A method and system for generating product tags based on multimodal product information analysis

Abstract

This invention provides a product tag generation method and system for multimodal product information analysis, belonging to the field of data processing technology. Specifically, it includes: a data acquisition layer: constructing a distributed crawler cluster to parse competitor product detail pages and extract multimodal product information; a modal understanding layer: extracting product attributes; a competitor analysis layer: extracting keywords from competitor titles and reviews, and generating competitor analysis feature vectors based on these keywords; a content generation layer: determining a method for pushing multiple sets of product tags; and a quality assessment and optimization layer: determining a method for pushing and updating the product tags, thereby improving the matching degree between product tags and user needs.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method and system for generating product labels based on multimodal product information analysis. Background Technology

[0002] In e-commerce, the quality of product information (including titles, descriptions, key selling points, image tags, etc.) directly impacts click-through rates and conversion rates. Traditional product information editing relies on manual writing or simple template filling, which is inefficient and struggles to guarantee consistent quality. While rule-based or simple NLP-based information extraction tools exist, they often fail to understand multimodal information (such as images, text, prices, reviews, etc.) and cannot dynamically optimize content based on competitor data. Therefore, there is an urgent need for a system that can automatically understand multimodal product information, integrate competitor analysis, and intelligently generate high-quality product descriptions.

[0003] To address the aforementioned technical issues, CN202211294577.9, "A Keyword Generation System and Method Based on Artificial Intelligence," extracts core product terms from competitor title data; combines this with a pre-set search term dataset to select a first core product term with a frequency higher than a preset frequency value; and generates keywords corresponding to the product based on the first core product term and keyword generation rules. However, the above technical solution has the following technical problems:

[0004] When generating product tags, there is usually a large amount of competitor information. Therefore, how to determine the product tag update and push scheme on multiple platforms based on the identification results of competitor information, and how to quickly obtain product tags that match user needs, so as to ensure that product tags are more in line with user needs, has become an urgent technical problem to be solved.

[0005] Therefore, there is an urgent need for a product label generation method and system based on multimodal product information analysis. Summary of the Invention

[0006] To achieve the objectives of this invention, the following technical solution is adopted:

[0007] Firstly, this application provides a product label generation system for multimodal product information analysis, specifically including:

[0008] Data Acquisition Layer: Construct a distributed crawler cluster to parse competitors' product detail pages and extract competitors' multimodal product information;

[0009] Multimodal Understanding Layer: Image Understanding: Extracts deep visual features from product main images and detail images, and uses the CLIP model finely tuned on e-commerce data to achieve zero-shot recognition of image-to-text tags; encodes product titles, descriptions, and reviews to obtain semantic vectors and extract product attributes;

[0010] Competitive analysis layer: Aggregates the captured product attributes by product category, extracts keywords from competitor titles and reviews, and generates a competitive analysis feature vector for the product based on the keywords;

[0011] Content generation layer: Using the multimodal feature vector of the product and the feature vector of the competitor analysis as conditions, the prompt words are concatenated and input into the large language model to guide the model to generate multiple sets of product tags for the product. Based on the similarity of the product tags of the product and combined with the data of the target platform, the push processing method of multiple sets of product tags is determined.

[0012] Quality assessment and optimization layer: Analyze competitor and product review data to determine the correlation between product tags and user review data, and determine the method for pushing and updating product tags by combining the idle status of product tags and push data.

[0013] The beneficial effects of this invention are as follows:

[0014] Full-process automation: It realizes an automated closed loop from data capture and analysis to content generation and evaluation, reducing the preparation time for single product information from hours to minutes.

[0015] Enhanced content quality and competitiveness: By integrating competitor market analysis and multimodal product understanding, the generated content is not only grammatically correct and informationally complete, but also market-targeted and competitively selling.

[0016] Cross-platform intelligent adaptation: Through rule engine and platform adaptation algorithm, it can automatically convert product information into a format that meets the requirements of different e-commerce platforms, achieving "one-time generation, multiple adaptations".

[0017] High robustness and accuracy: Semantic fuzzy matching is used to handle attribute diversity, and integrity judgment and automatic completion mechanisms are used to deal with data missing, which improves the practicality of the system in real and complex scenarios.

[0018] By analyzing the correlation between product tags and user review data, the idle status of product tags, and push data, a method for pushing and updating product tags was determined. This enabled dynamic updates of product tags across multiple platforms. While improving the exposure efficiency of various types of product tags, it also ensured that the updated product tags were more closely matched with user review data, laying the foundation for improving the optimal product tag recognition and processing efficiency.

[0019] It should be noted that the multimodal product information includes competitors' titles, prices, user reviews, and image URLs.

[0020] Furthermore, the competitor analysis feature vector for the product is generated, specifically including:

[0021] Based on competitors' titles and descriptions, marketing keywords are derived by sorting them using TF-IDF.

[0022] Based on the keyword extraction results from competitors' review data, positive and negative keywords are generated;

[0023] Based on the aforementioned marketing keywords, positive keywords, and negative keywords, a competitor analysis feature vector for the product is generated.

[0024] Secondly, this application provides a product label generation method for multimodal product information analysis, applied to the aforementioned product label generation system for multimodal product information analysis, specifically including:

[0025] S1 acquires multimodal product information of the product, determines the semantic similarity coefficient between the product and the competitor based on the product information, determines the product tag generation scheme of the product based on the semantic similarity coefficient between the product and the competitor of each push target platform, and when the generation scheme requires the generation of multiple sets of product tags, determines the reference competitor for generating tags in the competitor based on the comment data and product tags of the competitor.

[0026] S2 generates multiple sets of product tags for the product based on the generated tag reference competitors and all competitors. Based on the similarity of the product tags and combined with the data of the target platform, the push processing method for the multiple sets of product tags is determined.

[0027] S3 performs push processing of product tags according to the push processing method, analyzes the review data of competitors and products to determine the correlation between product tags and user review data, and determines the push update method of product tags by combining the idle status of product tags and push data.

[0028] How to update product labels, i.e., determining the target of the update process.

[0029] Furthermore, the product information includes the product type and product images.

[0030] Furthermore, the semantic similarity coefficient between the product and its competitors is determined based on the average of the similarity coefficients of multimodal product information.

[0031] Specifically, the competing products are those with a semantic similarity coefficient greater than a preset semantic similarity coefficient threshold.

[0032] Furthermore, the method for determining the push processing method for the multiple sets of product tags is as follows:

[0033] The product tags generated for the product are used as the generated product tags. The number of target platforms for pushing the product and the number of groups of generated product tags are obtained.

[0034] Based on the generated product tags of the group, determine the similarity between the generated product tags of the group and other groups, and determine the similarity coefficient between the product tags of the group and other groups based on the similarity.

[0035] Based on the number of target platforms for the product and the number of groups in which product tags are generated, and combined with the similarity coefficient between the generated product tags of the group and the product tags of other groups, the push processing method for the generated product tags of the group is determined.

