequent sentence generation stage, judging from three dimensions: content semantics, expression structure, and expression method. Specifically, by comparing the semantics of the generated text, it is determined whether it is consistent with the meaning of product advantages and user benefits reflected in the structured selling points; by analyzing the paragraph layout, selling point grouping, and hierarchical relationships of the final text, it is determined whether it continues the established information organization method; by analyzing the copywriting style, tone, and guiding form, it is determined whether it conforms to the expression strategy jointly determined by tone type, information organization method, and guiding method; when any dimension shows inconsistency, it is determined that there is an expression conflict between the stage results. For example, if the content organization stage establishes a structured order of selling points such as "quiet fan, easy to clean, and adaptable to multiple scenarios", and the copy generated in the sentence generation stage expands the description of the noise reduction effect in a way that does not conform to the product elements, disrupts the order of selling points, or adopts an expression that is inconsistent with the established tone, then the expression conflict can be identified in the corresponding dimension.

[0040] Based on this, the expression deviations identified during verification are categorized into conflict types according to the dimensions of the deviations: content semantic conflict, expression structure conflict, and expression mode conflict. Content semantic conflict manifests as inconsistencies between the core meaning or parameter description of the generated statement and the selling point elements; expression structure conflict manifests as deviations from the predetermined selling point organization sequence in paragraph order or selling point grouping; and expression mode conflict manifests as a failure to generate a tone or guiding style that follows the expression strategy. After classifying the conflicts, the statement generation process is adjusted accordingly. Based on the different conflict types, corresponding correction rules are matched to modify the wording logic, structural adaptation logic, and tone output logic in the statement generation stage, resulting in a corrected statement generation process. Consistency checks are continuously performed during the adjusted inference process until no new expression conflicts are detected. For example, if a structural conflict occurs, such as placing the cleanliness selling point first and the quietness selling point last in the generated copy, deviating from the intended organizational order, the sorting rules in the sentence generation stage can be adjusted to ensure that the generated text follows the intended order of selling points. If a semantic conflict occurs, such as describing the fan noise reduction parameters in a way that does not conform to the product elements, the parameter expression rules in the sentence generation stage can be adjusted to ensure that the sentence content is consistent with the product elements.

[0041] Following this, based on the adjusted sentence generation processing logic, the entire process of content organization and sentence generation is iteratively executed again, outputting revised stage results. Cross-stage expression conflicts are gradually eliminated through iterative verification, ensuring consistency in output results across stages. After the entire reasoning process has been iteratively verified and no valid expression conflicts remain, the final output content of all stages is collected to form a final stage result set. On this basis, the result set undergoes structured integration processing, unifying sentence format, paragraph structure, and transitional relationships across stages. This eliminates fragmentation and inconsistencies caused by multi-stage reasoning, and aggregates and optimizes scattered structured selling point data, modifiers, and scenario-based descriptions to form a clear, unified content structure. Ultimately, this generates target e-commerce content that is semantically accurate, structurally consistent, and appropriately toned, aligning with product attributes and user needs.

[0042] Furthermore, an e-commerce content generation system based on large model reasoning is proposed to implement any of the e-commerce content generation methods mentioned above, including: The selling point analysis module is used to perform structured analysis on the product description information of the target product, the behavioral preference information of the target user, and the content placement scenario information, and generate a set of product elements, a set of user needs, and a set of scenario constraints. The selling point filtering and combination module is used to extract candidate selling point elements based on the product element set, generate a selling point priority sequence according to preset rules, and perform matching analysis on the selling point priority sequence according to the user demand set to generate a target selling point combination. The expression strategy construction module is used to perform expression constraint analysis on the combination of target selling points based on the set of scenario constraints, and generate corresponding expression strategies. The expression strategies include tone type, information organization method and guidance method. The reasoning process construction module is used to divide the content generation process into stages based on the combination of target selling points and expression strategies, and to construct a content generation reasoning process that includes the content organization stage and the sentence generation stage. The phased execution module is used to acquire the pre-trained e-commerce content generation model and execute the inference step by step according to the content generation inference process, outputting the corresponding stage results at each stage. The consistency verification and correction module is used to verify the consistency of the output results of each stage during the phased execution process, identify expression conflicts and make corrections, and integrate the results of each stage after the reasoning is completed to generate the target e-commerce content.

