An intelligent commodity search system and method based on search word structured analysis

Abstract

This invention discloses an intelligent product search system and method based on structured search term analysis, belonging to the field of e-commerce data processing technology. By performing natural language processing on user-input search terms, multi-dimensional information is extracted and structured data is generated. Simultaneously, image recognition information is integrated to generate standardized image tags. Based on the integrated structured data, user search intent tags with multi-level weights are generated. Product tags are matched with user search intent tags, and dynamic tagging and weight adjustments are performed for products without corresponding preset tags. A personalized product recommendation list is then output. This invention can accurately understand user search intent, achieving precise tag-level matching between user needs and product features, effectively improving the accuracy and personalization of product searches.

Description

Technical Field

[0001] This invention relates to the field of e-commerce data processing technology, specifically to an intelligent product search system and method based on structured analysis of search terms. By performing structured analysis on user search terms and generating multi-dimensional intent tags, combined with multimodal fusion and personalized ranking, the system achieves precise matching between user search intent and product feature tags, effectively improving the accuracy of product search and user experience. Background Technology

[0002] In existing product search technologies, the mainstream approach typically employs keyword-based matching. Specifically, after a user enters a search term, the system breaks it down into several keywords through word segmentation, then performs keyword matching in the product database, returning a list of relevant products based on the degree of match. Additionally, some e-commerce platforms also incorporate historical user behavior data (such as clicks, favorites, and purchase history) to perform a simple ranking of search results to improve recommendation effectiveness.

[0003] In terms of product tag management, existing technologies mostly rely on manual pre-setting to build the tag system. Operations personnel manually assign several fixed tags to each product based on its category, material, brand, and other basic attributes. These tags serve as an important basis for product retrieval and recommendation. In addition, some platforms regularly update their tag libraries manually to adapt to the addition of new product categories and changes in user needs.

[0004] In the area of ​​multimodal search, existing technologies have begun to support image search functionality. When a user uploads a product image, the system uses image recognition technology to extract visual features from the image, identify information such as the product category, color, and style, then converts this information into a text description, and finally performs a search and matching based on text keywords.

[0005] However, the measures currently being taken have at least three problems: First, the search results are not highly relevant. Existing technology relies solely on keyword matching, which cannot accurately understand the user's true intent. For example, if a user enters "Nike sneakers," the system may not be able to distinguish whether the user wants to buy sneakers of a specific style, price range, or function, thus returning a large amount of irrelevant product information.

[0006] Second, the multimodal search capability is weak. For multimodal search methods such as image search, existing technologies simply convert image recognition results into text keywords, failing to deeply integrate visual information with the user's search intent. This results in insufficient recognition and matching accuracy, and the value of multimodal information is not fully realized.

[0007] Third, the labeling system is rigid. Product labels rely heavily on manual pre-setting, which is time-consuming and prone to inconsistencies due to subjective experience. Manual labeling only covers basic attributes such as category and material, making it difficult to uncover deeper product characteristics. The label library lacks an automatic update mechanism, making it unable to adapt to new product types and changes in user needs, and failing to achieve accurate label-level matching between users and products. Summary of the Invention

[0008] To address the three issues mentioned above, the purpose of this invention is to provide an intelligent product search system and method based on structured analysis of user search terms. By performing structured analysis on user search terms and generating multi-dimensional intent tags, combined with multimodal fusion and intelligent tagging technology, the system achieves accurate matching between user search intent and product feature tags, thereby improving the accuracy and personalization of product search.