[0036] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0038] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0039] Figure 1 This is a framework diagram of a product label generation system based on multimodal product information analysis;

[0040] Figure 2 This is a flowchart of a product tag generation method based on multimodal product information analysis;

[0041] Figure 3 This is a flowchart illustrating the method for determining the product label generation scheme.

[0042] Figure 4 This is a flowchart illustrating the method for determining the push processing method for multiple sets of product tags. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0044] Example 1

[0045] Specifically, such as Figure 1 As shown, this application provides a product label generation system for multimodal product information analysis, specifically including:

[0046] Data Acquisition Layer: Construct a distributed crawler cluster to parse competitors' product detail pages and extract competitors' multimodal product information;

[0047] Multimodal Understanding Layer: Image Understanding: Extracts deep visual features from product main images and detail images, and uses the CLIP model finely tuned on e-commerce data to achieve zero-shot recognition of image-to-text tags; encodes product titles, descriptions, and reviews to obtain semantic vectors and extract product attributes;

[0048] Competitive analysis layer: Aggregates the captured product attributes by product category, extracts keywords from competitor titles and reviews, and generates a competitive analysis feature vector for the product based on the keywords;

[0049] Content generation layer: Using the multimodal feature vector of the product and the feature vector of the competitor analysis as conditions, the prompt words are concatenated and input into the large language model to guide the model to generate multiple sets of product tags for the product. Based on the similarity of the product tags of the product and combined with the data of the target platform, the push processing method of multiple sets of product tags is determined.

[0050] Quality assessment and optimization layer: Analyze competitor and product review data to determine the correlation between product tags and user review data, and determine the method for pushing and updating product tags by combining the idle status of product tags and push data.

[0051] It should be noted that the multimodal product information includes competitors' titles, prices, user reviews, and image URLs.

[0052] This module is responsible for intelligently parsing and semantically refining the original product data (images and text), providing structured knowledge input for subsequent competitor analysis and content generation.

[0053] Semantic analysis of product images:

[0054] To overcome the limitations of traditional image classification models that rely on fixed categories and struggle to adapt to the massive volume of long-tail products in e-commerce, this system employs CLIP, a multimodal pre-trained model based on contrastive learning, as its infrastructure. We first fine-tuned the model for domain adaptation on our self-built large-scale e-commerce image-text pair dataset, enabling its visual encoder to more accurately capture the unique visual patterns of the products.

[0055] Zero-shot attribute recognition: The fine-tuned CLIP model possesses powerful zero-shot transfer capabilities. When processing images of new products, the system does not require retraining. It only needs to provide a text list containing candidate labels (such as "round neck," "V-neck," "polo neck," "lace," "solid color," "stripes," etc.), and the model can calculate the similarity between image features and the text features of each label, outputting the most relevant attribute labels and their confidence scores. This method supports flexible expansion of an unlimited number of attributes.

[0056] Fine-grained feature extraction: The multi-level feature maps output by the backbone visual encoder (such as Vision Transformer) are pooled and fused to generate a comprehensive deep visual feature vector. This vector not only contains global information but also encodes key local details of the product (such as logo, texture, and interface), which can be used by subsequent modules for more complex matching and generation tasks.

[0057] In-depth understanding and structuring of product text:

[0058] For official text such as product titles and descriptions, as well as massive amounts of user-generated content (UGC) reviews, this system adopts a layered processing strategy:

[0059] Semantic vectorization: Using a RoBERTa-large model pre-trained on a large-scale Chinese / multilingual corpus as a text encoder, unstructured text sequences (such as a product description) are converted into fixed-dimensional contextual semantic vectors. These vectors encode the full meaning of the text at a deeper level.

[0060] Structured attribute extraction: Based on semantic vectors, a Named Entity Recognition (NER) model finely tuned on e-commerce corpora is integrated to automatically identify and extract key product attribute entities from the text. These structured attributes are core to competitor benchmarking and generating accurate information.

[0061] Comment Sentiment and Opinion Mining: Fine-grained sentiment analysis is performed on user comments individually. Utilizing a Transformer-based sentiment analysis model, not only is the overall sentiment tendency (positive / negative / neutral) of the comments determined, but more importantly, aspect-level sentiment analysis techniques are used to identify the specific objects (aspects) of user evaluations. This data is a valuable source for uncovering the true strengths and weaknesses of products and generating compelling selling points.

[0062] Cross-modal semantic alignment:

[0063] By leveraging the shared multimodal embedding space learned during the CLIP model pre-training phase, this system achieves semantic alignment between images and text. This means that visual features extracted from images and semantic features extracted from text can be mapped to the same vector space for similarity comparison and joint inference to obtain product attributes. This technology forms the foundation for subsequent implementations of "image-to-text search," "image-to-text mutual inspection," and "multimodal feature fusion," ensuring that the system's understanding of products is comprehensive and consistent.

[0064] Furthermore, the competitor analysis feature vector for the product is generated, specifically including:

[0065] Based on competitors' titles and descriptions, marketing keywords are derived by sorting them using TF-IDF.

[0066] Based on the keyword extraction results from competitors' review data, positive and negative keywords are generated;

[0067] Based on the aforementioned marketing keywords, positive keywords, and negative keywords, a competitor analysis feature vector for the product is generated.

[0068] Furthermore, the system uses the "Women's Clothing - Cardigan" category pool as a basis to perform multidimensional analysis and generate feature vectors.

[0069] Price range analysis: The median price of best-selling products in this category is calculated to be $35.99, with the majority falling between $24.99 and $49.99. A competitive pricing range of $29.99 to $39.99 is recommended.

[0070] Keyword matrix construction:

[0071] Marketing keywords (from title / description): Based on TF-IDF sorting, the core selling points are: [Spring, New Arrival, Soft, Loose Fit, Versatile, Casual, Thin, French Style, Basic Style].

[0072] User mental keywords (from comments): High-frequency positive feedback keywords: [skin-friendly, no pilling, slimming, good drape]; High-frequency negative keywords: [shrinkage, static electricity].

[0073] Attribute distribution statistics: For example, 85% of the best-selling items are "loose" styles; "V-neck" styles account for 60%; and "cotton" is the mainstream material.

[0074] Generate a feature vector for the target product: Generate a competitive analysis feature vector for this "oatmeal-colored knitted cardigan".

[0075] The system integrates the following information into keywords and inputs them into the large language model:

[0076] Input conditions: 1) Multimodal features of this product (oatmeal color, V-neck, loose fit, cotton blend, shell buttons, etc.). 2) Feature vectors from competitor analysis.

[0077] Generation process and results:

[0078] Title Generation: The model learns from competitor title patterns such as "[Core Selling Point] + Product Name + [Attribute / Style] + [Applicable Scenarios]", and combines this with the characteristics of the product to generate multiple candidates. The selected title is: "2024 Spring New Women's French Loose Knit Cardigan Oatmeal V-Neck Long Sleeve Versatile Thin Jacket (Soft, Pill-Resistant, Slimming, Casual)". This title integrates marketing buzzwords (Spring New, French, Versatile), core attributes (Loose, V-Neck, Oatmeal), and user focus points (Soft, Pill-Resistant, Slimming).

