Fashion Item Recommendation Content Generation System

Abstract

Prompts are automatically generated from fashion product attribute information extracted through an artificial intelligence engine from existing information regarding fashion products, and these prompts are provided to a large language model to generate fashion product recommendation content from the results. Category and style information of fashion products can be extracted from product images based on a classification-type AI engine. Product attribute information can be extracted from various multimedia-based web documents regarding fashion products through an entity detection-based AI engine.

Description

Technology Field

[0001] Information processing technology, particularly artificial intelligence technology that automatically generates content recommending fashion products, is being disclosed. Background Technology

[0002] In addition to basic product information, online shopping malls provide recommendation content. Particularly for fashion products, this recommendation content offers information in a friendly tone, such as suitability for specific body types and age groups, as well as appropriate for specific times, places, and occasions. Since this recommendation content significantly impacts sales, shopping mall operators and fashion companies hire experts to create it, investing substantial resources in preparing content for a wide range of fashion products. The problem to be solved

[0003] The proposed invention aims to automatically generate fashion product recommendation content.

[0004] Furthermore, the proposed invention aims to enhance convenience by reducing the information input for creating fashion product recommendation content through the automatic extraction of fashion product attribute information from existing product information. means of solving the problem

[0005] According to one aspect of the proposed invention, fashion product recommendation content is generated by providing an automatically generated prompt to a large language model. Attribute information of fashion products necessary for generating the prompt is automatically extracted from existing information regarding fashion products through an artificial intelligence engine.

[0006] According to an additional aspect, attribute information of fashion products is extracted from product images based on a classification AI engine and can be reflected in a prompt requesting the generation of product recommendation content using a large language model.

[0007] According to additional aspects, category information and style information of fashion products can be extracted by two AI engines, each trained on the corresponding information, using a classification AI engine with the same structure.

[0008] Depending on additional aspects, characters found in product images can be extracted and reflected in product recommendation content.

[0009] Depending on additional aspects, product attribute information can be extracted from various multimedia-based documents regarding fashion products and reflected in product recommendation content. Depending on auxiliary aspects, an entity detection-based artificial intelligence engine can be applied to extract product attribute information from these multimedia-based documents. Effects of the invention

[0010] According to the proposed invention, recommendation content for fashion products can be automatically generated. Sufficient information about the fashion products necessary for content creation is provided through known fashion product detection or recognition models and attribute models, and information about the style of the fashion products is also provided. Consequently, effective recommendation content tailored to the relevant fashion products can be created using a large language model, while providing generally natural content. This allows for a reduction in the time and cost required to create recommendation content. Brief explanation of the drawing

[0011] FIG. 1 is a block diagram illustrating the configuration of a fashion product recommendation content generation system according to one embodiment. FIG. 2 is a block diagram illustrating the configuration of one embodiment of the product information extraction unit in FIG. 1. Specific details for implementing the invention

[0012] The foregoing and additional aspects are embodied through embodiments described with reference to the accompanying drawings. It is understood that the components of each embodiment may be combined in various ways within the embodiment or with components of other embodiments, unless otherwise stated or contradicted. Based on the principle that the inventor may appropriately define the concepts of terms to best describe his invention, the terms used in this specification and claims shall be interpreted in a meaning and concept consistent with the description or proposed technical idea. Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0013] <Description of Claim 1>

[0014] According to one aspect of the proposed invention, fashion product recommendation content is generated by providing an automatically generated prompt to a Large Language Model (LLM). Several Large Language Models, such as BERT, GPT, LaMDA, and OPT, are provided to developers and the general public via APIs. Attribute information of fashion products necessary for generating the prompt is automatically extracted from existing information regarding fashion products through an artificial intelligence engine.

[0015] FIG. 1 is a block diagram illustrating the configuration of a fashion product recommendation content generation system according to one embodiment. As illustrated, the fashion product recommendation content generation system according to one embodiment includes a seller interface (300), a product information extraction unit (100), a content prompt generation unit (510), a style prompt generation unit (530), and a large language model interface (700).