[0043] Furthermore, the selling point filtering and combination module includes: The field parsing unit is used to split the product description information into attribute fields, function fields, and parameter fields, and perform semantic annotation processing to generate structured product description data. The product element construction unit is used to classify attribute fields, function fields, and parameter fields based on structured product description data to generate a set of product elements. The user behavior analysis unit is used to analyze browsing history, click history and purchase history in the behavior preference information to extract user attention features; The user demand mapping unit is used to perform tag mapping based on user attention features to generate a set of user demands. The candidate selling point generation unit is used to combine and analyze attribute fields, functional fields, and parameter fields based on the product element set to generate candidate selling point elements. The selling point sorting and filtering unit is used to filter and sort candidate selling point elements based on preset rules, generate a selling point priority sequence, and perform matching analysis on the selling point priority sequence based on the user demand set to filter out the target selling point combination.

[0044] Furthermore, the expression strategy building module includes: The scene analysis unit is used to analyze the content delivery scene information, extract display location features, display format features, and interaction method features, and generate a scene feature set. The position and order analysis unit is used to perform position matching and sequential relationship analysis on the target selling point combination based on the display position characteristics, and generate a selling point display sequence. The form and structure mapping unit is used to classify the selling point display sequence based on the display form characteristics, determine the presentation structure type corresponding to each selling point element, and generate the selling point structure sequence. The expression mode configuration unit is used to adjust the expression content of the selling point structure sequence based on the interaction mode characteristics, and generate the selling point expression sequence. The organization method analysis unit is used to analyze the grouping and arrangement relationships of selling point elements based on the selling point structure sequence, and generate information organization method results; The strategy generation unit is used to perform tone type analysis and guidance method configuration processing based on the information organization method results and the selling point expression sequence to generate expression strategies.

[0045] Further verification and correction include: The stage result verification unit is used to verify the consistency of the output results between the content organization stage and the sentence generation stage, and to identify expression conflicts from three dimensions: content semantics, expression structure and expression mode. The conflict classification unit is used to classify expression conflicts based on the consistency verification results, generating content semantic conflicts, expression structure conflicts, and expression mode conflicts. The generation correction unit is used to adjust the statement generation process based on the conflict type and regenerate the corrected stage output results. The iterative verification unit is used to continuously verify the consistency of the stage output results after correction, and to continue to correct and verify when there are expression conflicts until the consistency requirements are met. The results integration unit is used to perform structured integration processing on the final stage output result set, generate a unified content expression structure, and output the target e-commerce content.

[0046] In summary, the advantages of this invention are as follows: By structurally analyzing product information, user behavior preference information, and content delivery scenario information, unified modeling and collaborative analysis of multi-source information are achieved, improving the completeness and consistency of information utilization; through a rule-based selling point screening and user demand matching mechanism, effective selection of candidate selling points is achieved, improving the relevance and hit rate between selling points and user attention points; by constructing a scenario-based expression strategy generation mechanism, tone type, information organization method, and guidance method are collaboratively controlled, enabling content expression to adapt to different delivery scenario requirements; by dividing the content generation process into a phased reasoning process of content organization stage and sentence generation stage, structured control and progressive refinement of the generation process are achieved, improving the logic and controllability of content generation; and by introducing an inter-stage consistency verification and conflict correction mechanism, semantic deviations, structural deviations, and expression deviations in the generation process are identified and corrected, improving the consistency and stability of the generation results; thereby comprehensively improving the accuracy, expression adaptability, and practical application effect of e-commerce content generation.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for generating e-commerce content based on large model inference, characterized in that, include: Obtain product description information, target user behavior preference information, and content delivery scenario information for the target product, and perform structured analysis to obtain a set of product elements, a set of user needs, and a set of scenario constraints. Based on the product element set, multiple candidate selling point elements are extracted, and each candidate selling point element is screened and sorted according to preset rules to form a selling point priority sequence. Then, based on the user demand set, the matching relationship between each selling point element in the selling point priority sequence and user demand is analyzed, and the target selling point combination is determined. Based on the set of scenario constraints, expressive constraints are imposed on the target selling point combination to determine the corresponding expressive strategy, which includes tone type, information organization method and guidance method. Based on the combination of target selling points and expression strategies, the content generation process is divided into stages, and a content generation reasoning process is constructed. The reasoning process includes a multi-stage processing of content organization and sentence generation. Obtain a pre-trained e-commerce content generation model, execute inference step by step according to the inference process, and output the corresponding stage results at each stage. During the step-by-step inference process, consistency checks are performed on the results of each stage to identify and correct expression conflicts. After completing the entire inference process, the output results of each stage are integrated to generate the target e-commerce content.