[0079] Five key selling points generation: The model generates targeted descriptions based on competitor selling point structures and user review focus areas.

[0080]

Skin-friendly and soft material

[0081]

Exquisite French Minimalist Style

[0082]

Slimming and non-bulky with a good drape

[0083]

Versatile and Practical

[0084] [Focus on Detail and Quality] Treated with an anti-pilling process, it is less prone to pilling and deformation; neat stitching reduces loose threads, committed to providing durable quality. (This directly addresses pain points mentioned in competitor reviews such as "many loose threads" and "prone to pilling").

[0085] Example 2

[0086] like Figure 2 As shown, this application provides a product tag generation method for multimodal product information analysis, specifically including:

[0087] S1 acquires multimodal product information of the product, determines the semantic similarity coefficient between the product and the competitor based on the product information, determines the product tag generation scheme of the product based on the semantic similarity coefficient between the product and the competitor of each push target platform, and when the generation scheme requires the generation of multiple sets of product tags, determines the reference competitor for generating tags in the competitor based on the comment data and product tags of the competitor.

[0088] Furthermore, the product information includes the product type and product images.

[0089] Furthermore, the semantic similarity coefficient between the product and its competitors is determined based on the average of the similarity coefficients of multimodal product information.

[0090] Specifically, the competing products are those with a semantic similarity coefficient greater than a preset semantic similarity coefficient threshold.

[0091] A cross-border e-commerce company plans to simultaneously list and sell a "clothing" product on three major e-commerce platforms: Platform A, Platform B, and Platform C. The core issue is: how many different sets of tags should be generated for this product? How should each set of tags be optimized for the characteristics of different platforms? Furthermore, considering the differences in demand due to the number of target platforms and the differences in the semantic similarity coefficients of competing products, which lead to varying demands for generating multiple sets of tags, the company aims to determine the optimal number of product tag sets to maximize product exposure.

[0092] Table 1 Competitor Data

[0093]

[0094] Specifically, such as Figure 3 As shown, the method for determining the product label generation scheme is as follows:

[0095] S11 determines the number of competitors in different push target platforms based on the semantic similarity coefficient between the competitors and the competitors.

[0096] In the above steps, the number of target platforms for push is determined, and it is determined whether the number of target platforms for push is greater than a preset threshold for the number of target platforms for push. If so, the product tag generation scheme for the product is determined to generate a preset number of product tags. If not, proceed to the next step.

[0097] Based on the number of competing products on different push target platforms, determine whether there are push target platforms with a number of competing products greater than a preset competing product number threshold. If so, determine that the product tag generation scheme for the product is to generate a preset number of product tags. If not, proceed to the next step.

[0098] Determine whether the number of competing products in different push target platforms is within the preset range of the number of competing products. If so, determine that the product tag generation scheme of the product is that only one set of product tags needs to be generated. If not, proceed to step S12.

[0099] 1. Determining the number of platforms

[0100] Judgment condition: Whether the number of target platforms pushed is greater than the preset threshold for the number of target platforms pushed (2). Actual data: Number of platforms = 3. Judgment result: 3>2, the condition is met, trigger decision: determine to generate a preset number of product tags (3 groups).

[0101] 2. Determining the threshold for the number of competing products

[0102] Assuming the above conditions are not met, for example, if the preset threshold for the number of target platforms is 3, the judgment condition is: Does the number of competing platforms exceed the preset threshold (100)? Actual data:

[0103] Platform A has 150 competing products (more than 100), Platform B has 85 competing products (less than 100), and Platform C has 15 competing products (less than 100). Result: Exists (Platform A). Trigger decision: Generate a preset number of product tags (3 groups).

[0104] 3. Determining the range of competitor quantity:

[0105] Assuming the preset threshold for the number of competitors is 150, the competition number range is determined. The determination condition is whether the number of competitors in different target platforms is within the preset competition number range [20, 100]. Actual data: Platform A: 150 > 20, Platform B: 85 > 20, → Yes, Platform C: 15 not greater than 20, → No. The determination result is that not all platforms are within the range. Proceed to the next step: Proceed to step S12.

[0106] The preset target platform number threshold is defined as the minimum number of platforms required to trigger the multi-platform differentiated tagging strategy. Its significance is to distinguish between simple single-platform operation and complex multi-platform operation scenarios.

[0107] Simplify decision-making: When there are few platforms, a general tagging strategy can be directly adopted to reduce decision-making complexity;

[0108] Efficiency optimization: Avoid performing unnecessary complex analyses in a limited number of platform scenarios;

[0109] Resource conservation: Reduce the workload of tag generation when operating on a single or dual platform;

[0110] Preset competitor quantity threshold, defined as: the critical value of the number of competitors to judge the intensity of market competition, and its significance: a quantitative standard for identifying red ocean markets (intense competition) and blue ocean markets (mild competition);

[0111] Competition Identification: Quickly locate highly competitive platforms and indicate the need for refined operations;

[0112] Strategy Tier: Platforms with a high number of competitors adopt more refined tagging strategies;

[0113] Risk warning: Too many competitors indicate a need for stronger differentiation positioning;

[0114] 3. Preset a range for the number of competitors, representing an ideal range of reference data. A number of competitors within this range is considered reliable data quality. Significance: This ensures decisions are based on sufficient but not inadequate reference data. Multiple sets of product tags are only needed when a certain number of competitors exist. Specifically, this includes the following:

[0115] Data quality assurance: Ensure that decisions are based on a sufficient amount of reliable competitor data;

[0116] Avoid extreme interference: Eliminate interference from very few competitors (insufficient data) or very many competitors (excessive noise);

[0117] Analysis effectiveness: Conducting analysis within a reasonable data range improves the accuracy of decision-making;

[0118] S12 uses the semantic similarity coefficient between the product and the competitor on different push target platforms to determine the average value of the semantic similarity coefficient between the product and the competitor on different push target platforms, and uses it as the reference similarity coefficient.

[0119] In the above steps, it is determined whether there is a target platform for push notification with a reference similarity coefficient greater than the preset similarity coefficient value. If yes, proceed to the next step; otherwise, it is determined that the product tag generation scheme for the product is to generate only 1 set of product tags.

[0120] Determine whether the number of competing products in the target platform whose reference similarity coefficient is greater than the preset similarity coefficient is within the preset range of the number of competing products. If yes, determine that the product tag generation scheme is to generate the product tags of the second preset number of groups. If no, proceed to step S13.

[0121] 1. Refer to the similarity coefficient calculation method: calculate the average semantic similarity coefficient between the product and the competitor's product on different push target platforms;

[0122] Actual data:

[0123] Platform A's reference similarity coefficient: 0.82;

[0124] Platform B's reference similarity coefficient: 0.78;

[0125] Platform C reference similarity coefficient: 0.65;

[0126] 2. Similarity coefficient threshold judgment, judgment condition: whether there is a target push platform with a reference similarity coefficient greater than the preset similarity coefficient value (0.75).