[0009] This was achieved through the following technical solutions: An intelligent product search system based on structured search term analysis includes: The structured analysis module is used to perform deep semantic analysis on the search terms input by the user using natural language processing (NLP) technology, extract multi-dimensional information, and generate structured data. Multimodal fusion module: used to extract visual information from images and convert it into standardized image labels, then fuse the standardized image labels with the structured data output by the structured analysis module to generate fused structured data; Intelligent tagging module: Connected to the structured analysis module and the multimodal fusion module, it generates user search intent tags based on the fused structured data and assigns weights to each tag; it matches the pre-labeled tags in the product library with the user search intent tags; API call module: Connected to the intelligent tagging module, it is used to call the backend search API to retrieve products. The retrieval criteria include user search intent tags and product feature tags, and outputs a preliminary product retrieval list. Personalized sorting module: Connected to the API call module, it sorts the initial product search list based on the matching degree between user search intent tags and product feature tags, as well as user historical preference data; Output module: Retrieves the sorted product search results and outputs a product list.

[0010] Optionally, the multi-dimensional information extracted by the structured analysis module includes at least one of the following: question relevance, keywords, brand, price range, functional requirements, usage scenarios, and audience characteristics. Through multi-dimensional information extraction, the user's search intent can be comprehensively captured, providing a rich data foundation for subsequent accurate matching and avoiding search result biases caused by missing information in traditional keyword matching.

[0011] Optionally, the structured analysis module execution process includes: word segmentation and part-of-speech tagging based on a customized e-commerce lexicon, named entity recognition fine-tuned from e-commerce corpus, semantic understanding and intent classification combined with a predefined intent classification system, and the step of encapsulating the extracted multi-dimensional information into structured data. This enables deep semantic analysis of user search terms, transforming unstructured natural language into structured machine-readable data, ensuring the accuracy and completeness of input data for subsequent intelligent tagging stages.

[0012] Optionally, the user search intent tags generated by the intelligent tagging module include primary core tags, secondary attribute tags, and tertiary personalized tags, and each tag is assigned a weight. Through a multi-level tagging system and weight configuration, a refined characterization of user intent can be achieved, capturing both the core needs of users and taking into account fine-grained personalized preferences, thereby improving the richness and accuracy of tag expression.

[0013] Optionally, the intelligent tagging module uses a hierarchical weighted cosine similarity algorithm to match the pre-labeled tags in the product database with the user's search intent tags, and outputs a matching score of 0%-100%.

[0014] Optionally, the intelligent tagging module is also used to dynamically supplement tags for products without corresponding preset tags, based on user search intent tags and product details information, while simultaneously adjusting the weights of existing product tags in real time. These supplementary tags are temporary tags that participate in subsequent product tag matching and weight adjustments. These supplementary tags will either become official tags or be eliminated as the frequency of user search behavior increases. This dynamic supplementary tagging mechanism solves the problems of rigidity and inability to cover new products in traditional tag systems. Real-time weight adjustments enable product tags to dynamically adapt to the user's current search needs, improving the timeliness and accuracy of tag matching.

[0015] Optionally, the personalized sorting module combines user historical behavior and preference data, integrating the matching degree and weight of user search intent tags and product feature tags to personalize the sorting of product search results. This prioritizes the display of products that highly match the user's current search intent, while also taking into account the user's long-term preferences and the real-time status of the products, achieving a balance between personalization and timeliness.

[0016] Furthermore, a smart product search method based on structured analysis of search terms is proposed based on the aforementioned system, including the following steps: Natural language processing is performed on the search terms entered by the user to extract multi-dimensional information and generate structured data; Visual information is extracted from images and converted into standardized image tags. These standardized image tags are then fused with structured data to generate fused structured data. User search intent tags are generated based on the fused structured data, and product tags are matched with user search intent tags to obtain fused tag data; Based on the fused tag data, the backend search API is called to retrieve products and output a preliminary product search list; the search criteria include user search intent tags and product feature tags. The preliminary product search list is sorted based on the matching degree between user search intent tags and product feature tags, as well as user historical preference data. Retrieve the sorted product search results and output a product list.

[0017] The beneficial effects of this invention compared to the prior art are: 1) First, build a full-link intelligent search system.

[0018] By performing structured analysis of user search terms, multimodal information fusion, intelligent tagging, and dynamic weight adjustment, a complete closed loop is formed from understanding user intent to matching product tags and then to personalized ranking, solving the core problems of low search result relevance, rigid tag system, and weak multimodal search capabilities in existing technologies.