[0016] The seller interface (300) receives information regarding a product to be used to generate recommendation content from the seller and outputs the generated content. Additionally, the seller interface (300) receives information regarding the style of the recommendation content from the seller and outputs the generated content. In one embodiment, the seller interface (300) is implemented as a graphical user interface, and the user can input an image of a fashion product by dragging it, input a URL of a web document containing information about the fashion product, or directly input information regarding the content or style of the product recommendation content.

[0017] The product information extraction unit (100) receives product image and product detail information in text form as input and processes it using artificial intelligence to extract product attribute information including the category and style of the fashion product. In one embodiment, the function of extracting product attribute information from the product image and the function of extracting product attribute information from product detail information in text form can be implemented as separate artificial intelligence engines. The product attribute information is attribute information required for generating fashion product recommendation content and may include information such as the category, style, functional attribute, expressive value attribute, aesthetic attribute, and design attribute of the fashion product. According to standard classification, the category information has a hierarchical structure of medium, small, and detailed classifications under major classifications of clothing, shoes, bags, accessories, and hats, and the category can be determined according to this hierarchical structure.

[0018] The content prompt generation unit (510) generates a prompt requesting a Large Language Model (LLM) to create fashion product recommendation content from product information entered through the seller interface (300) and product attribute information extracted from the product information extraction unit (100), and generates a content prompt that includes information regarding the content of the fashion product recommendation content to be created. Here, the content of the product recommendation content may include recommendations regarding product categories, brand introductions, style-centered recommendations, recommendations regarding TPO (Time Place Object) among product attributes, washing information, comments on protective functions such as durability and windproofing, etc.

[0019] In one embodiment, the content prompt generation unit (510) may refer to the content instruction template database (31). The content instruction templates may be prepared according to the categories of fashion products. Additionally, the content instruction templates may be prepared according to the attributes of the products. In one embodiment, the content prompt generation unit (510) may extract a content instruction template corresponding to the category of the fashion product and the style of the fashion product from the content instruction template database (31) among the fashion product information output from the product information extraction unit (100), and fill the template with the product attribute information output from the product information extraction unit (100) and the product information entered through the seller interface (300) to generate a content instruction prompt.

[0020] There are known tools that automatically generate prompts for large language models. For example, a tool called RAG (Retrieval-Augmented Generation) by LangChain converts user-specific knowledge—which is difficult to train a large language model in a short period—into a vector format that the model can understand and stores it in its own vector database. It then converts user queries into vectors of the same format and extracts query-related vectors from the database based on, for instance, cosine similarity, to generate prompts that serve as input for the large language model.

[0021] In another embodiment, the content prompt generation unit (510) can load the collected fashion product recommendation content into a prompt automatic generation tool such as RAG according to fashion product category or fashion product style and store it in a vector database, and generate a content instruction prompt by providing the product information output from the product information extraction unit (100) and the product information entered through the seller interface (300) to the prompt automatic generation tool.

[0022] The style prompt generation unit (530) generates a prompt requesting a large language model to create fashion product recommendation content from product information entered into the seller interface (300) and product attribute information obtained from the product information extraction unit (100), and generates a style prompt that includes information regarding the style of the fashion product recommendation content to be created. Here, the style information refers to the style of the expression of the content, and may include, for example, the format of endings such as ~yo, ~hae in recommendation text, and a writing style that takes into account the gender and age of the target customers of the fashion product.

[0023] In one embodiment, the style prompt generation unit (530) may refer to the style instruction template database (33). Style instruction templates may be prepared for each category of fashion products. Additionally, style instruction templates may be prepared for each attribute of the product. In one embodiment, the style prompt generation unit (530) may extract a style instruction template corresponding to the category of the fashion product and the style of the fashion product from the style instruction template database (33) among the fashion product information output from the product information extraction unit (100), and fill the template with the product attribute information output from the product information extraction unit (100) and the product information entered through the seller interface (300) to generate a style instruction prompt.

[0024] In another embodiment, the style prompt generation unit (530) loads the collected fashion product recommendation content into a prompt automatic generation tool such as RAG according to the fashion product category or the style of the fashion product and stores it in a vector database, and can generate a style instruction prompt by providing the product information output from the product information extraction unit (100) and information regarding the style of the content input through the seller interface (300) to the prompt automatic generation tool.