[0127] Actual data:

[0128] Platform A: 0.82 > 0.75 → Exists;

[0129] Platform B: 0.78 > 0.75 → Exists;

[0130] Platform C: 0.65 < 0.75 → Does not exist;

[0131] Judgment result: Exists (Platform A and Platform B).

[0132] Next step: Continue to assess

[0133] 3. Determining the number of competing products on similar platforms

[0134] Judgment criteria: Among platforms with a similarity coefficient greater than a preset value, are the number of competing products all within the preset competing product number range [20, 100]?

[0135] Actual data:

[0136] Platforms that meet similar conditions: Platform A and Platform B

[0137] Number of competitors on platform A: 150 ∉ [20,100] → No

[0138] Number of competitors on platform B: 85 ∈ [20,100] → Yes

[0139] Result: Not all are within the interval.

[0140] Proceed to the next step: Proceed to step S13

[0141] Reference similarity coefficient:

[0142] Definition: The arithmetic mean of the semantic similarity coefficients between this product and all competing products on a specific platform;

[0143] Significance: To quantify the degree of overall competition and homogeneity of products on a specific platform;

[0144] Beneficial effects:

[0145] Competitive landscape assessment: A higher value indicates that the product is more conventional and the demand for differentiation is lower;

[0146] Understanding platform characteristics: Reflecting the product ecosystem features of different platforms;

[0147] Strategic guidance: Emphasize cost-effectiveness when similarity is high, and emphasize uniqueness when similarity is low;

[0148] 2. Preset value for similarity coefficient;

[0149] Definition: The semantic similarity threshold for determining whether products are highly homogeneous;

[0150] Significance: A quantitative standard to distinguish between innovative products and improved products;

[0151] Beneficial effects:

[0152] Product positioning identification: Quickly determine the degree of innovation of a product in the target market;

[0153] Tagging strategy adjustment: More refined keyword optimization is needed when there is high similarity;

[0154] Differentiation needs assessment: When the similarity is low, it indicates that the product's unique selling points need to be highlighted;

[0155] 3. Similar target platforms for push notifications;

[0156] Definition: A set of target platforms for push notifications with a similarity coefficient greater than a preset value;

[0157] Significance: To identify platform groups with similar competitive environments and the ability to share labeling strategies;

[0158] Beneficial effects:

[0159] Strategy grouping: Similar platforms can be merged for processing, simplifying the tag generation process;

[0160] Resource optimization: Reduce redundant analysis work on highly similar platforms;

[0161] Efficiency improvement: Adopt a unified or slightly modified labeling strategy for similar platforms;

[0162] S13 determines the product tag generation scheme based on the number of competing products in different push target platforms and the reference similarity coefficient.

[0163] Understandably, the product tag generation scheme for the product is determined based on the number of competing products on different target platforms and the reference similarity coefficient, specifically including:

[0164] The target platforms with a similarity coefficient greater than the preset similarity coefficient value are designated as similar target platforms. The weight value of the similar target platforms is determined based on the number of competing products in the similar target platforms. When the sum of the weight values ​​of the similar target platforms is greater than the preset weight threshold, the product tag generation scheme is determined to generate a preset number of product tags. Otherwise, the product tag generation scheme is determined to generate a second preset number of product tags.

[0165] Similar push target platform identification, identification standard: refer to platforms with a similarity coefficient greater than the preset similarity coefficient value (0.75), actual results: platform A and platform B.

[0166] 2. Weight value calculation;

[0167] Calculation method: The weight value is determined based on the number of competitors in the similar push target platform; Weight calculation formula: Weight = Number of competitors on the platform / Total number of competitors;

[0168] Actual calculation:

[0169] Total number of competitors = 150 + 85 + 15 = 250;

[0170] Platform A's weight = 150 / 250 = 0.60;

[0171] Platform B's weight = 85 / 250 = 0.34;

[0172] Weighted sum = 0.60 + 0.34 = 0.94;

[0173] 3. Weight threshold determination

[0174] Judgment criteria: Whether the sum of the weight values ​​of similar target platforms is greater than the preset weight threshold (0.6);

[0175] Actual data: 0.94 > 0.6

[0176] Judgment result: The condition is true.

[0177] 4. Final Decision

[0178] Decision rule: When the sum of the weights is greater than a preset weight threshold, generate a preset number of product tags;

[0179] Final solution: Generate 3 sets of product tags (customized independently for each platform);

[0180] S13 Step Keyword Definitions and Significance

[0181] 1. Target platform weighting for similar push notifications;

[0182] Definition: A quantitative value representing the importance of similar platforms in overall decision-making, calculated based on the number of competing products;

[0183] Significance: To measure the proportion and influence of similar platforms in the overall competitive environment;

[0184] Importance Quantification: Shifting the assessment of platform importance from qualitative to quantitative calculation.

[0185] Decision objectification: Allocating resources based on data rather than subjective judgment

[0186] Prioritization: Identify which similar platforms require special attention.

[0187] 2. Preset weight thresholds;

[0188] Definition: The cumulative weight threshold for determining whether similar platforms are sufficiently important;

[0189] Significance: Determines whether it is necessary to customize the demarcation criteria for tags for similar platforms;

[0190] Beneficial effects:

[0191] Decision-making standardization: a unified standard for judging importance;

[0192] Strategy consistency: Ensure that similar strategies are adopted under similar circumstances;

[0193] Resource allocation criteria: to guide the optimization of resource allocation within a limited number of tags;

[0194] 3. The sum of the weights;

[0195] Definition: The sum of the weight values ​​of all similar target platforms;

[0196] Significance: To quantify the overall importance of similar platforms in the entire product competitive environment;

[0197] Beneficial effects:

[0198] Overall impact assessment: Measures the combined impact of similar platforms on the overall strategy;

[0199] Comprehensive decision-making basis: Considering the cumulative effect of multiple similar platforms;

[0200] Strategy balance: Finding a balance between complete customization and moderate integration.

[0201] The S11-S13 three-step decision-making method demonstrated in this embodiment transforms complex cross-platform tagging strategy decisions into a standardized calculation process through a progressively advancing judgment logic. This method ensures both the scientific rigor and accuracy of the decision-making process while significantly improving efficiency, providing replicable and quantifiable tagging strategy decision support for e-commerce companies operating on multiple platforms.

[0202] Furthermore, the method for determining the competitive products' generated tags is as follows:

[0203] A clothing brand plans to generate product tags for its new "Women's Autumn / Winter Wool Blend Knitwear". Sixty relevant competing products have been identified on the target e-commerce platform. These products differ significantly in sales volume, number of reviews, and product characteristics (such as material, style, color, and fit). The task is to select the most valuable "reference competitors for tag generation" from these 60 products.

[0204] Competitor data example:

[0205] Competitor P01: Women's wool blend knit sweater. Ranked #1 with 12,500 reviews. Average similarity coefficient: 0.65. Style: Simple and suitable for commuting.