[0019] 2) Second, achieve dynamic self-adaptation of the labeling system.

[0020] Through intelligent tagging technology, the system can dynamically generate user intent tags, accurately match and dynamically supplement product tags, and adjust tag weights in real time based on the matching degree. This allows the tag library to automatically iterate and update with user search behavior, adapting to the massive number of products on e-commerce platforms and the diverse and dynamic search needs of users.

[0021] 3) Third, construct a weight-ranking collaborative mechanism to improve search accuracy and personalized experience.

[0022] The intelligent tagging module dynamically adjusts the weight of product tags based on the matching degree between user intent and product tags, enabling product tags to adapt to user needs in real time. The personalized sorting module calculates the comprehensive matching degree based on the adjusted tag weights and uses this matching degree as the core sorting criterion. Combined with user historical preferences and real-time product operation data (such as conversion rate, positive review rate, and inventory adequacy rate), the module uses a multi-level priority sorting logic to ensure that products that highly match the user's current search intent are displayed first. This achieves a complete closed loop from tag optimization to sorting output, significantly improving the relevance of search results and user satisfaction.

[0023] 4) Fourth, enhance multimodal search fusion capabilities.

[0024] By converting image recognition information into standardized tags and deeply integrating them with text search intent tags, the problem of "easy recognition but difficult matching" of multimodal information is solved, significantly improving the matching accuracy and practicality of multimodal search.

[0025] 5) Improve system operating efficiency and business value.

[0026] By calling backend APIs using structured JSON format data based on integrated tag data, invalid requests and search results are reduced, thus lowering system resource consumption. At the same time, the intelligent tag system provides e-commerce platforms with refined data support for product operation and user segmentation, helping the platform achieve accurate recommendations and marketing campaign pushes. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the module architecture of an intelligent product search system based on structured analysis of search terms according to the present invention.

[0028] Figure 2 This is a flowchart illustrating an intelligent product search method based on structured analysis of search terms according to the present invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] like Figure 1 As shown, an intelligent product search system based on structured search term analysis has an architecture comprising: a structured analysis module, a multimodal fusion module, an intelligent tagging module, an API call module, a personalized ranking module, and an output module. The system performs structured analysis on user search terms, integrates multimodal information from text and images, generates multi-level user search intent tags, and accurately matches these tags with product tags, achieving end-to-end intelligent search from understanding user intent to product retrieval and personalized ranking. The following provides a detailed description of each module: The structured analysis module employs a customized NLP process optimized for e-commerce scenarios to perform deep semantic analysis on user-input search terms, extract multi-dimensional information, and generate structured data. The workflow of the structured analysis module includes: (1) Data preprocessing: Clean the search terms entered by the user, remove invalid characters, special symbols, extra spaces, etc., and unify the text format; (2) Word segmentation and part-of-speech tagging: Based on a customized e-commerce thesaurus, the jieba framework was used to segment the preprocessed search terms and tag the part of speech of each word. The jieba framework is a third-party Chinese word segmentation library that supports custom thesaurus expansion and can accurately segment a text into Chinese words; the customized e-commerce thesaurus includes common brand words, category words, attribute words, etc. in the e-commerce field to improve the accuracy of word segmentation. (3) Named entity recognition: A lightweight version of BERT-Base model fine-tuned with e-commerce corpus is used to identify named entities in search terms, including brand entities (such as "Nike"), category entities (such as "sports shoes"), and functional entities (such as "shock absorption"). Among them, the BERT-Base model is a pre-trained language model based on the Transformer architecture, which can capture the contextual semantic information in the text and is suitable for handling complex situations such as abbreviations, aliases, and colloquial expressions that may exist in user search terms. After being fine-tuned with e-commerce corpus, the model can accurately identify key entities such as brands, categories, and functions in e-commerce scenarios. (4) Semantic understanding and intent classification: Semantic matching is achieved by combining the SimBERT model to perform deep semantic understanding of search terms and determine the user's search intent type; among which, the SimBERT model is a BERT-based semantic matching model that focuses on calculating the semantic similarity between texts and is suitable for matching user search terms to predefined intent categories; the user search intent classification system includes: Purchase intents: direct purchase, precise selection (specific brand / specification / function), price comparison selection, bulk purchase; Information-related intents: product parameter query, usage instructions, after-sales policy consultation, brand background information; Reference-based purposes: to find similar / identical items, to refer to popular recommendations, to check customer reviews, and to get styling suggestions.

