Product information generation system, server and program

Abstract

Achieve both quality and work efficiency when creating product descriptions and attribute information using product images and supplementary information. [Solution] Product images and supplemental information sent from user devices are received by a server, and then integrated and input into a generation AI using the same inference process to generate product information including product descriptions. An external product database is searched based on the generated product information, and category information is identified for the top candidates using statistical criteria consisting of at least one of majority voting, weighted voting, or rank integration (reliability values ​​are integrated as necessary), and attribute items and attribute information associated with the category are generated. The user interface detects inconsistencies and highlights or guides the user to make corrections depending on their importance (based on their essentiality, sales impact, and frequency of occurrence). Furthermore, a configuration is adopted in which description generation and attribute generation are performed independently of each other, aiming to localize errors and reduce re-execution time.

Description

[Technical Field]

[0001] The present invention relates to a technology for automating product registration operations, and in particular to a product information generation system that automatically generates product descriptions and product attribute information using generation AI, and related technology. [Background technology]

[0002] Previously, when registering products on e-commerce sites or online shops, staff had to manually enter information about each product, such as the product name, description, materials, brand, use, and country of manufacture.

[0003] To improve the efficiency of such registration work, automatic product information generation systems that apply image recognition technology and natural language generation technology have been proposed.

[0004] For example, Patent Document 1 discloses a technology in which a product image is input and attribute information of the product is generated based on features extracted from the image.

[0005] Furthermore, Patent Document 2 discloses a system that extracts text, logos, and the like from product packaging images and registers them in a product database.

[0006] Furthermore, Patent Document 3 discloses a technology for analyzing product images using a convolutional neural network (CNN) and classifying them into categories. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] US Patent Application Publication No. 2019 / 0311210 [Patent Document 2] U.S. Patent No. 9,262,686 [Patent Document 3] U.S. Patent No. 10,817,749 Summary of the Invention [Problem to be solved by the invention]

[0008] However, because these conventional technologies all rely on the analysis of product images, the content of the product information generated is limited, and there are cases where the product information intended by e-commerce site owners (such as the granularity and tone of expression, attributes that should be emphasized for sales purposes) cannot be fully obtained.

[0009] Such insufficient generated results make it necessary to make individual corrections or additional inputs when registering products, which increases the number of input steps and hinders efficient product information input.

[0010] Therefore, the present invention aims to provide a product information generation system and related technologies that improve the suitability of the content and expression of product information used for product registration, while reducing the burden required for inputting such information, and enable e-commerce site merchants to efficiently register products. [Means for solving the problem]

[0011] The present invention relates to a product information generation system, server, and program that streamlines product registration work by inputting product images and supplemental information about the product into a generation AI and automatically generating product descriptions and attribute information in natural language.

[0012] According to one aspect of the present invention, there is provided a product information generation system including a user terminal and a server capable of communicating with the user terminal via a network.

[0013] The user terminal includes a transmitting unit that receives product images and supplemental information about the products and transmits them to a server.

[0014] The server includes a receiving unit that receives the product image and the supplemental information, and a product information generating unit that inputs the product image and the supplemental information into a generation AI and integrates and processes them in the same inference process to generate product information including at least a product description in natural language. Here, the integration and processing method may be at least one of a method of listing image tokens and text tokens side by side in the same prompt, or a method of combining image embedding and text embedding.

[0015] The server includes an identification unit that searches an external product database based on the product information generated by the product information generation unit, identifies category information, and identifies attribute items associated with the category information.

[0016] The identification unit may extract key words included in the product information, input them to an external product database search API, extract a plurality of category candidates based on the search results, and perform analysis based on at least one statistical criterion of majority voting, weighted voting, or rank integration to identify category information. The configuration may also be such that category information is identified by majority voting based on the appearance frequency of the top N category candidates (N is a positive integer) in the search results.

[0017] The identification unit may also identify category information based on an integrated score calculated by integrating the appearance frequency, score, and output reliability of the generation AI for multiple category candidates obtained from search results of an external product database. This allows for improved classification accuracy by combining the output reliability of the generation AI with external statistical data.

[0018] The server may further include an attribute information generation unit that uses the generation AI to generate attribute information consisting of text to be input into the attribute items based on the category information and attribute items identified by the identification unit. This generation is preferably performed in a two-stage configuration in which product information is generated in a first cycle and attribute information is generated in a second cycle. Furthermore, the output of the first cycle may not be included in the input of the second cycle, and the two cycles may be performed independently (non-dependently) from each other.

[0019] The product information generation unit and the identification unit may execute processing by calling an external AI model or an external API, thereby flexibly linking with a general-purpose generation service or an external product database and easily updating the model or data.

[0020] The server transmits the generated product information and attribute information to the user terminal, and the user terminal has a correction input unit that displays the information on a display unit and accepts confirmation and correction input from the user.

[0021] The correction input unit may display a reliability value indicating the likelihood of the estimation result by the generation AI, and may present a message prompting correction if the reliability is below a predetermined threshold. Furthermore, if an inconsistency is detected between the estimation content based on the product image and the supplemental information, the correction input unit may present the possibility and a correction policy depending on the type and severity of the inconsistency.

[0022] Furthermore, the correction input unit may present multiple category candidates obtained by the identification unit and determine the category information based on a user's selection. When an inconsistency is detected, the importance may be calculated based on the degree of essentiality, the degree of impact on sales, and the frequency of occurrence, and the inconsistency may be presented in the form of a highlight or a message according to the importance. This automatically organizes the priority of review work and enables efficient correction.

[0023] The product information generation unit may be configured to input keywords or prompts that control the format and content of the output when inputting the information to the generation AI. Also, the product information generation unit may receive product information from the generation AI according to a predetermined data structure (structured format including JSON) specified by the prompt, and store the content for each item of the data structure.

[0024] Furthermore, the server may record the operation results of the correction input unit as an audit log, and statistically compile the log to use for relearning the reliability threshold or the output parameters of the generation AI, thereby enabling the system to improve its performance in a self-improving manner based on data from actual operation.