[0206] Competitor P02: 9,800 reviews, ranked second. Average similarity coefficient: 0.72. Style: Retro and artistic.

[0207] Competitor P03: 8,700 reviews, ranked third. Average similarity coefficient: 0.68. Style: casual and relaxed.

[0208] Competitor P04: 7,600 reviews, ranked fourth. Average similarity coefficient: 0.81. Style: Slim-fit base layer.

[0209] Competitor P05: 6,500 reviews, ranked fifth. Average similarity coefficient: 0.58. Style: niche design.

[0210] S21 determines the amount of the competitor's review data based on the competitor's review data, and determines the ranking of the competitor among all competitors based on the amount of the review data;

[0211] In the above steps, it is determined whether the ranking result of the competitor among all competitors is before the preset position. If yes, proceed to step S22; otherwise, it is determined that the competitor does not belong to the generated label reference competitor.

[0212] Based on the initial screening of comment data volume, the term "comment data volume" is explained below.

[0213] Definition: Refers to the number of user reviews of competitors on e-commerce platforms. Significance: When there is a large amount of review data, the competitor's product tags are of higher reference value, so it is necessary to prioritize the selection of products with higher reference value.

[0214] Beneficial effect: Ensures that products with more reference data and higher reference value can be effectively screened.

[0215] Execution process.

[0216] Count the number of reviews for all competing products and sort them from largest to smallest. Set the default sorting position to the top 36.

[0217] Competitors P01, P02, P03, P04, and P05 are all ranked within the top 36, so they all proceed to step S22.

[0218] If a competitor is ranked 37th or lower, it is directly determined that it is not a competitor to be referenced.

[0219] S22 determines the semantic similarity coefficient between the competitor and other competitors of the product based on the competitor's product tags;

[0220] In the above steps, the semantic similarity coefficient between the competitor and other competitors of the product is used to determine whether there are other competitors with a semantic similarity coefficient greater than a preset similarity coefficient threshold. If yes, proceed to the next step; otherwise, proceed to step S23.

[0221] Other competitors with semantic similarity coefficients greater than a preset similarity coefficient threshold are considered as similar competitors. It is determined whether the number of similar competitors is greater than a preset similar competitor number threshold. If so, it is determined that the competitor is not a reference competitor for generating tags. If not, proceed to the next step.

[0222] Based on the average semantic similarity coefficient between the competitor and other competitors of the product, the mean similarity coefficient is determined. It is then determined whether the mean similarity coefficient is less than a preset coefficient threshold. If so, proceed to step S23; otherwise, it is determined that the competitor does not belong to the generated tag reference competitor.

[0223] Semantic similarity analysis and homogenization check; Definition: Semantic similarity coefficient, the degree of textual similarity between competing products in terms of product descriptions, attribute tags, etc.

[0224] Significance: The lower the similarity to other competing products, the more conducive it is to generating personalized product tags, thus making the generated product tags more valuable. Beneficial effect: Helps generate personalized tags and increases exposure.

[0225] Execution process:

[0226] Sub-step 1: Identification of highly similar competitors. For competitor P04, calculate its semantic similarity coefficient with other competitors. Four competitors were found to have similarity coefficients greater than 0.75 with P04. These four competitors were marked as similar competitors. The number of similar competitors equals the threshold of four. Therefore, competitor P04 was determined not to be a reference competitor for generating labels.

[0227] Judgment logic: In the apparel industry, if competitors have too many highly similar products, it indicates that their product labels are highly similar, reducing their reference value.

[0228] Sub-step 2: Similarity coefficient mean judgment. For competitor P05, calculate its similarity coefficient mean as 0.58. Determine if it is less than the preset coefficient threshold of 0.65. If 0.58 is less than 0.65, the condition is met. Competitor P05 proceeds to step S23.

[0229] Industry-specific significance: In the apparel industry, a certain degree of differentiation is necessary to ensure that products can generate personalized product labels. A low average similarity coefficient may indicate that the product label is more personalized, but it may still provide unique label inspiration.

[0230] S23 determines whether the competitor is a reference competitor for generating tags based on the ranking result of the competitor among all competitors and the semantic similarity coefficient between the competitor and other competitors of the product.

[0231] Specifically, the ranking result is determined by sorting the competitor's review data from largest to smallest among all competitors.

[0232] Specifically, based on the ranking of the competitor among all competitors and the semantic similarity coefficient between the competitor and other competitors of the product, it is determined whether the competitor is a reference competitor for generating tags, including:

[0233] Based on the ranking of the competitor among all competitors and the average similarity coefficient of the competitor, the matching coefficient of the competitor is determined, and based on the matching coefficient, it is determined whether the competitor is a reference competitor for generating tags.

[0234] It is understood that the matching coefficient is determined based on the ranking of the competitor among all competitors and the average similarity coefficient of the competitors. The higher the ranking of the competitor among all competitors and the lower the average similarity coefficient of the competitors, the larger the matching coefficient.

[0235] Specifically, based on the product tag generation scheme, the number of product tag generation groups is determined, and the competitor with the largest matching coefficient (number of generation groups minus 1) is used as the reference competitor for generating tags.

[0236] Comprehensive matching coefficient calculation and final screening. Definition of terminology: matching coefficient.

[0237] Definition: A comprehensive score combining competitor ranking and the mean similarity coefficient.

[0238] The calculation formula is based on the following logic: the matching coefficient is inversely proportional to the ranking and inversely proportional to the mean similarity coefficient.

[0239] Significance: To balance market enthusiasm with product uniqueness.

[0240] Beneficial effects: Provides quantitative comparison data to support final selection decisions.

[0241] Execution process:

[0242] Example of matching coefficient calculation: Competitor P01 ranks first with a mean similarity coefficient of 0.65, indicating a relatively high matching coefficient. Competitor P05 ranks fifth with a mean similarity coefficient of 0.58, indicating the highest matching coefficient.

[0243] Final selection decision: Based on the product tag generation scheme, three sets of tags need to be generated. Therefore, the two competing products with the highest matching coefficient are selected as reference competing products.

[0244] Assuming the calculation results show that competitor P05 has the highest matching coefficient and competitor P01 has the second highest matching coefficient, it is ultimately determined that competitor P05 and competitor P01 are selected as reference competitors for generating tags.

[0245] IV. Final Solution and Industry Application Value:

[0246] 4.1 Application of generating tags to reference competitors:

[0247] The value of competitor P05 as a reference lies in providing differentiating labels, such as "designer original" and "niche and unique." It expands the boundaries of style and targets niche audiences.

[0248] The value of competitor P01 as a reference lies in providing mainstream market labels such as "workplace attire" and "basic and versatile," ensuring market acceptance and covering a wide range of people.

[0249] 4.2 Tag Generation Strategy:

[0250] The first set of tags was generated based on competitor P05. The style is design-oriented and niche. Example tags include "designer collaboration" and "textured knitwear".

[0251] The second set of tags was generated based on competitor P01. The style is minimalist and commuter-friendly. Example tags include "basic workplace style" and "versatile knitwear".