[0025] The large language model interface (700) provides the generated content prompt and style prompt to the large language model, receives the response, and outputs it to the seller interface as fashion product recommendation content. An interface for developers, such as an API, is provided for each large language model. The large language model interface (700) may include such known APIs.

[0026] <Description of Claim 2>

[0027] FIG. 2 is a block diagram illustrating the configuration of an embodiment of the product information extraction unit in FIG. 1. According to an additional aspect, attribute information of a fashion product is extracted from a product image based on a classification type artificial intelligence engine and can be reflected in a prompt requesting the creation of product recommendation content using a large language model. In the illustrated embodiment, the product information extraction unit (100) may include a classification-based fashion product attribute extraction unit (110). The classification-based fashion product attribute extraction unit (110) receives a product image as input and extracts category information and style information of the fashion product based on a classification type artificial intelligence engine.

[0028] <Description of Claim 3>

[0029] According to additional aspects, category information and style information of fashion products can be extracted by two AI engines, each trained on the corresponding information, using a classification AI engine with the same structure.

[0030] In one embodiment, the classification AI engine may be the model described in Ge, et al. “Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images”, Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2019. This classification AI engine utilizes a Mask R-CNN model, in which fashion product regions are first detected in an image, and features are extracted using a convolutional neural network. However, while this model is known to be top-tier in item detection, it is difficult to apply to fields requiring real-time prediction due to this type of feature extraction.

[0031] In another embodiment, the classification AI engine may be the RTMDet model described in Lyu, et al., “Rtmdet: An empirical study of designing real-time object detectors”, arXiv preprint arXiv:2212.07784 (2022). This model has a backbone-neck-heads structure, where the backbone and neck have harmonious capacities and both have large-kernel depth-wise convolutions as their basic building blocks. A multiscale feature pyramid extracted from the backbone is input into the neck, and the pyramid feature map is improved through bottom-up and top-down feature propagation. The head predicts object bounding boxes and predicts their categories based on the feature maps of each scale. This model can extract fashion product categories at high speed.

[0032] In another embodiment, the classification AI engine may be a Vision Transformer-based AI model. A Vision Transformer is a deep learning model that performs well in image recognition and processing tasks. By using a Transformer structure instead of existing Convolutional Neural Networks (CNNs), it can appropriately capture and process global image information. The baseline model of the SOTA model adopted by Imagenet, a benchmark in the field of image classification, uses this Vision Transformer as its backbone.

[0034] In the embodiment illustrated in FIG. 2, the classification-based fashion product attribute extraction unit (110) may include a classification-based category information extraction unit (111) and a classification-based style information extraction unit (113). The classification-based category information extraction unit (111) receives a product image and extracts category information of the fashion product based on a classification-type artificial intelligence engine. The classification-based style information extraction unit (113) receives a product image and extracts style information of the fashion product based on a classification-type artificial intelligence engine. Depending on the characteristic aspects, the classification-based category information extraction unit (111) and the classification-based style information extraction unit (113) may be models trained by the aforementioned RTMDet model with category-labeled training data and style-labeled training data, respectively.

[0035] <Description of Claim 4>

[0036] In the proposed invention, the categories of fashion products output by the classification-based category information extraction unit (111) may have a hierarchical structure of major categories, medium categories, and sub-categories. The major categories may be divided, for example, into clothing, shoes, bags, accessories, hats, and others.

[0037] For example, the medium and detailed classifications for clothing can have the following system.

[0038] Outers

[0039] Item 2 Item 3 Item 4 Coats / Jackets Basic Tailored Trench Baseball Trucker Poncho Trench coat, duffle coat, cape, short coat, trench coat, winter coat, wool coat, cocoon coat, double-breasted coat, half coat, leather coat, long coat, mouton coat, rain coat, wrap coat, fur coat Blazer jacket, Leather jackets, Mustang jacket, Denim jacket, Lightweight jacket, Athletic jacket, Fleece jacket, Outdoor jacket, Stadium jacket, Safari jacket, Coach jacket, Winter jacket, Vests, Blouson, Bike jacket, Single-breasted jacket, Double-breasted jacket, Military jacket, Tailored jacket, Windbreaker, Poncho, Fur jacket jumpers Padded anorak zip-up Zip-up hoodies, down jumpers, windbreakers, fleece jumpers, parkas, shawls, bombers, truckers cardigans Cardigan bolero Cardigans, Boleros Best (vests) Best Gilets & Waistcoats, Padded Vest, Fur Vest, Bustier