[0025] The present invention may also be realized as the above-mentioned server itself, i.e., a product information generation server comprising a receiving unit, product information generating unit, and specifying unit (and an attribute information generating unit, if necessary), or as a program that causes the server to execute (a) a receiving process, (b) a process for generating product information by integrating and processing both in the same inference process, and (c) a process for searching an external product database and specifying category information and attribute items.Furthermore, the present invention may be realized as a product information generation system comprising: an image input means for inputting product images; a supplemental information input means for inputting supplemental information; a generation means for inputting the product images and supplemental information to a generation AI and integrating and processing both in the same inference process to generate product descriptions; and a specification means for searching an external product database based on the product information generated by the generation means to specify category information and specify attribute items associated with the category information.

[0026] According to each aspect of the present invention, the accuracy of generation is improved by integrating and inputting product images and supplementary information into the generation AI in the same inference process, and by combining operations using statistical criteria such as majority voting, weighted voting, or rank integration, and thresholds based on reliability values, with category identification, attribute generation, inconsistency detection, and log relearning, it is possible to provide a product registration support technology that combines efficiency, reliability, and self-improvement in actual operations. [Effects of the Invention]

[0027] According to this invention, by inputting supplemental information along with product images into the generation AI, the content and expression of the generated product information is closer to the intentions of the e-commerce site owner and is less limited than when it is based solely on images. As a result, the amount of additions and corrections required during the confirmation stage is reduced, allowing for more efficient product information entry during product registration. Furthermore, because category information and corresponding attribute items can be identified based on publicly available information from an external product database, it becomes easier to present and select items to be entered, improving the overall throughput of the registration process.

[0028] It should be noted that the above-mentioned Patent Documents 1 to 3 do not describe specific configurations for improving the compatibility of the content and expression of product information used in product registration while reducing the burden of inputting the information, and enabling e-commerce site merchants to efficiently register products, as in the product information generation system, server, and program of the present invention. In particular, while these conventional technologies generate information by relying solely on the analysis of product images, the present invention differs in that it employs a configuration in which product images and supplemental information are input into a generation AI in an integrated manner, thereby relaxing the limitations on the content and expression of the generated product information and realizing more efficient input work. [Brief explanation of the drawings]

[0029] [Figure 1] 1 is a block diagram showing the overall configuration of a product information generation system according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram showing a sequence of a product information generation process according to an embodiment of the present invention. [Figure 3] FIG. 10 is a diagram showing an example of a product information input screen according to one embodiment of the present invention. [Figure 4] FIG. 10 is a diagram showing an example of a product information confirmation / editing screen (excerpt of product description section) according to one embodiment of the present invention. [Figure 5] FIG. 1 is a diagram showing an example of an input configuration (first cycle (A) and second cycle (B)) to a generation AI according to one embodiment of the present invention. [Figure 6]FIG. 2 is a diagram illustrating an overview of communication with an external API according to an embodiment of the present invention. [Figure 7] FIG. 10 is a diagram showing the flow of a category determination process according to an embodiment of the present invention. [Figure 8] FIG. 10 illustrates an example of an inconsistency detection and correction prompting process according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] <1. Embodiment> The product information generation system according to the present invention operates as a product registration support function in an e-commerce site management system. Target e-commerce sites include, but are not limited to, Rakuten Ichiba, Yahoo! Shopping, and Amazon. This function can be incorporated into the existing product master registration flow provided in the e-commerce site management system to simplify the input work during registration.

[0031] With the introduction of this function, product registration staff can input only product images and supplementary information, instead of manually entering detailed descriptions and attribute information. The server integrates and processes this input in the generation AI using the same inference process to automatically generate product information, and displays the results on the registration screen of the e-commerce site's management system. The staff member can check the content on the screen, make any necessary corrections, and then update the product master using the normal registration procedure.

[0032] This function utilizes the existing implementation of each mall integration (API, CSV, etc.) that the e-commerce site management system has. In other words, this function is responsible for generating and finalizing product information, and the final registration and synchronization with the mall is carried out by the standard functions of the e-commerce site management system. This makes it possible to streamline product information input without changing existing operations or integration specifications.

[0033] Additionally, category identification and attribute item determination are performed by referencing publicly available information from an external product database. There are no restrictions on the specific vendors or algorithms of the generating AI or external database, and these can be selected according to the operational environment of the e-commerce site management system.

[0034] Here, as an example of a product information generation system according to the present invention, a product information generation system 10 shown in Fig. 1 is illustrated. Also, as an example of a server according to the present invention, a server 12 shown in Fig. 1 is illustrated. Furthermore, as an example of a program according to the present invention, a program executed by the server 12 shown in Fig. 1 is illustrated.

[0035] As shown in Fig. 1, the product information generation system 10 comprises a user terminal 11 and a server 12. The two can communicate with each other via a network N, which is connected to the Internet including public lines.

[0036] The user terminal 11 is an information processing device operated by an EC site merchant, and is exemplified by a laptop (11A, 11B). The terminal 11 is equipped with a processor (CPU) and memory (RAM, ROM, etc.), and the processor executes programs stored in the memory to realize various functions. The memory is a computer-readable recording medium that stores the programs and processing data. If necessary, the terminal 11 may also be equipped with non-volatile storage (SSD, HDD, etc.).

[0037] A display device corresponding to the display unit 111 is connected to the user terminal 11, and input devices such as a keyboard, mouse, and touch panel are provided corresponding to the correction input unit 112. Furthermore, the user terminal 11 is provided with a network interface and performs two-way communication with the server 12 via the Internet.

[0038] The user terminal 11 is provided with a display unit 111, a correction input unit 112, and a transmission unit 113. The display unit 111 is a processing unit that displays the input screen of FIG. 3 and the confirmation / correction screen of FIG. 4. The correction input unit 112 is a processing unit that accepts instructions such as correction, selection, and registration confirmation on the displayed screen. The transmission unit 113 is a communication processing unit that transmits and receives data to and from the server 12.

[0039] The server 12 is provided with a receiving unit 121, a product information generating unit 122, an identifying unit 123, and an attribute information generating unit 124. The server 12 is equipped with a processor (CPU) and memory (RAM, ROM, etc.), and the processor executes a program stored in the memory to realize the functions of the receiving unit 121, product information generating unit 122, identifying unit 123, and attribute information generating unit 124. The memory is a computer-readable recording medium that stores the program and processing data such as product information, category information, attribute items, etc. If necessary, the server 12 may be equipped with non-volatile storage for holding logs and caches.