2. The e-commerce content generation method based on large model inference according to claim 1, characterized in that, The process involves extracting multiple candidate selling point elements based on a set of product elements, filtering and sorting these elements according to preset rules to form a priority sequence of selling points, and then analyzing the matching relationship between each selling point element in the priority sequence and user needs based on a set of user needs to determine the target combination of selling points. Specifically, this includes: The product description information is split into fields to extract attribute fields, function fields and parameter fields, and semantic annotation is performed on each field to obtain structured product description data. Based on structured product description data, attribute fields, function fields, and parameter fields are categorized and organized to form a set of product elements; Behavioral preference information is analyzed and processed to extract browsing history, click history and purchase history, and user attention features are extracted based on the product feature information corresponding to each record; Based on user attention features, label mapping is performed on user attention features to form a set of user needs; The content delivery scenario information is analyzed and processed to extract the display location, display format and interaction method, forming a set of scenario-specific conditions. Based on the product element set, attribute fields, function fields and parameter fields are combined and analyzed to generate multiple candidate selling point elements; Based on candidate selling point elements, the candidate selling point elements are filtered and sorted according to preset rules to obtain a selling point priority sequence. The preset rules include statistical analysis of attribute fields, function fields and parameter fields in the candidate selling point elements, retaining candidate selling point elements whose statistical count is greater than or equal to the minimum field count requirement, and sorting them from high to low according to the statistical count. Based on the selling point priority sequence and the user demand set, the selling point elements in the selling point priority sequence are matched with the user attention features in the user demand set to determine the number of fields in each selling point element that meet the user attention features. The selling point elements in the priority sequence are filtered based on the number of fields that meet the user's attention characteristics, and the selling point elements with the highest number of fields that meet the user's attention characteristics are combined to obtain the target selling point combination.

3. The e-commerce content generation method based on large model inference according to claim 2, characterized in that, The step of applying expression constraints to the target selling point combination based on the set of scenario-defined conditions and determining the corresponding expression strategy specifically includes: The set of scene constraints is parsed and processed to extract features of display location, display form, and interaction method, resulting in a set of scene features. Based on the display location characteristics, the display location matching process is performed on each selling point element in the target selling point combination to determine the display location element corresponding to each selling point element, and the display location comparison process is performed on each selling point element to determine the display order relationship between each selling point element. Based on the order of presentation of each selling point element, the selling point elements in the target selling point combination are arranged sequentially to obtain the selling point presentation sequence. Based on the features of the selling point display sequence and display format, the display format type is extracted from the display format features. According to the content presentation requirements corresponding to each display format type, each selling point element in the selling point display sequence is classified and processed, and each selling point element is assigned to the corresponding display format type. Based on the display format type of each selling point element, the presentation structure of each selling point element is divided, the corresponding presentation structure type of each selling point element is determined, and the structure of each selling point element is adjusted to obtain the selling point structure sequence. Based on the characteristics of the interaction methods, obtain the expression requirements under different interaction methods; Based on the expression requirements under different interaction methods, the expression content of each selling point element in the selling point structure sequence is adjusted, the expression method corresponding to each selling point element is determined, and each selling point element is expressed according to the corresponding expression method to obtain the selling point expression sequence. Based on the selling point structure sequence, the selling point elements in the selling point structure sequence are grouped and processed. According to the arrangement order of each selling point element in the selling point structure sequence, the combination relationship and the order relationship between each selling point element are determined to obtain the selling point organization sequence. Based on the selling point organization sequence, the grouping results and arrangement relationships of each selling point element in the selling point organization sequence are analyzed to determine the information organization method. Based on the results of the selling point expression sequence and information organization method, the tone features of the expression content and organizational structure of each selling point element in the selling point expression sequence are analyzed to determine the tone type corresponding to each selling point element and obtain the tone type results. Based on the tone type results, the expression content of each selling point element in the target selling point combination is analyzed to guide the demand. According to the guidance method rules corresponding to different tone types, the guidance method of each selling point element is configured to obtain the guidance method results. The expression strategy is obtained by combining the results of information organization method, tone type and guidance method.