[0041] Tops

[0042] Item 2 Item 3 Item 4 Woven tops 　 Silk Tops, Tunic Tops, Blouses, Tees, Shirts, Business Shirts, Tanks knit tops 　 Compression Tops, Sweatshirts, Heavy Knit, Medium Knit, Fine Knit, Polo Shirts, T-Shirts, Print T-Shirts, Tilden Vest styles of tops 　 Vintage Tops, Crop Tops Tower structures (patterns of tops) 　 Camisole Tops, Kimono Tops, Hoodies, Off-shoulder Tops, Long-sleeve Tops, Sleeveless, Short-sleeve Tops, Tanks, Vest Tops

[0043] Pants

[0044] Item 2 Item 3 Item 4 Woven pants 　 Skinny Fit pants, Denim Shorts, Chinos, Ski Bottoms, Cargo Trousers, Hotpants knit pants 　 Joggers & Sweats, Leggings, Sweatpants, Relaxed Jeans Structure of pants (patterns of pants) 　 Flare pants, Bootcut pants, Loose Fit pants, Slim Fit pants, Straight Leg pants, Shorts, Tapered Jeans, High-Waisted Jeans, Wide Leg Jeans, Cropped Jeans, Boyfriend Jeans, Ankle Jeans, Baggy

[0045] skirts

[0046] Item 2 Item 3 Item 4 Styles of skirt Basic pleated flare Leather skirts, denim skirts, business skirts, casual skirts, kilt skirts, wrap skirts, tulle skirts length of skirt 　 Mini Skirts, Maxi Skirts skirt structure (patterns of skirt) 　 Pleated Skirts, A-Line Skirts, Pencil Skirts, Bubble Hem Skirts, Culottes, Gored Skirt, Peplum Skirt, Tiered Skirt, Trumpet Skirt

[0047] Dress

[0048] Item 2 Item 3 Item 4 dress Dress jumpsuit Off-Duty Dresses, Co-ordinates, Gowns, Wrap Dresses, Loungewear Dresses, Slip Dresses, Shift Dresses, Night Out, Vintage Dresses, Day to Night, Bodycon, Tie-Dye Dresses, Formal Dresses, Casual Dresses, Bra Bridal Dresses, Cocktail Dresses, Black Dresses, Evening Dresses, Daytime Dresses, Denim Dresses, Knitted Dresses, Jersey Dresses, Shirt Dresses, Work Dresses, Sweater Dresses, Halter Neck Dresses, Wedding Dresses length of dress 　 Midi dresses, Maxi dresses, Knee-length dresses, Mini dresses

[0049] Others

[0050] Item 2 Item 3 Item 4 nightwear 　 Sleepwear, Dressing Gowns, Nightwear, Bathrobes, Pajama Tops underwear 　 Sports Bras, Base Layers, Underwear, Undershirts, Shapewear, Bustier, Camisole sportswear 　 Activewear Outerwear, Activewear Jackets, Tracksuits 　 Activewear Tops, Sports Bras, Football Shirts, Runner Shorts 　 Activewear Bottoms, Tracksuits

[0051] In addition to clothing, the footwear subcategory among the major categories includes boots, heels, flats, sandals, mules, slippers, sneakers, and sportswear; the bag subcategory includes backpacks, shoulder bags, waist bags, handbags, and carriers; the accessory subcategory includes scarves, gloves, belts, eyewear, socks, stockings, jewelry, and watches; and the hat subcategory may include beanies, bucket hats, berets, hinting caps, caps, and fedoras.

[0052] <Description of Claim 5>

[0053] In the proposed invention, the style information of a fashion product output by the classification-based style information extraction unit (113) may have the following system for, for example, clothing.