[0040] The receiving unit 121 is a communication processing unit that receives product images and supplemental information from the user terminal 11. The product information generating unit 122 inputs the images and supplemental information into the generation AI 30 and generates product information including a product description in natural language. The identifying unit 123 analyzes the search results of the external product database 20 and identifies category information and corresponding attribute items. The attribute information generating unit 124 uses the generation AI 30 to generate text to be entered into each attribute item.

[0041] In this specification, each term is defined as follows (1) to (4).

[0042] (1) Integrated processing: This refers to the input format in which features based on product images (image tokens or image embeddings) and text based on supplemental information are combined and provided to the generative AI in the same inference process. Specific examples include (i) the same prompt containing image tokens and text tokens, and (ii) the combination of image embeddings and text embeddings.

[0043] (2) Statistical criteria: Criteria calculated by at least one of majority voting, weighted voting, or rank integration based on the frequency of appearance or score of category candidates included in candidates in an external product database. If necessary, the reliability of the generated AI output may also be integrated.

[0044] (3) Reliability value: An evaluation value between 0 and 1 based on the output of the generation AI, etc., and controls regeneration, warning display, or review priority based on a predetermined threshold value.

[0045] (4) Importance: An evaluation value calculated based on the necessity of the inconsistency, the impact on sales, and the frequency of occurrence, and is used to emphasize the display and control the review order.

[0046] In this specification, "category information" refers to the classification categories of products in an external product database, and refers to information that classifies products hierarchically or by a code system. For example, it may include hierarchical names such as "Fashion > Men's > T-shirts and cut-and-sew," "Home appliances > Audio > Earphones and headphones," or "Food > Beverages > Coffee," as well as the category codes associated with these. The category naming rules and code system may conform to each mall's public API (e.g., genre search API, product API, etc.), but the present invention is not limited to these. In this specification, "category information" and "attribute items" are used as conceptual components that are independent of specific database structures or mall specifications.

[0047] In this specification, "attribute item" refers to the name of an item that is required to be entered in a product master or mall linkage in a specific category, and the text value actually entered in that item is called "attribute information." Examples of attribute items are as follows (all are illustrative and not limiting):

[0048] (1) Apparel (T-shirts): Size, color, material, pattern, sleeve length, neck shape, brand, country of origin, season, model number, JAN / UPC, and washing instructions.

[0049] (2) Home appliances (wireless earphones): Connection method (Bluetooth / wired), supported codec, presence or absence of noise cancellation, waterproof rating, driver diameter, continuous playback time, weight, charging terminal type, accessories, model number, presence or absence of technical conformity mark.

[0050] (3) Food (coffee beans): Contents, bean shape (beans / ground), roasting level, country of origin / producing area, expiration date, storage method, allergen information, whether or not the product is organically certified, and JAN.

[0051] (4) Household goods (frying pans): Nominal diameter, material, coating type, compatible heat source (IH / gas), dishwasher safe, heat resistance temperature, whether or not a lid is provided, weight, and manufacturer's warranty.

[0052] (5) Cosmetics (lotion): expected skin type, main ingredients, whether it is alcohol-free or not, whether it is paraben-free or not, content amount, how to use, and quasi-drug classification.

[0053] Upon identifying category information, the identification unit 123 determines a set of attribute items associated with the category. The attribute information generation unit 124 generates a text value for each attribute item. Normalization and validation may be performed, as necessary, by referencing a controlled vocabulary of terms (e.g., standardization of color names and size symbols) and input constraints (e.g., numerical ranges and options). This enables mutual consistency between product information and attribute information and stable mapping to mall-linked items, and also ensures compatibility when directly mapping the results automatically generated by the generation AI to input items in external systems.

[0054] The server 12 is equipped with a network interface and is connected to the Internet. This allows communication with the external product database 20 and the generation AI 30 (S4 to S8 in FIG. 2, FIG. 6). This communication supports either API calls or CSV linkage and is not dependent on a specific vendor. The generation AI 30 is a generic term for an external generation AI model, and is not limited to a specific vendor or method. The external product database 20 is also generic, and provides a search API and an attribute item acquisition API.

[0055] The functions of the present invention are realized as a program executed on the server 12. The program may be supplied to the server 12 via a network, or may be provided by being recorded on a non-transitory computer-readable recording medium (e.g., semiconductor memory, optical disk, magnetic disk). The processor reads the program from the memory and executes it to realize the functions of the receiving unit 121, product information generating unit 122, identifying unit 123, and attribute information generating unit 124. The program executed on the user terminal 11 also realizes the functions of the display unit 111, correction input unit 112, and transmission unit 113, thereby embodying each of the means recited in the means claims.

[0056] Next, the processing sequence in the product information generation system will be described with reference to the sequence diagram of FIG.

[0057] In step S1, the sending unit 113 sends the product image and supplemental information (material, brand, use, country of manufacture, etc.) specified by the user to the server 12. In step S2, the product information generation unit 122 inputs the received image and supplemental information into the generation AI 30 and requests the generation of product information. In step S3, the generation AI 30 replies with product information consisting of product descriptions, etc., which is acquired by the product information generation unit 122.

[0058] In step S4, the identification unit 123 sends the main words and phrases extracted from the product information in step S3 as search parameters to the search API of the external product database 20. In step S5, the external product database 20 returns the top N (e.g., 30) search results, which are received by the identification unit 123. In step S6, the identification unit 123 identifies category information using statistical criteria such as majority voting based on appearance frequency, and identifies attribute items associated with the category by referring to the cache and internal tables in the external product database 20 or the server 12. Note that extraction of main words and phrases can be achieved by known methods such as morphological analysis, N-grams, or statistical weighting (TF-IDF, etc.).

[0059] In step S7, the attribute information generation unit 124 requests the generation AI 30 to generate attribute information using the category information and attribute items from step S6 as input conditions. In step S8, the generation AI 30 returns text for each attribute item, which is acquired by the attribute information generation unit 124. The server 12 integrates the product information and attribute information, and in step S9, transmits the integration result to the user terminal 11.