[0252] The third set of tags was generated based on a comprehensive analysis of all competing products. The overall approach combines mainstream trends with differentiation. Example tags include "autumn / winter wool blends" and "layered outfits".

[0253] 4.3 Unique and Beneficial Effects for the Apparel Industry:

[0254] Ensure style diversity. Avoid overly concentrated labels or a single style, and cover different consumer groups.

[0255] Precise seasonal matching. Ensures reference is made to autumn / winter competitors, generating tags that meet seasonal needs.

[0256] Professionalize material descriptions. Learn professional material description terminology to improve the visibility of your materials in search results.

[0257] Size system optimized. We referenced the size description methods of top-selling competitors to reduce returns and exchanges.

[0258] Enhanced style details. Utilizing unique style description vocabulary to improve visibility in detailed search results.

[0259] S2 generates multiple sets of product tags for the product based on the generated tag reference competitors and all competitors, and determines the push processing method for the multiple sets of product tags according to the similarity between the product tags and the product tags of competitors.

[0260] Specifically, generating multiple sets of product tags for the aforementioned products includes:

[0261] Based on the competitor's tag generation reference, generate a corresponding set of product tags, and then combine all the competitor's tags to generate a corresponding set of product tags.

[0262] A clothing brand has generated three sets of product tags for its new "Women's Autumn / Winter Wool Blend Knitwear." These three sets of tags are optimized for different style positioning and consumer groups. Now, these three sets of tags need to be pushed to different e-commerce platforms, but there is a problem of mismatch between the number of platforms and the number of tag sets. A scientific push strategy needs to be developed to maximize the utilization efficiency of the tags and avoid high-quality tags being idle.

[0263] Specifically, such as Figure 4 As shown, the method for determining the push processing method for the multiple sets of product tags is as follows:

[0264] Product tag groups generated: three groups in total.

[0265] The first set of tags targets niche, design-oriented styles. Tags include: "Designer Collaboration," "Textured Knitwear," and "Artistic Pleated Designs."

[0266] The second set of tags targets a minimalist work style. It includes tags such as "basic workplace pieces," "versatile knitwear," and "simple solid color design."

[0267] The third group of tags: Mixed style. Includes tags such as "Autumn / Winter wool blend", "Layered outfits", and "Warm and comfortable".

[0268] Assume the number of target platforms for push notifications is two.

[0269] Platform A: A niche e-commerce platform for designers.

[0270] Platform B: A comprehensive e-commerce platform for the general public.

[0271] Calculation of inter-group similarity coefficient:

[0272] The similarity coefficient between the first and second groups is 0.32.

[0273] The similarity coefficient between the first and third groups is 0.48.

[0274] The similarity coefficient between the second and third groups is 0.69.

[0275] S31 uses the product tags generated for the product as the generated product tags, and obtains the number of target platforms for pushing the product and the number of groups of generated product tags;

[0276] S32 determines the similarity between the generated product tags of the group and the generated product tags of other groups, and determines the similarity coefficient between the product tags of the group and other groups based on the similarity.

[0277] S33 determines the push processing method for the generated product tags of a group based on the number of target platforms for pushing the product and the number of groups in which product tags are generated, and in combination with the similarity coefficient between the generated product tags of the group and the product tags of other groups.

[0278] Example 1: It is understood that in one possible embodiment, if the number of target platforms for pushing the product is equal to the number of groups of generated product tags, and the number of target platforms for pushing is greater than a preset threshold for the number of target platforms, then it is determined that different generated product tags are used in different target platforms for pushing.

[0279] Step 1: Basic Condition Judgment

[0280] Judgment condition 1: Whether the number of target platforms pushed to is equal to the number of groups of generated product tags.

[0281] Actual data: There are two platforms and three tag groups. These two figures are not equal. The condition is not met.

[0282] Judgment condition two: Whether the number of target platforms pushed to is greater than the preset threshold for the number of target platforms.

[0283] Preset threshold: three platforms. Actual number of platforms: two. Two is no more than three. Condition not met.

[0284] Conclusion: Since neither of the two conditions is met, proceed to the detailed analysis step.

[0285] Example 2: S311 If the number of target platforms for pushing the product is not equal to the number of groups for generating product tags, or if the number of target platforms for pushing is not greater than a preset target platform number threshold, based on the similarity coefficient between the generated product tags of the group and the product tags of other groups, determine whether the average similarity coefficient between the generated product tags of the group and the product tags of other groups is greater than a preset similarity coefficient threshold. If yes, then the group is determined to be an available push group; otherwise, proceed to the next step.

[0286] Sub-step 1: Calculate the average similarity coefficient between groups.

[0287] For the first group of labels: calculate the average similarity coefficient with the second and third groups.

[0288] The average similarity coefficient of the first group is (0.32 plus 0.48) divided by 2 equals 0.4.

[0289] Determine if the similarity coefficient exceeds the preset threshold of 0.75. If the condition is not met, proceed to the next step.

[0290] S312 identifies other groups whose generated product tags have a similarity coefficient greater than a preset similarity coefficient threshold as similar groups, determines whether there are similar groups for the group, and proceeds to the next step if yes, otherwise, the group is identified as an available push group.

[0291] Sub-step 2: Identify similar groups.

[0292] Other groups with a similarity coefficient greater than 0.75 with the first group of labels are considered similar groups.

[0293] The similarity coefficient between the first group and the second group is 0.32, which is no greater than 0.75.

[0294] The similarity coefficient between the first group and the third group is 0.48, which is no greater than 0.75.

[0295] Therefore, there are no similar groups in the first group of labels.

[0296] According to the rules: if there are no similar groups, then that group will be used as an available push group.

[0297] The first group of tags has been identified as an available push notification group.

[0298] S313 groups groups that are similar to each other into the same group combination, and determines the push processing method in the group combination based on the similarity coefficient between the groups in the group combination and other groups.

[0299] For available push groups that do not have similar groups, the push will be processed on a single push target platform.

[0300] Since the three groups of tags do not have similar categories, there are no group combinations that need to be merged. Therefore, each group of tags is an independent and usable push group.

[0301] Push platform resource allocation:

[0302] Available push target platforms: Platform A and Platform B, a total of two platforms.

[0303] Available tag groups: Group 1, Group 2, and Group 3, for a total of three groups.

[0304] For available push groups that have similar groups, the average similarity coefficient between the groups in the group combination and other groups is determined, and the average similarity coefficient is used as the adaptation factor in the group combination.

[0305] Specifically, based on the adaptation factor, the target platforms for push notifications of available push groups in the group combination are determined from low to high, until no available push target platforms are found.

[0306] The adaptation factor is divided into three groups: the first group is 0.4, the second group of tags has an average similarity coefficient of (0.32 plus 0.69) divided by 2 equals 0.505, and the third group of tags is 0.585. Therefore, the push tags at this time are the second group of tags and the first group of tags.

[0307] Specifically, the idle status of the product tags is determined based on the groups that have not been pushed to the target platform.