[0054] Traditional Mannish Feminine ethnic Contemporary Natural Gender fluid sporty subculture Classic, Preppy Mannish, Tomboy Feminine, romantic, sexy Hippie, Western, Oriental Modern, Sophisticated, Avant-garde Country, Resort Genderless sporty Retro, Kitsch, Kidult, Hip Hop, Punk

[0055] Style information can have at least one of these 22 style values.

[0056] <Description of Claim 6>

[0057] Depending on additional aspects, characters found in product images may be extracted and reflected in product recommendation content. In the embodiment illustrated in FIG. 2, the product information extraction unit (100) may further include a character recognition information extraction unit (130). The character recognition information extraction unit (130) receives a product image and extracts text contained within the image. The character recognition information extraction unit (130) first detects areas of characters in the input product image and applies character recognition technology, commonly known as OCR technology, to the detected character areas to extract characters.

[0058] <Description of Claim 7>

[0059] Depending on additional aspects, product attribute information can be extracted from various multimedia-based documents regarding fashion products and reflected in product recommendation content. In the embodiment illustrated in FIG. 2, the product information extraction unit (100) may further include an entity detection-based fashion product attribute extraction unit (150). The entity detection-based fashion product attribute extraction unit (150) can receive multimedia-based documents as input and extract attribute information of fashion products based on an entity detection type artificial intelligence engine. Typical examples of multimedia-based documents include web documents containing images, videos, graphics, and text.

[0060] <Description of Claim 8>

[0061] In one embodiment, the attributes of a fashion product may follow a standard classification system of the fashion industry. According to this classification system, fashion attributes can be broadly divided into customer-perspective attributes and product-perspective attributes. Furthermore, customer-perspective attributes can be divided into functional, expressive value, and aesthetic attributes, while product-perspective attributes include design and brand attributes. Customer-perspective attributes include the following:

[0062] division Attribute name Value example Functional protection Durability, Windproof, Waterproof, Antibacterial, Water Repellent, Thermal Insulation, Heat Generation, Pest Control, Disinfection, Functional donning Button, buttonless, zipper, band, elastic banding, Velcro, hook-and-loop fastener, snap button, string, snap, hook Functional care Washable, water wash, no ironing, wrinkle-free, low wrinkle, wrinkle-resistant, Functional comfort Soft, warm, cool, cozy, chewy, crisp, cooling Functional mobility Stretch, elasticity, lightweight, light, seamless, stitchless Functional bodyfit 163cm, 45kg ExpressiveValue tpo Official, Daily, ExpressiveValue eco Eco-friendly, reuse, recycling, recycle, organic, ethical, sustainable, environmental certification, eco, ExpressiveValue brand Nike, Adidas, ExpressiveValue price 59,900 won Aesthetic style Basic, Casual, Aesthetic personal color Springwarmbright,Springwarmlight,Summercoollight,Summercoolmute,Autumnwarmmute,Autumnwarmdeep,Wintercooldeep,Wintercoolbright Aesthetic bodytype Short and skinny build, short and chubby build...

[0063] The attributes of the customer perspective include the following:

[0064] Design color Black, White, Design neckline Round, V, 쪋 Design collar China, Bow, 쪋 Design detail A-line, Mini / Mepsi etc. Design fit Tight, Normal, 쪋 Design category T-shirt, top, Design size_detail Total length: 92cm, Waist circumference: 34cm, .. Design length Crop, mini, 쪋. Design size XXS, XS, 쪋 Design print Check, zebra, 쪋 Design fabric Denim, Polyester. brand Nike, Adidas, and related brands

[0065] <Description of Claim 9>

[0066] Multimedia-based documents input to the entity detection-based fashion product attribute extraction unit (150) are typically web documents containing images, text, and graphics. These web documents may be posted on a shopping mall for fashion products, or they may be review articles or news articles about the fashion products. This information can be collected by a web crawler. The entity detection-based fashion product attribute extraction unit (150) extracts text portions from the input web documents and annotates them to extract fashion attributes. In the illustrated embodiment, text extracted from a product image (11) may be added as input to the entity detection-based fashion product attribute extraction unit (150).