[0031] (5) Information Structuring: The multi-dimensional information extracted in the above steps is structured to achieve deep semantic analysis of user search terms, and unstructured natural language is transformed into structured machine-readable data (JSON format). The extracted multi-dimensional information includes at least one of the following: question relevance, keywords, brand, price range, functional requirements, usage scenarios, and audience characteristics.

[0032] By extracting information from multiple dimensions, we can fully capture users' search intent, providing a rich data foundation for subsequent accurate matching. This ensures that the input data in the subsequent intelligent tagging process is accurate and complete, avoiding the search result deviations caused by missing information in traditional keyword matching.

[0033] Multimodal fusion module: Receives images uploaded by users, extracts visual information from the images and converts it into standardized image tags, then fuses the standardized image tags with structured data to generate fused structured data, which serves as input to the intelligent labeling module.

[0034] The multimodal fusion module employs a multimodal model fusion architecture combining a lightweight convolutional neural network (CNN) and contrastive language-image pre-training (CLIP). Simultaneously, the neural network uses the MobileNetV3-Large model for basic image feature extraction, extracting the low-level visual features of the image (such as color, texture, contour, and material). A CLIP-ViT-B / 32 model, finely tuned with e-commerce product corpora, is used for fine-grained recognition, focusing on enhancing the recognition capabilities of product categories, brand logos, and style features, thus addressing the issue of low accuracy in brand and specific style recognition by pure CNN models. The training data for the MobileNetV3-Large and CLIP-ViT-B / 32 models are based on the product image libraries owned by e-commerce platforms and publicly available e-commerce product datasets. The images cover all e-commerce product categories, including cosmetics, apparel, digital products, and food. Each image is annotated in multiple dimensions, including category, brand, color, style, function, and material. After optimization, the recognition time for a single image is ≤100ms, making it suitable for real-time online search needs.

[0035] The identified multi-dimensional labels adopt the naming rule of "first-level category label_second-level sub-label_core attribute label_feature description label" and are uniformly encoded in JSON format, such as "sports shoes_running shoes_Nike_white and blue color blocking_shock absorption and breathability" and "beauty_lipstick_Dior_shade 999_matte". The standardized image label is encoded using JSON key-value pairs. The top-level key is "image visual label", and the lower level is divided into second-level keys according to "category level", "core attribute", "visual feature" and "brand identity". The encoding format must be fully compatible with the JSON format structured data output by the structured analysis module to ensure uniform format and smooth connection when fusion of multimodal information.

[0036] Subsequently, the standardized image tags are integrated with the structured data output from the structured analysis module. The standardized image tags are hierarchically divided into primary categories / brands (core tags), secondary attributes / styles (attribute tags), and tertiary visual feature descriptions (personalized tags), with initial weighting rules (core 80%, attribute 15%, personalized 5%). A lightweight self-attention mechanism is then introduced, using the "matching degree between user search intent tags and image tags" as the weighting basis to dynamically allocate weights to each level of tags in both image tags and text structured data. This involves increasing the attention weight of tags that overlap between text and images, retaining the initial weight of core tags that are not mentioned in the text but are highly recognizable in the image, and treating them as supplementary intent tags. Conversely, the attention weight of low-recognition, blurry visual tags in the image is reduced to minimize interference from invalid information.