[0060] In step S10, the display unit 111 displays the integrated result as a confirmation screen, and the user confirms and corrects it using the correction input unit 112. Here, the correction input unit 112 calculates or acquires a reliability value for the generated result, displays the reliability value on the display unit 111, and executes processing to display a message prompting correction if the reliability is below a predetermined threshold. The reliability value is calculated based on the score returned by the generation AI 30 or a statistical indicator (such as similarity or match rate) in the server 12. The predetermined threshold is determined by empirical setting, optimization based on verification data, or an operation policy, and can be set, for example, in the range of 0.7 to 0.9 (an example obtained by experimental verification based on trained evaluation data).

[0061] Furthermore, when an inconsistency is detected between the estimated content based on the product image and the supplemental information, the correction input unit 112 causes the display unit 111 to display the possibility of the inconsistency and a correction policy. The detection of the inconsistency can be implemented by a known method such as comparing item values, determining keyword consistency, or detecting missing information. Furthermore, multiple category candidates may be presented, and the category may be determined based on the user's selection.

[0062] As mentioned above, by integrating and processing product images and supplementary information in the same inference process, product descriptions that simultaneously reflect image findings and promotional intent can be generated reliably, and because they are output in a specified data structure (e.g., JSON), the number of times proofreading and regeneration is reduced, improving work efficiency.

[0063] If the user makes any corrections, the corrected content and a registration instruction are sent in step S11A, and if no corrections are required, only the registration instruction is sent in step S11B. Finally, in step S12, the server 12 executes the registration process to the product master of the EC site management system.

[0064] Next, examples of each screen and the input / output configuration will be described with reference to Figures 3 to 8. On the input screen of Figure 3, the display unit 111 displays an image input area and a supplemental information (e.g., keyword) input field. When the user inputs and transmits the required information, the transmission unit 113 starts S1.

[0065] The display unit 111 may accept input by drag and drop or file selection into the image input area. The correction input unit 112 may generate a thumbnail immediately after input and display it on the display unit 111, and may accept operations to delete, replace, or rearrange multiple images. The transmission unit 113 transmits the image data to the server 12 only when the user explicitly performs a transmission operation.

[0066] The image acceptance specifications may be limited to those that satisfy a predetermined file format (e.g., JPEG, PNG) and upper limit size. If the conditions are not satisfied, the correction input unit 112 displays an error message on the display unit 111 and prompts the user to reselect. If multiple images are input, the correction input unit 112 may allow the user to select a "main image," and the transmission unit 113 may transmit the designation information together.

[0067] The input field for supplemental information may be configured to be a free-form keyword field, but may also be configured to allow for the addition of optional items such as brand, material, model number, etc. The correction input unit 112 may perform a simple validity check of the input value (detection of missing required items, upper limit of character count, allowable character types, etc.) and display an inline warning on the display unit 111.

[0068] During a transmission operation, the transmission unit 113 displays the progress status (progress bar, transmitting indicator), and if an exception such as a communication interruption occurs, the display unit 111 may present an option for automatic retransmission or retry. After transmission is complete, the screen may be configured to wait until the processing result of the next step is received in conjunction with S1 in Fig. 2, and transition to the confirmation and correction screen in Fig. 4 upon reception.

[0069] From the viewpoint of privacy and prevention of erroneous transmission, the image may be configured to be previewed only within terminal 11 and not automatically transmitted in the background until an explicit operation is performed on transmission unit 113. If necessary, pre-processing such as reduction and recompression may be performed on terminal 11 before transmission to reduce the amount of communication traffic.

[0070] To ensure accessibility, the display unit 111 may be provided with a focus order that allows image selection and transmission to be completed using only the keyboard, and an alternative text input field. This improves the operability of the transmission unit 113 and contributes to reducing input errors.

[0071] On the confirmation and correction screen of FIG. 4, the display unit 111 displays the product description, category information, and attribute information in an easy-to-read format. If necessary, reliability information, inconsistency indications and correction guidelines, and a list of category candidates are also displayed. The user corrects the wording, selects a candidate, and confirms the registration using the correction input unit 112. The confirmation and correction screen of FIG. 4 may be configured to display a list of the generated product information and attribute information, with a correction field (text box, pull-down menu, etc.) to the right of each item. The correction input unit 112 may change the background color or icon depending on the reliability value and display a message prompting correction for items below a predetermined threshold. Furthermore, when an inconsistency is detected, correction candidates may be presented in a pop-up format for the corresponding item, and the user may select one to confirm the content. If multiple category candidates are presented, the selection may be accepted using a pull-down list, radio buttons, etc.

[0072] The display unit 111 may display product description, category information, and attribute information in separate sections as editable areas for each item. The correction input unit 112 may provide input controls such as inline editing for the description, single selection of candidates for category information, and text, pull-downs, and check boxes for attribute information. Edits may be instantly reflected locally on an item-by-item basis, and unsaved items may be visually identified.

[0073] The correction input unit 112 may check the validity of the input value (missing required fields, upper limit of character count, allowable character types, numerical range, etc.), and if there is a violation, may present an inline error on the display unit 111. When the category information is changed, the identification unit 123 may present an option to re-input or regenerate (attribute information) for an item that does not match the set of attribute items identified.

[0074] When presenting the reliability values, the display unit 111 may add color coding, a warning icon, or an explanatory tooltip according to a threshold value. The correction input unit 112 may preferentially guide scrolling and focus movement to items with low reliability, thereby reducing the review burden on the user.

[0075] If an inconsistency is detected, the display unit 111 may highlight the relevant part and display a correction policy (e.g., "Material candidates: cotton / polyester" or "Model number needs to be reconfirmed"). The correction input unit 112 may accept a selection from the suggested candidates, free description, or a request to reacquire only the relevant item (request to regenerate attribute information).

[0076] After the corrections have been confirmed, the correction input unit 112 may accept a "register" operation and transmit the final content including the edited differences (S11A in FIG. 2). If no editing has been performed, the correction input unit 112 may accept a "register as is" operation and transmit the final content after presenting a confirmation dialog (S11B). The display unit 111 may display the progress and result (success, failure, retry) during transmission, and may transition to a completion screen upon receiving a notification from the server 12 that product master registration has been completed (S12).

[0077] To visualize the operation history, the display unit 111 may optionally display the difference between before and after the change (deletion, addition, replacement) for each item. The correction input unit 112 may accept undo and redo of the previous operation, making it easy to recover from an incorrect operation. To ensure accessibility, a focus order and shortcuts may be provided so that all operations can be completed using only the keyboard.