[0067] In one embodiment, the entity detection type artificial intelligence engine may apply the BERT-CRF architecture model, which is a named entity recognition model utilizing the OpenTag model concept. This model is a deep learning-based Named Entity Recognition (NER) model and is used as a baseline model in various product attribute extraction studies. The Viterbi Algorithm may be applied for decoding. This technology is described in detail, for example, in Xu, et al., “Scaling up open tagging from tens to thousands: Comprehension empowered attribute value extraction from product title”, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019.

[0068] However, the proposed invention is not limited to this, and other AI-based annotation tools or ultra-large AI-based fashion attribute extraction models may also be applied.

[0070] <Description of Claim 10 Invention>

[0071] In the illustrated embodiment, the product information extraction unit (100) may further include a rule-based fashion product category information extraction unit (170). The rule-based fashion product category information extraction unit (170) receives product information in text form and can extract classification information of fashion products based on a heuristic algorithm. Here, the category information may be the same information as described in the classification-based category information extraction unit (111). This method is a traditionally widely used method for extracting fashion attributes of products, and it utilizes domain-specific knowledge to design and apply regular expressions. This rule-based information extraction is easy to apply to web documents that are standardized into a specific format. Since attributes exist by region, attributes can be extracted by separating the regions after parsing. In the illustrated embodiment, text extracted from a product image (11) may be added as input to the rule-based fashion product category information extraction unit (170).

[0073] The present invention is implemented through a plurality of artificial intelligence engines controlled by a controller, some of which may be implemented as program instructions, and others as dedicated artificial intelligence processors or dedicated artificial intelligence information processing devices. Each task may be executed sequentially or independently with concurrency. The claims relating to methods in the appended claims are methods executed in these information processing devices, the substance of which is sufficiently understood by those skilled in the art from the aforementioned invention related to the device, so a detailed description is omitted.

[0075] Although the present invention has been described above with reference to embodiments with reference to the accompanying drawings, it is not limited thereto and should be interpreted to encompass various variations that can be obviously derived from them by those skilled in the art. The claims are intended to encompass such variations. Explanation of the symbols

[0076] 11 : Product Image 13 : Product Information 31 : Content Directive Template DB 33 : Style Directive Template DB 100 : Product Information Extraction Unit 110 : Classification-based fashion product attribute extraction unit 111 : Classification-based category information extraction unit 113 : Classification-based style information extraction unit 130 : Character recognition information extraction unit 150 : Entity Detection-Based Fashion Product Attribute Extraction Unit 170 : Rule-based fashion product category information extraction unit 300 : Seller Interface 510 : Content Prompt Generation Section 530 : Style prompt generation section 700 : Large Language Model interface