[0037] During the integration process, the core tag layer (brand / category) uses text tags as the core reference, with image tags serving as verification and supplements. If there is a conflict between the two, the text tags take precedence. In the attribute tag layer (function / scenario / style), text and image tags have equal weight, and the attention mechanism is dynamically adjusted based on feature recognition. The personalized tag layer (color / material / audience) uses image visual tags as the core reference, with text tags serving as supplements, to meet the user's core need of "searching for the same / similar styles by looking at images".

[0038] Finally, the image tags adjusted by the attention mechanism and the text structured data are integrated in a unified JSON format to form fused structured data containing "tag level, tag content, fused weight, and information source (text / image / fusion)," which serves as the input for the intelligent tagging module.

[0039] The intelligent tagging module connects to the structured analysis module and the multimodal fusion module, generating user search intent tags based on the fused structured data. It matches pre-labeled tags from the product database with these user search intent tags. Specifically, the intelligent tagging module uses a pre-defined tag lexicon combined with a deep learning tag model to generate multi-level tags for user search intent. The pre-defined tag lexicon is a multi-level structured lexicon specific to the e-commerce field, dynamically expanding in scale. The initial version covers mainstream e-commerce categories (beauty, apparel, digital products, home furnishings, food, bags, etc.), with a total vocabulary exceeding 20,000 entries, and is further subdivided by category for lexicon management.

[0040] The tag database employs a hybrid update mechanism: high-frequency words are added weekly based on user search behavior and new product data mining. These new high-frequency words undergo manual review to confirm their standardization and scenario suitability before being officially incorporated into the tag database. It also supports emergency real-time updates for major promotions, new product launches, or trending keywords. Furthermore, the tag database evaluates tag activity quarterly, downgrading or removing tags that have not been searched for a long time or are not associated with products.

[0041] The intelligent labeling module outputs user intent labels based on a deep learning labeling model, employing an end-to-end framework of "BiLSTM + Attention + multi-layer fully connected layers". The fused structured data is vectorized by concatenating textual semantic features with image visual features to form the input feature vector for the deep learning labeling model. The core layer of the model uses BiLSTM to capture the contextual semantic relationships of the input features and the correlation between multimodal features. A self-attention layer strengthens the feature representation of core intent features and weakens ineffective redundant features. Finally, fully connected layers and a Softmax layer achieve hierarchical classification and initial weight allocation of the labels. The model outputs a multi-level set of user search intent labels, including label hierarchy, label content, and initial weight values.

[0042] The model training uses user search logs (text + images) annotated by the platform as the core data. Each sample contains user search input, fused structured data, and manually annotated multi-level intent tags, with a total training sample size exceeding 5 million. The generated user search intent tags include primary core tags (e.g., brand: Nike, category: athletic shoes), secondary attribute tags (e.g., function: shock absorption, scenario: running), and tertiary personalized tags (e.g., target audience: teenagers, price expectation: mid-to-high-end), and each tag is assigned a weight. The core tag weight is 80%, the attribute tag weight is 15%, and the personalized tag weight is 5%.

[0043] The intelligent tagging module also employs a hierarchical weighted cosine similarity algorithm to calculate the matching degree between user search intent tags and preset product tags, outputting a matching score ranging from 0% to 100%. Specifically, both user search intent tags and product tags are converted into word vectors, and then the cosine similarity of the first, second, and third level tags is calculated separately. Finally, a weighted sum is performed based on the preset weights of each level of tags to obtain the overall matching degree. If a certain level does not have a corresponding tag, the weight of that level is proportionally distributed to other levels to ensure that the total weight is 1.

[0044] When a product lacks preset tags that correspond to a user's search intent, the intelligent tagging module dynamically adds tags based on the user's search intent tags and product details (including parameters, functions, scenarios, etc.). It also adjusts the product tag weights in real time according to the tag matching degree, generates supplementary tags, marks them as temporary tags and incorporates them into the product tag system, and participates in subsequent product tag matching and weight adjustment. The supplementary tags will be converted into formal tags or eliminated as the frequency of user search behavior increases.