[0078] To confirm in advance whether the data conforms to the external mall specifications, the display unit 111 may display the results of the preliminary check, such as whether or not required attribute items have been satisfied and character count restrictions (e.g., title upper limit). The correction input unit 112 may allow the user to select whether or not to apply automatic formatting (trimming, replacing prohibited symbols, etc.) to satisfy the restrictions.

[0079] 5 is a conceptual example of the input configuration provided to the generation AI 30, showing the input / output relationship between the first cycle (A) and the second cycle (B). In the first cycle (A), the product information generation unit 122 inputs product images and supplemental information to the generation AI 30, which then outputs product information (product name, product description, etc.) in natural language. In the second cycle (B), the attribute information generation unit 124 inputs the category information and attribute items identified by the identification unit 123 to the generation AI 30, and outputs text values ​​(attribute information) to be input into each attribute item.

[0080] In the first cycle shown in Figure 5(A), the generation AI 30 processes both the product image and the supplemental information as input conditions, integrating them in the same inference process. The output consists of natural language text such as the product name and product description, and does not include the generation of category information or attribute information.

[0081] In the second cycle shown in Figure 5(B), the category information and attribute items identified by the identification unit 123 are input, and the generation AI 30 outputs the values ​​of each attribute item (e.g., size, material, country of origin, etc.) as natural language text.

[0082] In the second cycle (B), the generation AI 30 generates attribute information consisting of text to be input into each attribute item, based on the category information identified by the identification unit 123 and the attribute items associated with the category information. At this time, the product information (description, etc.) generated in the first cycle (A) may not be included in the input of the second cycle.

[0083] The control variables of the prompt may include the output language, tone, numeric format, unit of measure, digit separator, symbol formatting, line break convention, etc. The product information generation unit 122 may use these specifications to obtain output in a format that conforms to the display unit 111 and mall linkage specifications.

[0084] If the response from the generation AI 30 does not conform to the specified data structure, the product information generation unit 122 may issue a second instruction (request for format modification) or perform minimal post-structuring processing (e.g., normalizing key names, filling in missing keys), thereby ensuring consistency in mapping processing in downstream processes.

[0085] Corresponding to the two-stage cycle (S2 to S3, S7 to S8 in FIG. 2), the prompt for generating product information and the prompt for generating attribute information may be logically separated, and category information and a set of attribute items may be added to the latter as prerequisites.

[0086] From the viewpoint of localization, the product information generation unit 122 may include a vocabulary policy for each market / mall (e.g., recommended tag expressions for Rakuten Ichiba) in the prompt. This allows for optimization for each mall from the same base information.

[0087] The control parameters of the generation AI 30 (e.g., upper limit of output length, etc.) do not depend on a specific implementation, and the product information generation unit 122 may set general parameters to obtain stable output. These setting values ​​are not limited to fixed values, and may be adjusted based on operational quality indicators.

[0088] FIG. 6 is a diagram illustrating an overview of communication with an external API. The identification unit 123 may extract key words from product information, send a request to a product search API (keyword search), receive the top N results (N is a positive integer), and identify category information based on the results by majority vote. An example of a default value for N is 30, and N may be increased stepwise if it is below a predetermined threshold. Furthermore, in collaboration with the external product database 20, the system may be configured to call (1) the product search API (keyword search), (2) the category information acquisition API, and (3) the attribute item acquisition API in this order. The attribute information generation unit 124 acquires attribute items associated with the identified category and calls the generation AI 30 to generate attribute information consisting of text to be entered into the attribute items. Note that, in this specification, the symbol N (network) and the symbol N (number parameter) for the top N items are the same, but are clearly distinguishable depending on the context.

[0089] The identification unit 123 may send a request to the external product database 20 using at least one or more of a product search API (keyword search), a category information acquisition API, and an attribute item acquisition API. Each request may include locale information (language, currency, region) and authentication information (API key, etc.) according to the market or mall. Hereinafter, in this embodiment, these three types of APIs are collectively referred to as "product search API, etc."

[0090] The identification unit 123 may extract category identifiers of product candidates from the response of the product search API (keyword search), normalize duplicate hits of the same product, and then aggregate the frequency of appearance. This reduces statistical bias in terms of majority rule. If a predetermined threshold is not met, auxiliary rules such as weighting the rankings of top candidates may be applied. Note that if the frequency of appearance is the same, weighting may be applied to the top rankings or a re-search may be performed to identify category information.

[0091] When the identification unit 123 detects a network anomaly or a rate-limited response, it may select one of the following: retry with exponential backoff, retain partial results, or skip or postpone subsequent processing. This allows the robustness of processing within the server 12 to be maintained against failures of external dependencies in S4 to S6 of FIG. 2.

[0092] After the category information is determined, the identification unit 123 may acquire a set of attribute items associated with the category from an attribute item acquisition API in the external product database 20 or from a cache or internal table in the server 12. If both are available, external acquisition may be prioritized, and if acquisition is not possible, internal information may be used as a fallback. Note that if the return specifications of the product search API (keyword search), category information acquisition API, and attribute item acquisition API differ, the server 12 may normalize the item names and value formats according to mapping rules.

[0093] The attribute information generation unit 124 may assign the set of attribute items received from the identification unit 123 as input conditions to the generation AI 30 and request the generation of text for each item. If there are missing items in the response, the items may be temporarily accepted as null values, and the user may then complete them in the screen confirmation (FIG. 4).

[0094] The identification unit 123 may perform mapping between the category identifiers acquired from the external product database 20 and the category system of the management system of the EC site. In this case, the mapping may be changeable via a setting file or a management screen. This allows the system to maintain internal consistency while following changes in external specifications.

[0095] The server 12 may store a summary of external API requests and responses (request type, time, number of requests, processing time, and decision result) as a log. This helps to reset N based on quality indicators and fine-tune the operation of prompt policies.

[0096] Figure 7 shows the detailed flow of category determination. Product information or a group of keywords output by the generation AI 30 is input, and the process is carried out in the following order: main phrase extraction, API call, top N counting, and frequency of appearance majority vote.