[0308] S3 performs push processing of product tags according to the push processing method, and determines the push update method of product tags based on the deviation of multiple sets of product tags and the competitor distribution data of the target platform for pushing product tags.

[0309] A clothing brand's new "Women's Autumn / Winter Wool Blend Knitwear" has completed its initial tag push. Platform A (a niche designer platform) pushed the second set of tags (simple commuter style). Platform B (a general platform) pushed the first set of tags (design-conscious niche style). The first set of tags (general style) was left idle due to insufficient platform resources.

[0310] After two weeks of operation, we have accumulated a certain amount of user comment data. Now, based on actual comment feedback, we need to evaluate whether the existing push tags need to be updated and optimized, especially considering whether to replace the unused first set of tags with the existing platform.

[0311] Effect verification requirement: The initial push is based on prediction and needs to be verified by actual effect data.

[0312] Utilization of idle resources: The first set of high-quality tags is not used, which may result in a waste of resources.

[0313] Dynamic optimization requirements: Market feedback may show that the actual effects of different labels do not match expectations.

[0314] Timing of updates: When and how to update require a scientific basis for decision-making.

[0315] Furthermore, the method for determining the push update method for the product tags is as follows:

[0316] Current push notification status:

[0317] Platform A: Push the second set of tags L2;

[0318] Platform B: Push the third group of tags, L3;

[0319] First tag L1: Idle;

[0320] Comment data statistics (cumulative over 14 days):

[0321] Total number of comments: Ntotal=1200;

[0322] Number of comments related to L1: n1=350;

[0323] Number of comments related to L2: n2=500;

[0324] Number of comments related to L3: n3=250;

[0325] Other irrelevant comments: noneer=100;

[0326] S41 determines the groups that have not been pushed out based on the idle status of product tags;

[0327] Step S41: Identify idle resources – Determine unpushable tag groups

[0328] Execution Process: The system first reviews the push status of all generated product tag groups (L1, L2, L3). The review record shows that L2 has been pushed to platform A, L3 has been pushed to platform B, while L1 has not been used on any platform. Therefore, the system determines that the set of groups that have not undergone push processing is U={L1}.

[0329] Significance of this step: This step is the starting point for resource inventory. The key term "groups not processed for push notifications" refers to complete sets of tags that have been optimized and generated but have not been allocated due to platform limitations or initial strategy arrangements. Its significance lies in clearly identifying available "strategic reserve resources," providing candidate targets for subsequent optimization updates. The beneficial effect is avoiding the complete idleness of high-quality tag resources and creating the possibility for dynamic optimization.

[0330] S42, based on the push data, determines the push target platform for available push groups that do not have similar groups, and uses it as the update target platform;

[0331] Define the optimization scope – determine the target platform for the update;

[0332] Execution process: The system then reviews all platforms currently pushing tags. Platform A is pushing L2, and platform B is pushing L3. These two platforms and their currently pushed tag groups are both considered "available push groups". The system defines the set of these platforms as the update target platform set Pupdate={PA,PB}.

[0333] Significance of this step: This step defines the scope of this optimization operation. The key term "update target platforms" specifically refers to e-commerce platforms that have already been tagged and do not have similar groups. Its significance lies in narrowing the optimization focus from all platforms to those without similar groups. Since these platforms are less representative, updating them effectively improves the efficiency of matching user needs and ensures that update decisions are based on comparable existing data. The beneficial effects are improved targeting and operability of the optimization, avoiding interference from irrelevant factors such as expanding to new platforms.

[0334] S43 determines the method for pushing and updating the product tags based on the groups that have not undergone push processing and the target platform for updating, and in conjunction with the correlation with the comment data.

[0335] Furthermore, the method for determining the push update of the product tags based on the group that has not undergone push processing and the target platform for updating specifically includes:

[0336] By analyzing the review data of competitors and products on the target platform, the correlation coefficient with product tags in different groups is determined;

[0337] It should be noted that the correlation coefficient is determined based on the ratio of the number of reviews containing product tags in the group to the total number of reviews.

[0338] S431 Determine whether the correlation coefficients of the groups that have not undergone push processing are all greater than the correlation coefficients of the groups of the existing updated target platforms. If so, replace the group with the smallest correlation coefficient of the existing product label with the group that has not undergone push processing, that is, perform push processing on the push target platform of the group with the smallest correlation coefficient of the existing product label. If not, proceed to the next step;

[0339] Sub-step S431: Global correlation coefficient comparison

[0340] Execution process: The system calculates the global correlation coefficients of each label group. Based on the two-week review data, r1 = 0.292, r2 = 0.417, and r3 = 0.208 are obtained. Then, it is determined whether the correlation coefficients of the idle group L1 are all greater than the correlation coefficients of the current labels of the existing two platforms. It is found that r1 > r3 holds, but r1 < r2r1 does not hold. Therefore, the condition "all greater than" is not satisfied, and the decision process enters the next sub-step.

[0341] Significance of the step: This step is a quick screening mechanism. Its significance is that if the idle label is significantly better than all existing used labels in all dimensions, it should be given a push opportunity, which is a quick decision-making path in the "obvious advantage" scenario. The beneficial effect is to achieve quick response when the advantage is extremely obvious and simplify the decision-making process.

[0342] S432 When the number of groups that have not undergone push processing is greater than the number of updated target platforms, then it is determined whether the correlation coefficient of the product label corresponding to the updated target platform is greater than the preset correlation coefficient threshold. If so, proceed to the next step. If not, as long as the correlation coefficient of the group that has not undergone push processing within the second time period is greater than the correlation coefficient of the group of the updated target platform, then perform product label update processing on the updated target platform, that is, use the product label of the group that has not undergone push processing to replace the group with the smallest correlation coefficient of the existing product label;

[0343] Execution process: The system compares the number of idle groups ∣U∣ = 1 with the number of target platforms ∣Pupdate∣ = 2. Since 1 is less than 2, it meets the condition that "the number of groups that have not undergone push processing is not greater than the number of updated target platforms", so the decision logic of sub-step S434 is executed.

[0344] Step significance: This step selects different decision-making paths based on the quantitative relationship between idle resources and the platform to be optimized. The key term "quantity not greater than" describes a state of relatively scarce or balanced resources, that is, the number of idle label groups is no more than the platforms to be investigated. Its significance lies in adopting differential strategies for different resource abundances: when there are more idle labels (abundant resources), if not handled more strictly, commodity labels without similar groups may be frequently swapped, making it difficult for highly personalized commodity labels to be effectively exposed, and thus difficult to form effective independent personalized commodity labels; when there are fewer idle labels (tight resources), the decision needs to more precisely search for local optimization opportunities. The beneficial effect is to achieve the adaptability of the strategy, making the decision-making logic match the current resource endowment.

[0345] S433 If the correlation coefficients of the group that has not been pushed within the most recent preset duration are all greater than the correlation coefficient of the commodity label corresponding to the updated target platform, then use the group that has not been pushed to replace the group of the commodity label with the smallest current correlation coefficient, that is, perform a push process on the push target platform of the group of the commodity label with the smallest current correlation coefficient.