[0045] The intelligent tagging module combines user search intent tag weights, tag matching degrees, and real-time product interaction data (clicks, favorites, transactions, etc.) to adjust the weights of existing product tags in real time: When the matching degree is ≥80%, the weight is increased by a fixed 40% based on the user search intent tag level weight, ensuring that the weight of a single tag does not exceed the maximum weight percentage of that level; when the matching degree is between 30% and 80%, the product weight remains unchanged; if the product has multiple user clicks or favorites within the past hour, the matching tags are increased by a fixed 5%; when the matching degree is ≤30%, the weight of the product's matching tags is decreased by a fixed 30%, ensuring that the weight of a single tag is not less than 0.5%. If multiple tags match at the same level, the adjustment ratio is allocated according to the matching degree percentage of each tag at its single level, ensuring the rationality of tag weight adjustments within the same level.

[0046] Weight adjustments take effect immediately, directly connecting to subsequent API retrieval and personalized ranking. For the same product and the same tag, only one weight adjustment is allowed per minute to avoid weight fluctuations caused by high-frequency user behavior. If a product has no matching user search intent for 24 consecutive hours, the tag weight will be reset to the initial preset weight.

[0047] Through a multi-level tagging system, dynamic tagging mechanism, and real-time weighting adjustment, the intelligent tagging module achieves a refined depiction of user intent. It can not only grasp the core needs of users, but also take into account fine-grained personalized preferences, thereby improving the richness and accuracy of tag expression. It also solves the problem of the rigidity of traditional tagging systems and their inability to cover new products, enabling product tags to dynamically adapt to the user's current search needs and improve the timeliness and accuracy of tag matching.

[0048] API Call Module: Connects with the intelligent tagging module, merges the user search intent tags and product feature tags output by the intelligent tagging module, and encapsulates them into standardized JSON format fused tag data. The fused tag data is used as the core retrieval basis to call the backend search API to retrieve products, replacing the traditional single keyword matching method.

[0049] The structured JSON data of the integrated tag data adopts a hierarchical nested structure, containing three core modules: basic retrieval information, core integrated tag information, and weight configuration. The core integrated tag information is organized hierarchically into first-level core tags, second-level attribute tags, and third-level personalized tags. Each tag includes a tag name, tag type, weight value, and tag source (text / image / integrated). When the tag array is empty, the API returns a general category recommendation; after a timeout, it returns a downgraded search result (keyword matching result) to adapt to the real-time search experience requirements of e-commerce.

[0050] The data returned by the backend search API includes three main modules: request and response information, core product retrieval data, and pagination information. The core product retrieval data includes the product's unique identifier, name, price, brand, category, and the product's total match score with the integrated tags, as well as tag matching details. The initial product search list returned by the API will serve as input for the personalized sorting module.

[0051] Calling APIs using structured JSON data based on integrated tag data can reduce invalid API requests and search results, lower system resource consumption, and improve the operating efficiency of the backend search system.

[0052] Personalized sorting module: Connects to the API call module, obtains the preliminary product search list returned by the API call module, and combines it with user historical preference data. It integrates the matching degree and weight of user search intent tags and product feature tags to personalize the product search results, prioritizing the display of products with high tag matching and those that align with user historical preferences. This achieves a triple sorting logic of "structured analysis + tag matching + personalized preference." The criterion for high tag matching degree is a matching degree ≥ 0.7.

[0053] User historical preference data comes from the e-commerce platform's user behavior tracking system, extracting valid behavioral data from the past 90 days. This data is categorized by behavioral value into core behaviors (e.g., purchases, favorites, add-to-cart), important behaviors (e.g., prolonged browsing, precise clicks, viewing reviews), and general behaviors (e.g., short-term browsing, category page turning, clicking on related recommendations). The data is aggregated based on product dimensions, extracting user preferences for brands, categories, functions, price ranges, styles, and other characteristics to form a personalized user preference tag library.