[0097] The identification unit 123 may aggregate search results acquired from the external product database 20 after normalizing overlaps of the same product, spelling variations, and language differences (Japanese, English, Roman alphabet, etc.). This prevents results belonging to the same category from being counted in a dispersed manner, improving the stability of determination based on appearance frequency.

[0098] When aggregating results, weighting may be applied according to the ranking of the top N results. For example, a monotonically decreasing function that gives greater weight to the top results can be used to reflect the tendency of the top search results in determining the category. Whether to use this or a simple majority vote without weighting can be selected depending on operational requirements.

[0099] When the appearance frequencies of multiple categories are the same, a tie-breaking rule may be applied, such as mapping the representative category to a higher hierarchy based on the hierarchy information, or determining the representative category using the degree of compatibility (partial match rate, similarity index, etc.) with the main phrase output by the product information generation unit 122. This makes it possible to unify the determination in the case of a tie.

[0100] If the search results do not reach the specified number of items, or if the frequency of occurrence of any category does not reach the specified threshold, a re-search (expanding or contracting the main phrase) can be performed, or a fallback process can be performed in which a predefined general category (e.g., Other) is provisionally adopted and confirmation is requested from the correction input unit 112.

[0101] The identification unit 123 may assign a reliability index to the determined category information. The index is calculated based on the relative difference in appearance frequency, the weighted sum, the number of re-searches, etc., and can be linked to review prioritization on the display unit 111.

[0102] If there are multiple external product databases 20, the top N results may be tallied for each database, and category candidates may be extracted, and the final category information may be integrated using vote integration or reliability weighting. This allows for decisions that are less susceptible to bias in individual databases.

[0103] Note that the value of N is not limited to a fixed value, but may be variable as described above. For example, a stepwise search may be performed in which a small value of N is used initially, and if the reliability is less than a threshold value, N is gradually increased.

[0104] The determined category information may depend on the category system of the external product database 20, and may be mapped to an internal standard category system as needed. This allows the attribute information generating unit 124 to consistently acquire and organize attribute items.

[0105] If the query to the external product database 20 fails or a rate limit is detected, the most recent decision history recorded in the cache or the previous decision result with a time limit may be provisionally applied, and this may be notified to the display unit 111 and confirmation may be obtained from the correction input unit 112.

[0106] Figure 8 shows an example of the process of detecting inconsistencies and prompting corrections. The output of the generation AI 30 is compared and judged, a correction message is generated as necessary, and after confirmation input from the user, the correction is reflected or the final output is reached.

[0107] When determining whether there is an inconsistency, the correction input unit 112 may compare the item value output by the product information generation unit 122 with the keywords in the supplemental information and the category information determined by the identification unit 123. The comparison can be implemented by known methods such as vocabulary matching, a synonym dictionary, regular expressions, validity of a numerical range, unit consistency, etc.

[0108] The types of inconsistencies may be classified into value discrepancies (e.g., color and size), omissions (e.g., no brand listed), format violations (e.g., digits and units), category inconsistencies (e.g., home appliance attributes assigned to clothing category), etc. The correction input unit 112 presents template wording according to the type on the display unit 111 and guides the user through the correction operation.

[0109] The display unit 111 may visualize the inconsistency with highlighting, a warning icon, a tooltip, or the like depending on its importance. The importance may be calculated based on the essentiality of the relevant item, its impact on sales, frequency of occurrence, etc., and the most important inconsistencies may be displayed preferentially at the top of the screen. Here, the impact on sales refers to the degree to which practical disadvantages may occur in the display ranking, exposure opportunities, or conversion rate of the product page due to non-fulfillment of essential items, the possibility of a violation of the listing terms, or the absence of customer-attracting terms (e.g., main search terms, promotional terms), etc.

[0110] The correction input unit 112 may automatically suggest candidate values. The candidate values ​​are presented based on other outputs of the generation AI 30, past confirmation history, representative values ​​in the external product database 20, etc., and the user can confirm them by selection or free input. Detected inconsistencies may be prioritized based on their impact or frequency, and the presentation order may be controlled.

[0111] If the inconsistency is not resolved, the correction input unit 112 may set a hold flag for the item and issue a warning at the time of registration. In the hold state, product master registration may be possible with only the minimum required items, and a message urging the user to add the items at a later date may be displayed on the display unit 111.

[0112] The results of the correction operation may be recorded in the server 12 as an audit log. The log includes information such as the detection type, initial value, final value, operator, and time, which will be useful for future verification and quality improvement. Here, "operator" refers to information including the identification information of the user who performed the correction operation (user ID, terminal identifier, account name, etc.). This makes it possible to clarify the scope of responsibility for each correction operation, improving the auditability and operational manageability of the entire product registration process. If necessary, the data may be anonymized and aggregated to generate a statistical report.

[0113] The detection and resolution status of inconsistencies may be reflected in the reliability value. For example, if there are any unresolved high importance inconsistencies, the reliability is decremented, and when all inconsistencies are resolved, the reliability is recalculated and updated and displayed on the display unit 111.

[0114] The reliability value may be determined based on a predetermined threshold value, which is set appropriately depending on the application.

[0115] When multiple inconsistencies occur in a chain (for example, when a group of attributes changes due to category reselection), the correction input unit 112 may analyze the dependency relationships and present a list of items to be recalculated. This can prevent duplicate corrections and loss of consistency.

[0116] If the generation AI 30 or the external product database 20 is temporarily unavailable, the server 12 may temporarily present the most recent cached attribute candidate or default template, and may perform fallback processing by clearly indicating this on the display unit 111 and requesting the user's confirmation.

[0117] To ensure accessibility, the display unit 111 may be provided with a display that takes color vision diversity into consideration, corrections that can be made only by keyboard operations, voice-readable labels, etc. This can improve the reliability of correction operations and work efficiency.

[0118] The threshold value and target items for inconsistency detection may be adjustable from the management screen in accordance with the operation policy. The setting changes are stored in the server 12 and dynamically reflected in the processing of the product information generation unit 122 and the identification unit 123. The setting changes may also include the weighting of the evaluation factors used to select category candidates. Specifically, the weighting coefficients and judgment thresholds for the appearance frequency, search score value, and output reliability of the generation AI for each category candidate may be adjusted from the management screen, and the adjustment results may be stored in the server 12 and dynamically reflected in the processing of the product information generation unit 122 and the identification unit 123. This makes it possible to identify category information based on an index obtained by combining the evaluation factors using a predetermined weighting rule (hereinafter, this index will be referred to as the "integrated score").