4. The e-commerce content generation method based on large model inference according to claim 3, characterized in that, The content generation process is divided into stages based on the combination of target selling points and expression strategies, and a content generation reasoning process is constructed. This reasoning process includes a multi-stage processing procedure for content organization and sentence generation, specifically including: The elements of each selling point in the target selling point combination are organized and processed according to the information organization method. The elements of each selling point are combined according to the arrangement structure corresponding to the information organization method to obtain the organized selling point content. This process is regarded as the first processing stage. Based on the tone type and guidance method results, the corresponding tone and guidance expression form are matched to each selling point element in the organized selling point content, and the sentence generation rules corresponding to each selling point element are determined. According to the statement generation rules, each selling point element in the organized selling point content is processed to generate corresponding content statements, and the statement generation process is regarded as the second processing stage. The first processing stage is defined as the content organization stage, and the second processing stage is defined as the statement generation stage. The organized selling points obtained in the content organization stage are used as input data for the statement generation stage, and a content generation reasoning process is constructed based on the data transfer relationship between the content organization stage and the statement generation stage.

5. The e-commerce content generation method based on large model inference according to claim 4, characterized in that, The pre-trained e-commerce content generation model is acquired, and inference is performed stage by stage according to the inference process. At each stage, the corresponding stage result is output, specifically including: Obtain a pre-trained e-commerce content generation model and perform inference-driven initialization of the e-commerce content generation model according to the content generation inference process; The content generation reasoning process is analyzed according to the processing logic of content organization and sentence generation, and it is determined that it consists of a content organization stage and a sentence generation stage. The processing order and dependencies between the content organization stage and the sentence generation stage are analyzed to determine the sequential dependencies between each stage. Based on sequential dependencies, each inference execution stage is bound to the execution order of the model to form a stage control relationship of the model; The model execution process is scheduled and controlled in stages based on stage control relationships, so as to drive the model to perform inference in the order of stages; Generate the corresponding stage output results in each inference execution stage.

6. The e-commerce content generation method based on large model inference according to claim 5, characterized in that, During the stage-by-stage inference process, consistency checks are performed on the results of each stage to identify and correct expression conflicts. After completing the entire inference process, the output results of each stage are integrated to generate the target e-commerce content, specifically including: During the step-by-step reasoning process, consistency verification is performed on the output results of the content organization stage and the statement generation stage. The results of each stage are identified from three dimensions: content semantics, expression structure, and expression mode to identify whether there are expression conflicts between the results of each stage. Based on the expression conflict, the results of the conflict stage are divided into different conflict types, namely, content semantic conflict, expression structure conflict, and expression mode conflict. The statement generation process is adjusted according to the conflict type to obtain the corrected statement generation process, and consistency checks are continuously performed during the adjusted statement generation process. Based on the revised statement generation process, the reasoning process is re-executed in the statement generation stage. After completing the entire reasoning process, obtain the final set of results for each stage; The final stage results are structured and integrated to form a unified content expression structure and generate target e-commerce content.