[0054] The weights of user preference tags are calculated using a combination of behavior type weighting, time decay, and frequency normalization. Higher weight values ​​indicate a stronger user preference for the tag. The user preference tag library employs a real-time incremental update mechanism, updating the tag library within 5 minutes of a user generating a new valid behavior, and cleaning historical data every 30 days.

[0055] Personalized ranking employs a combination of fixed base proportions and dynamic fine-tuning based on specific scenarios. Under the fixed base proportions, tag matching accounts for 60%, user historical preference accounts for 30%, and real-time product operational data accounts for 10%. Dynamic fine-tuning is applied across different search scenarios: for precise searches, tag matching weight is increased to 70%, and user historical preference weight is decreased to 20%; for generalized searches, user historical preference weight is increased to 40%, and tag matching weight is decreased to 50%; for new or anonymous users, tag matching weight is increased to 80%, and operational data weight is increased to 20%. Corresponding fine-tuning is also applied to different product category characteristics: for categories with high brand sensitivity, brand tag weight is increased; for categories with high personalization sensitivity, user historical preference weight is increased; and for standardized product categories, operational data weight is increased.

[0056] The final score is calculated using a weighted summation method, and the system sorts the products from highest to lowest score. The sorted product list is then output to the output module.

[0057] Output module: Used to obtain the product search results returned by the backend search API and output a product list sorted by the personalized sorting module.

[0058] Furthermore, a smart product search method based on structured analysis of search terms is proposed, such as... Figure 2 As shown, it includes the following steps: Step S1: The user enters a search term, and the system determines whether the search term contains an image. If the user input contains an image, proceed to step S2; otherwise, proceed directly to step S3.

[0059] Step S2: Image recognition and information extraction, converting images into standardized image tags. The system performs feature recognition and content analysis on user-uploaded images, extracts visual information such as product brands, styles, and functions from the images, and converts them into standardized image tags.

[0060] Step S3: NLP Semantic Parsing and Structured Analysis Natural language processing is performed on the user's search terms to extract multi-dimensional information and generate structured data. The extracted multi-dimensional information includes at least one of the following: question relevance, keywords, brand, price range, functional requirements, usage scenarios, and audience characteristics.

[0061] Step S4: Multi-channel information integration The standardized image labels generated in step S2 are merged with the structured data generated in step S3 to generate merged structured data.

[0062] Step S5: Generate user search intent tags Based on the fused structured data, user search intent tags are generated using a pre-defined tag library and a deep learning tag model. The generated user search intent tags include primary core tags, secondary attribute tags, and tertiary personalized tags, with each tag assigned a weight.

[0063] Step S6: Product Feature Tag Matching and Dynamic Labeling The system matches pre-defined tags in the product database with user search intent tags. For products without corresponding pre-defined tags, it dynamically adds tags based on user search intent tags and product details, generating supplementary tags and incorporating them as temporary tags into the product tagging system.

[0064] Step S7: Real-time adjustment of tag weights The weight of product tags is adjusted in real time based on the match between user search intent tags and product tags. The specific adjustment rules are as follows: When the match rate is ≥80%, the weight is increased by 40% of the weight corresponding to the intent tag, based on the original weight; when the match rate is between 30% and 80%, the product tag weight remains unchanged, but if the product has been clicked or favorited multiple times within the past hour, the matching tag weight is increased by a fixed 5%; when the match rate is <30%, the product tag weight is decreased by 30%. The adjusted product tag weight will serve as an important basis for the personalized ranking module to calculate the overall match rate.

[0065] Step S8: Encapsulate JSON data and call the backend search API. By merging user search intent tags with product feature tags and encapsulating them into standardized JSON format data, the system accurately calls the backend product search API to obtain a preliminary product search list.

[0066] Step S9: Personalized sorting By combining users' historical preference data and integrating the matching degree and weight of users' search intent tags and product feature tags, the initial product search results are personalized and prioritized for displaying products with high tag matching degree and in line with users' historical preferences.