[0119] In addition to the above-described embodiment, the following supplements additional operational effects of each subordinate configuration, all of which are based on specific cooperation between the components (display unit 111, correction input unit 112, transmission unit 113, reception unit 121, product information generation unit 122, identification unit 123, attribute information generation unit 124, external product database 20, and generation AI 30).

[0120] By displaying product information on the display unit 111 and receiving correction instructions through the correction input unit 112, corrections can be reflected on the spot while leaving the confirmation of the generated content to manual review. This allows stable finalization in accordance with the operational policy and notation regulations of each mall where the store is located.

[0121] The correction input unit 112 displays the reliability value on the display unit 111 and prompts correction when the reliability value is below a predetermined threshold, which makes it possible to visualize the review priority. This allows the person in charge to concentrate on cases that require attention, ensuring both quality and shortening the processing time.

[0122] By detecting inconsistencies between estimates based on product images and supplementary information and presenting correction guidelines, it is possible to quickly identify discrepancies in color, model number, brand, etc. This makes it possible to prevent incorrect registration and present guidelines for correction work.

[0123] The identification unit 123 presents multiple category candidates on the display unit 111, and the selection is confirmed by the correction input unit 112. This allows the domain knowledge of the person in charge to be incorporated into the decision-making process. As a result, it is possible to simultaneously improve classification accuracy and work efficiency.

[0124] The product information generation unit 122 uses prompts or keywords in input to the generation AI 30, which stabilizes the format, granularity, tone, and comprehensiveness of required items in the output. This reduces the need for post-processing formatting adjustments and item mapping revisions.

[0125] The product information generation unit 122 receives output in a predetermined data structure and stores it by item, which facilitates automatic mapping to product master items and each mall linkage item. This improves the interoperability and maintainability of API / CSV linkage.

[0126] The identification unit 123 is configured to analyze multiple category candidates included in the search results of the external product database 20 using statistical criteria, thereby enabling category determination that is robust against individual noise. This makes it possible to suppress misclassification and ensure the reproducibility of the determination.

[0127] By using a majority vote based on the frequency of appearance of category candidates in the top N search results, it is possible to reduce the impact of outliers and make decisions with a reduced amount of calculation, ensuring an appropriate balance between operational response speed and accuracy.

[0128] Category determination may be performed using at least one of the following methods: majority voting, weighted voting, or rank integration. Category information may also be determined using a score that combines these with the reliability value of the generating AI. This configuration enables highly reproducible category determination that is robust against individual noise and database bias.

[0129] The attribute information generation unit 124 generates attribute text using the generation AI 30 based on category information and attribute items, which allows automatic completion of descriptions such as size, material, target, country of origin, etc. This reduces the time required to fill in attribute fields and ensures consistency of expression.

[0130] This two-stage process, in which product information is determined in the first cycle and attribute information is generated in the second cycle, allows the second stage to be optimized based on the results of the first stage, making it possible to localize error locations, flexibly rerun the process, and produce a highly consistent final output.

[0131] By configuring the description generation (first cycle) and attribute generation (second cycle) to be independent (non-dependent) from each other and executing them so that the output of the first cycle is not included in the input of the second cycle, the effect of errors can be localized and processing time can be reduced by re-executing only the necessary parts.

[0132] By extracting key words from the results of the product information generation unit 122 and inputting them as search terms in the external product database 20, it is possible to achieve both accuracy and responsiveness compared to a word-for-word search, thereby improving the category hit rate and reducing search costs.

[0133] The product information generation unit 122 and the identification unit 123 are configured to call the APIs of the external generation AI 30 and the external product database 20, so that the latest generation capability and product catalog information can be continuously retrieved. This enables scalable operation while reducing the maintenance load on the server 12 side.

[0134] The additional effects based on the above-mentioned dependent configuration not only provide the basic effects (simplification of input and improvement of generation quality), but also gradually improve review efficiency, classification stability, consistency of attribute completion, and scalability of linked operations. This ensures a high level of overall processing quality and productivity in actual operations.

[0135] <2. Modifications> The product information generating system, server, and program according to the present invention are not limited to the above-described embodiments, and various modifications and improvements are possible within the scope of the claims.

[0136] In the above-described embodiment, the order in which the product image and the supplemental information are input is arbitrary, and the supplemental information may be analyzed first and used for prompt control, and then the image may be input.

[0137] Furthermore, the criteria for identifying a category are not limited to majority voting, and ranking weighting, scoring, threshold value determination, etc. can also be applied.

[0138] In addition, attribute items can be obtained by referencing internal server data or by using an external attribute API.

[0139] Furthermore, confirmation and correction input may be accepted on the local screen as well as the cloud management screen.

[0140] Furthermore, if the product information generation unit 122 and the attribute information generation unit 124 share the same calling procedure and prompt structure as the generation AI 30, the division of functions is merely an example, and they may be integrated into a single generation unit. In this case, the generation AI 30 may simultaneously generate product descriptions, category information, and attribute information according to the input information, and the server 12 may automatically identify and distribute the type of each piece of information.

[0141] Alternatively, the terminal 11 may be configured to execute part or all of the category identification process. In this case, the terminal 11 temporarily stores the output content of the generation AI 30 and can use local computing resources to extract key words or perform majority voting. This reduces the communication load and enables part of the process to be performed offline.

[0142] Furthermore, the configuration of the server 12 can be applied to any of a cloud environment, an on-premise environment, or a hybrid configuration thereof.

[0143] When identifying a category, instead of using a majority vote for the top N results, the average score or maximum weighted score of each candidate category may be used. Also, the reliability of the results returned from the external product database 20 may be applied as a weight.

[0144] The reliability value may be calculated using a statistical reliability index (e.g., a known index such as a Jaccard coefficient, a BLEU score, or a similarity index) calculated by the server 12, in addition to the probability value returned by the generation AI 30. Such an index can also be used as basic data for evaluating the consistency of the generated content and for re-learning feedback.