7. An e-commerce content generation system based on large model inference, used to implement the e-commerce content generation method as described in any one of claims 1-6, characterized in that, include: The selling point analysis module is used to perform structured analysis on the product description information of the target product, the behavioral preference information of the target user, and the content placement scenario information to generate a set of product elements, a set of user needs, and a set of scenario limiting conditions. The selling point filtering and combination module is used to extract candidate selling point elements based on the product element set, generate a selling point priority sequence according to preset rules, and perform matching analysis on the selling point priority sequence according to the user demand set to generate a target selling point combination. An expression strategy construction module is used to perform expression constraint analysis on the target selling point combination based on a set of scenario-limiting conditions, and generate a corresponding expression strategy. The expression strategy includes tone type, information organization method and guidance method. The reasoning process construction module is used to divide the content generation process into stages based on the combination of target selling points and expression strategies, and to construct a content generation reasoning process that includes a content organization stage and a sentence generation stage. The phased execution module is used to acquire the pre-trained e-commerce content generation model, and execute the inference step by step according to the content generation inference process, and output the corresponding stage results at each stage. The consistency verification and correction module is used to verify the consistency of the output results of each stage during the stage-by-stage execution process, identify expression conflicts and make corrections, and integrate the results of each stage after the reasoning is completed to generate the target e-commerce content.

8. The e-commerce content generation system based on large model inference according to claim 7, characterized in that, The selling point filtering and combination module includes: The field parsing unit is used to split the product description information into attribute fields, function fields and parameter fields, and perform semantic annotation processing to generate structured product description data. A commodity element construction unit is used to classify attribute fields, functional fields and parameter fields based on structured commodity description data to generate a commodity element set. The user behavior analysis unit is used to analyze browsing history, click history and purchase history in behavior preference information to extract user attention features; The user demand mapping unit is used to perform tag mapping processing based on user attention features to generate a set of user demands. The candidate selling point generation unit is used to combine and analyze attribute fields, functional fields and parameter fields based on the product element set to generate candidate selling point elements. The selling point sorting and filtering unit is used to filter and sort candidate selling point elements based on preset rules, generate a selling point priority sequence, and perform matching analysis on the selling point priority sequence based on the user demand set to filter and obtain the target selling point combination.

9. The e-commerce content generation system based on large model inference according to claim 7, characterized in that, The expression strategy construction module includes: The scene analysis unit is used to analyze the content delivery scene information, extract display location features, display format features and interaction method features, and generate a scene feature set. The position and order analysis unit is used to perform position matching and sequential relationship analysis on the target selling point combination based on the display position characteristics, and generate a selling point display sequence; A form and structure mapping unit is used to classify the selling point display sequence based on the display form characteristics, determine the presentation structure type corresponding to each selling point element, and generate a selling point structure sequence. An expression mode configuration unit is used to adjust the expression content of the selling point structure sequence based on the interaction mode features to generate a selling point expression sequence. The organization method analysis unit is used to analyze the grouping and arrangement relationships of selling point elements based on the selling point structure sequence, and generate information organization method results; The strategy generation unit is used to perform tone type analysis and guidance method configuration processing based on the information organization method results and the selling point expression sequence to generate expression strategies.

10. The e-commerce content generation system based on large model inference according to claim 7, characterized in that, The verification and correction include: The stage result verification unit is used to verify the consistency between the output results of the content organization stage and the sentence generation stage, and to identify expression conflicts from three dimensions: content semantics, expression structure and expression mode. A conflict classification unit is used to classify expression conflicts based on consistency verification results, generating content semantic conflicts, expression structure conflicts, and expression mode conflicts. A correction generation unit is used to adjust the statement generation process based on the conflict type and regenerate the corrected stage output results. An iterative verification unit is used to continuously verify the consistency of the stage output results after correction, and to continue to correct and verify when there are expression conflicts until the consistency requirements are met. The result integration unit is used to perform structured integration processing on the final stage output result set, generate a unified content expression structure, and output the target e-commerce content.

Images