[0067] Step S10: Output the final search results Retrieve the sorted product search results and output a product list.

[0068] In summary, this invention constructs a full-link intelligent search system—from understanding user intent to matching product tags and then to personalized ranking—through structured analysis of user search terms, multimodal information fusion, intelligent tagging, and dynamic weight adjustment. This solution effectively addresses the problems of low search result relevance, rigid tagging systems, and weak multimodal search capabilities in existing technologies. It significantly improves the relevance and accuracy of search results, optimizes the tagging system for dynamic adaptation, strengthens multimodal search fusion capabilities, upgrades the personalized recommendation experience, enhances commercial conversion rates and refined operational capabilities, and improves backend API call efficiency. Therefore, compared with existing technologies, this invention represents a significant advancement.

[0069] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. An intelligent product search system based on structured analysis of search terms, characterized in that, include: The structured analysis module is used to perform deep semantic analysis on the search terms input by the user using natural language processing (NLP) technology, extract multi-dimensional information, and generate structured data. Multimodal fusion module: used to extract visual information from images and convert it into standardized image labels, then fuse the standardized image labels with the structured data output by the structured analysis module to generate fused structured data; Intelligent tagging module: Connected to the structured analysis module and the multimodal fusion module, it generates user search intent tags based on the fused structured data and assigns weights to each tag; it matches the pre-labeled tags in the product library with the user search intent tags; API call module: Connected to the intelligent tagging module, it is used to call the backend search API to retrieve products, where the retrieval criteria include user search intent tags and product feature tags; Personalized sorting module: Connected to the API call module, it sorts product search results based on the matching degree between user search intent tags and product feature tags, as well as user historical preference data; Output module: Used to obtain the product search results returned by the backend search API and output a product list.

2. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The multi-dimensional information extracted by the structured analysis module includes at least one of the following: problem relevance, keywords, brand, price range, functional requirements, usage scenarios, and audience characteristics.

3. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The structured analysis module execution process includes: word segmentation and part-of-speech tagging based on a customized e-commerce lexicon, named entity recognition fine-tuned from e-commerce corpus, semantic understanding and intent classification combined with a predefined intent classification system, and the step of encapsulating the extracted multi-dimensional information into structured data.

4. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The intelligent tagging module generates user search intent tags, including primary core tags, secondary attribute tags, and tertiary personalized tags, and assigns weights to each tag.

5. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The intelligent tagging module uses a hierarchical weighted cosine similarity algorithm to match the pre-labeled tags in the product database with the user's search intent tags and outputs a matching score of 0%-100%.

6. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, For products without corresponding preset tags, the intelligent tagging module dynamically adds tags based on user search intent tags and product details information. The added tags are marked as temporary tags to participate in subsequent product tag matching and weight adjustment, and are converted into formal tags or eliminated based on the frequency of user search behavior.

7. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The intelligent tagging module is also used to adjust the weight of product tags in real time based on the matching degree between search intent tags and preset product tags.

8. The intelligent product search system based on structured search term analysis according to claim 1, characterized in that, The personalized ranking module combines users' historical behavior and preference data, and integrates the matching degree and weight of users' search intent tags and product feature tags to personalize the ranking of product search results.

9. An intelligent product search method based on structured search term analysis, operating using the intelligent product search system as described in any one of claims 1 to 8, characterized in that, Includes the following steps: Natural language processing is performed on the search terms entered by the user to extract multi-dimensional information and generate structured data; Visual information is extracted from images and converted into standardized image tags. These standardized image tags are then fused with structured data to generate fused structured data. User search intent tags are generated based on the fused structured data, and product tags are matched with user search intent tags to obtain fused tag data; Based on the fused tag data, the backend search API is called to retrieve products and output a preliminary product search list. The retrieval criteria include user search intent tags and product feature tags; The preliminary product search list is sorted based on the matching degree between user search intent tags and product feature tags, as well as user historical preference data. Retrieve the sorted product search results and output a product list.

Images