[0346] It should be noted that the preset duration is greater than the second duration.

[0347] S434 When the number of groups that have not been pushed is not greater than the number of updated target platforms, then as long as the correlation coefficient of the combination of groups that have not been pushed within the second duration is greater than the correlation coefficient of the commodity label corresponding to the updated target platform, then perform an update process on the commodity label of the updated target platform, that is, use the commodity label of the combination of groups that have not been pushed with a greater correlation coefficient to replace the group of the commodity label with the smallest correlation coefficient.

[0348] Sub-step S434: Fine-grained replacement based on short-term trends

[0349] Execution process: This is the final decision point. The system retrieves the short-term data of the correlation coefficients of each label group within the past second duration (7 days): r1(7)=0.35, r2(7)=0.39, r3(7)=0.22. The core judgment logic is: to find whether there is a target platform whose global correlation coefficient of the current label is lower than the short-term correlation coefficient of the idle label.

[0350] For platform A (current label L2, r2 = 0.4177): r1(7)(0.35) < r2(0.417), not satisfied.

[0351] For platform B (current label L3, r3=0.208): r1(7)(0.35)>r3(0.208), which satisfies the condition. Furthermore, among all existing labels, L3 has the smallest correlation coefficient r3=0.208. Therefore, the system makes the final decision: update the labels on platform B, replacing the existing third set of labels L3 with the idle first set of labels L1.

[0352] Significance of this step: This step is the final implementation of the update strategy. The key term "group of product tags with the lowest correlation coefficient" refers to the group of tags with the lowest correlation to user comments and potentially the worst effect among all currently pushed tags. Its significance lies in establishing a "survival of the fittest" update principle: prioritizing the replacement of the least effective tags to achieve the greatest overall improvement with minimal changes. The beneficial effect is ensuring that each update brings a definite and quantifiable improvement (the overall average correlation coefficient increases from 0.3125 to 0.3545), maximizing optimization efficiency.

[0353] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0354] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0355] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A commodity label generation system for multi-modal commodity information analysis, characterized by, Specifically, it includes: Data Acquisition Layer: Construct a distributed crawler cluster to parse competitors' product detail pages and extract competitors' multimodal product information; Multimodal Understanding Layer: Image Understanding: Extracts deep visual features from product main images and detail images, and uses the CLIP model finely tuned on e-commerce data to achieve zero-shot recognition of image-to-text tags; encodes product titles, descriptions, and reviews to obtain semantic vectors and extract product attributes; Competitive analysis layer: Aggregates the captured product attributes by product category, extracts keywords from competitor titles and reviews, and generates a competitive analysis feature vector for the product based on the keywords; Content generation layer: Using the multimodal feature vector of the product and the feature vector of the competitor analysis as conditions, the prompt words are concatenated and input into the large language model to guide the model to generate multiple sets of product tags for the product. Based on the similarity of the product tags of the product and combined with the data of the target platform, the push processing method of multiple sets of product tags is determined. Quality assessment and optimization layer: Analyze the review data of competitors and products to determine the correlation between product tags and user review data, and determine the push update method for the product tags by combining the idle status of product tags and push data; The method for determining the push update method for the product tags is as follows: Based on the idle status of product tags, determine the groups that have not been pushed out. Based on the push data, determine the target platform for available push groups that do not have similar groups, and use it as the update target platform; The method for pushing and updating product tags is determined based on the groups that have not undergone push processing and the target platform for updates, combined with the correlation with comment data.

2. The multi-modal commodity information analysis generated commodity label system of claim 1, wherein, Generate a competitive analysis feature vector for the product, specifically including: Based on competitors' titles and descriptions, marketing keywords are derived by sorting them using TF-IDF. Based on the keyword extraction results from competitors' review data, positive and negative keywords are generated; Based on the aforementioned marketing keywords, positive keywords, and negative keywords, a competitor analysis feature vector for the product is generated.

3. The method of claim 1 or 2, wherein the method is applied to the system of claim 1 or 2. Specifically, it includes: The system acquires multimodal product information, determines the semantic similarity coefficient between the product and its competitors based on the product information, determines the product tag generation scheme based on the semantic similarity coefficient between the product and its competitors on various push target platforms, and when the generation scheme requires the generation of multiple sets of product tags, it determines the reference competitors for generating tags among the competitors based on the comment data and product tags of the competitors. Based on the generated tags and all competing products, multiple sets of product tags for the products are generated. The push processing method for the multiple sets of product tags is determined according to the similarity of the product tags and the data of the target platform. The product tag push processing is performed according to the push processing method. The review data of competitors and products is analyzed to determine the correlation between product tags and user review data. The push update method of the product tags is determined by combining the idle status of product tags and push data.

4. The multi-modal commodity information analysis commodity label generation method of claim 3, wherein, The product information includes the product type and product images.

5. The multi-modal commodity information analysis-based commodity label generation method of claim 3, wherein, The semantic similarity coefficient between the product and its competitors is determined based on the average similarity coefficients of the product information in the multimodal context.

6. The multi-modal commodity information analysis commodity label generation method of claim 3, wherein, The method for determining the product label generation scheme is as follows: The number of competitors in different push target platforms is determined based on the semantic similarity coefficient between them and the competitors. The average semantic similarity coefficient between the product and the competitor on different push target platforms is determined by the semantic similarity coefficient between the product and the competitor on different push target platforms, and this average value is used as the reference similarity coefficient. Based on the number of competing products on different push target platforms and the reference similarity coefficient, the product tag generation scheme is determined.

7. The multi-modal commodity information analysis commodity label generation method of claim 6, wherein, Based on the number of competing products on different push target platforms and the reference similarity coefficient, a product tag generation scheme is determined, specifically including: The target platforms with a similarity coefficient greater than the preset similarity coefficient value are designated as similar target platforms. The weight value of the similar target platforms is determined based on the number of competing products in the similar target platforms. When the sum of the weight values ​​of the similar target platforms is greater than the preset weight threshold, the product tag generation scheme is determined to generate a preset number of product tags. Otherwise, the product tag generation scheme is determined to generate a second preset number of product tags.

8. The multi-modal commodity information analysis commodity label generation method of claim 3, wherein, Generate multiple sets of product tags for the aforementioned products, specifically including: Based on the competitor's tag generation reference, generate a corresponding set of product tags, and then combine all the competitor's tags to generate a corresponding set of product tags.

9. The multi-modal commodity information analysis commodity label generation method of claim 3, wherein, The method for determining the push processing method for the multiple sets of product tags is as follows: The product tags generated for the product are used as the generated product tags. The number of target platforms for pushing the product and the number of groups of generated product tags are obtained. Based on the generated product tags of the group, determine the similarity between the generated product tags of the group and other groups, and determine the similarity coefficient between the product tags of the group and other groups based on the similarity. Based on the number of target platforms for the product and the number of groups in which product tags are generated, and combined with the similarity coefficient between the generated product tags of the group and the product tags of other groups, the push processing method for the generated product tags of the group is determined.

Images