Claims

Claim 1 A fashion product recommendation content generation system comprising: a seller interface that receives information regarding a fashion product to be recommended content from a seller and outputs the generated content; a product information extraction unit that receives product detailed information in the form of product images and text, processes it using artificial intelligence to extract product attribute information including the category and style of the fashion product; a content prompt generation unit that generates a prompt requesting a Large Language Model (LLM) to create fashion product recommendation content from product information entered through the seller interface and product attribute information extracted by the product information extraction unit, and generates a content prompt that includes information regarding the content of the fashion product recommendation content to be created; a style prompt generation unit that generates a prompt requesting a Large Language Model to create fashion product recommendation content from product information entered through the seller interface and product attribute information extracted by the product information extraction unit, and generates a style prompt that includes information regarding the style of the fashion product recommendation content to be created; and a Large Language Model interface that provides the generated content prompt and style prompt to the Large Language Model, receives the response, and outputs the fashion product recommendation content to the seller interface. Claim 2 A fashion product recommendation content generation system according to claim 1, wherein the product information extraction unit comprises: a classification-based fashion product attribute extraction unit that receives a product image as input and extracts category information and style information of a fashion product based on a classification-type artificial intelligence engine. Claim 3 A fashion product recommendation content generation system according to claim 2, wherein the product information extraction unit comprises: a classification-based category information extraction unit that receives a product image and extracts category information of a fashion product based on a classification-type artificial intelligence engine; and a classification-based style information extraction unit that receives a product image and extracts style information of a fashion product based on a classification-type artificial intelligence engine. Claim 4 A fashion product recommendation content generation system according to claim 3, wherein category information comprises at least one value of a hierarchical structure including: major categories of clothing, shoes, bags, accessories, and hats; medium categories including outerwear, tops, pants, skirts, jumpsuits, and dresses in clothing; boots, heels, flats, sandals, mules, slippers, sneakers, and sports shoes in shoes; backpacks, shoulder bags, waist bags, handbags, and carriers in bags; scarves, gloves, belts, eyewear, socks, stockings, jewelry, and watches in accessories; and medium categories including beanies, bucket hats, berets, hinting caps, caps, and fedoras in hats; and subcategories for each medium category. Claim 5 A fashion product recommendation content generation system according to claim 3, wherein style information includes at least one value among: Traditional: (Classic, Preppy), Mannish: (Mannish, Tomboy), Feminine: (Feminine, Romantic, Sexy), Ethnic: (Hippie, Western, Oriental), Contemporary: (Modern, Sophisticated, Avant-garde), Natural: (Country, Resort), Genderfluid: (Genderless), Sporty: (Sporty), Subculture: (Retro, Kitsch, Kidult, Hip Hop, Punk). Claim 6 The fashion product recommendation content generation system according to claim 1, wherein the product information extraction unit further comprises: a character recognition information extraction unit that receives a product image as input and extracts texts included in the image. Claim 7 A fashion product recommendation content generation system according to claim 1, wherein the product information extraction unit comprises: a classification-based fashion product attribute extraction unit that receives product information in text form and extracts attribute information of a fashion product based on an entity detection type artificial intelligence engine. Claim 8 Claim 7, a fashion product recommendation content generation system comprising fashion product attributes including customer perspective attributes including functional, expressive value, and aesthetic attributes, and product perspective attribute values ​​including design and brand. Claim 9 Claim 7, wherein the entity detection-based fashion product attribute extraction unit comprises a BERT-CRF architecture model, a fashion product recommendation content generation system. Claim 10 Claim 1, wherein the product information extraction unit comprises: a rule-based fashion product category information extraction unit that receives product information in text form and extracts classification information of fashion products based on a heuristic algorithm; a fashion product recommendation content generation system. Claim 11 A method for generating fashion product recommendation content, comprising: a seller instruction input step for receiving information regarding a fashion product to be recommended content from a seller; a product information extraction step for receiving product detailed information in the form of product images and text, processing it with artificial intelligence to extract product attribute information including the category and style of the fashion product; a content prompt generation step for generating a prompt requesting a Large Language Model (LLM) to create fashion product recommendation content from the product information entered in the seller instruction input step and the product attribute information extracted in the product information extraction unit, wherein the content prompt includes information regarding the content of the fashion product recommendation content to be created; a style prompt generation unit for generating a prompt requesting a Large Language Model to create fashion product recommendation content from the product information entered in the seller instruction input step and the product attribute information extracted in the product information extraction step, wherein the style prompt includes information regarding the style of the fashion product recommendation content to be created; and a content generation step for providing the generated content prompt and style prompt to the Large Language Model, receiving the response, and outputting the fashion product recommendation content to the seller. Claim 12 A method for generating fashion product recommendation content according to claim 11, wherein the product information extraction step comprises: a classification-based fashion product attribute extraction step that receives a product image as input and extracts category information and style information of a fashion product based on a classification-type artificial intelligence engine. Claim 13 A method for generating fashion product recommendation content according to claim 12, wherein the product information extraction step comprises: a classification-based category information extraction step that receives a product image and extracts category information of a fashion product based on a classification-type artificial intelligence engine; and a classification-based style information extraction step that receives a product image and extracts style information of a fashion product based on a classification-type artificial intelligence engine. Claim 14 A method for generating fashion product recommendation content according to claim 11, wherein the product information extraction step comprises: a classification-based fashion product attribute extraction step that receives product information in text form and extracts attribute information of fashion products based on an entity detection type artificial intelligence engine. Claim 15 A method for generating fashion product recommendation content according to claim 11, wherein the product information extraction step comprises: a rule-based fashion product category information extraction step that receives product information in text form and extracts classification information of fashion products based on a heuristic algorithm.

Images