[0145] The mismatches can be indicated not only by highlighting them in the text but also by voice guidance or a chatbot-type assistive user interface, in which case the user can give correction instructions via voice or natural language input.

[0146] In addition, the correction results and reliability logs can be visualized in a dashboard format or output in CSV format and sent to an external system, allowing the operations manager to continuously monitor the registration accuracy and generation quality.

[0147] The trigger for the product information generation process may be automatically initiated by scheduled execution or batch processing, rather than by user operation. This allows for efficient automatic generation even when updating a large amount of product data.

[0148] The type of generated AI 30 is not limited to a specific vendor or model, and may utilize a local inference model, an API service, or a lightweight model on an edge device.

[0149] The data format of the external product database 20 may be any structure, such as a relational database, a NoSQL database, or a spreadsheet.

[0150] Furthermore, the generated product information may be translated into multiple languages ​​and registered simultaneously in multiple malls. In this case, the generation AI 30 applies a translation model or a multilingual generation model for each language to generate product descriptions suitable for the local market.

[0151] These modifications do not deviate from the technical concept of the present invention, and similarly achieve the object of the present invention of improving the efficiency of automatic generation, confirmation, and registration of product information. [Explanation of symbols]

[0152] 10... Product information generation system 11... User terminal 11A, 11B ... Laptop (example) 111...Display section 112 ... Correction input section 113 ... Transmitter 12 … Server 121 ... Receiver 122 … Product information generation department 123 … Specific section 124 … Attribute information generation unit 20 … External product database 30…Generation AI N...Network S1~S12 ... Processing steps

Claims

1. A product information generation system, A user terminal and a server capable of communicating with the user terminal via a network, the user terminal includes a transmission unit that receives product images and supplemental information related to the products and transmits the images to the server; The server a receiving unit that receives the product image and the supplemental information; a product information generation unit that inputs the product image and the supplemental information into a generation AI, integrates and processes both in the same inference process by the generation AI, and generates product information including at least a product description in natural language; an identification unit that searches an external product database based on the product information generated by the product information generation unit, identifies category information, and identifies attribute items associated with the category information; Equipped with The product information generation system is characterized in that the identification unit identifies the category information based on an integrated score calculated by integrating the appearance frequency, score value, and output reliability of the generation AI of each of multiple category candidates obtained from the search results of the external product database.

2. In the system described in claim 1, The product information generation system is characterized in that the identification unit extracts key words from the product information generated by the product information generation unit and inputs the key words as search terms into an external product database search API.

3. A product information generation system, A user terminal and a server capable of communicating with the user terminal via a network, the user terminal includes a transmission unit that receives product images and supplemental information related to the products and transmits the images to the server; The server a receiving unit that receives the product image and the supplemental information; a product information generation unit that inputs the product image and the supplemental information into a generation AI, integrates and processes both in the same inference process by the generation AI, and generates product information including at least a product description in natural language; an identification unit that searches an external product database based on the product information generated by the product information generation unit, identifies category information, and identifies attribute items associated with the category information; Equipped with The product information generation system is characterized in that the server further comprises an attribute information generation unit that generates attribute information consisting of text to be input into the attribute items using the generation AI based on the category information and attribute items identified by the identification unit.

4. A product information generation system, further comprising a configuration in which the server performs generation using the generation AI as a process including (i) a first cycle of generating the product information based on the product image and the supplementary information, and (ii) a second cycle of generating text to be input into each attribute item based on the category information and the attribute item, and the output of the first cycle is not included in the input of the second cycle.

5. In the system described in claim 3, the server transmits the product information generated by the product information generation unit and the attribute information generated by the attribute information generation unit to the user terminal; The product information generating system is characterized in that the user terminal displays the product information and attribute information on a display unit and is equipped with a correction input unit that accepts correction instructions from the user as necessary.

6. In the system described in claim 5, The correction input unit quantifies a reliability value indicating the likelihood of the estimation result by the generation AI and displays it on the display unit, and if the reliability value is below a predetermined threshold, presents a message to the user urging them to make corrections. This is a product information generation system characterized by the above.

7. In the system described in claim 6, The display unit calculates the importance of each detected inconsistency based on its degree of necessity, its degree of impact on sales, and its frequency of occurrence, A product information generation system characterized by presenting inconsistencies in the form of highlights or messages according to the importance.

8. In the system described in claim 5, The product information generation system is characterized in that, when an inconsistency is detected between the estimated content based on the product image and the supplementary information, the correction input unit presents a message indicating the possibility and correction policy depending on the type and importance of the inconsistency.

9. In the system described in claim 8, The product information generating system is characterized in that the importance is an evaluation value calculated based on the degree of necessity, the degree of sales impact, and the frequency of occurrence.

10. In the system described in claim 5, The product information generation system is characterized in that the correction input unit presents a plurality of category candidates obtained by the identification unit, accepts a user's selection, and confirms the category information based on the selection.

11. In the system described in claim 10, The server records the operation result of the correction input unit as an audit log, A product information generation system characterized by having a configuration in which the logs are statistically compiled and used to re-learn the reliability threshold or the output parameters of the generation AI.

12. A server capable of communicating with a user terminal via a network, a receiving unit that receives product images and supplemental information related to the products from the user terminal; a product information generation unit that inputs the product image and the supplemental information into a generation AI, integrates and processes both in the same inference process by the generation AI, and generates product information including at least a product description in natural language; an identification unit that searches an external product database based on the product information generated by the product information generation unit, and identifies category information and attribute items associated with the category information; an attribute information generation unit that generates attribute information consisting of text to be input into the attribute items using the generation AI based on the category information and the attribute items identified by the identification unit; A product information generation server comprising:

13. A server that can communicate with user terminals via a network, (a) receiving product images and supplemental information about the products from the user terminal; (b) inputting the product image and the supplemental information into a generation AI, and integrating and processing both in the same inference process by the generation AI to generate product information including at least a product description in natural language; (c) a process of searching an external product database based on the product information generated in the process (b), identifying category information, and identifying attribute items associated with the category information; (d) generating attribute information consisting of text to be input into the attribute items using the generation AI based on the category information and the attribute items identified by the process of (c); A program to execute.

14. A product information generation system, image input means for inputting product images; a supplemental information input means for inputting supplemental information about the product; A generation means for inputting the product image and the supplemental information into a generation AI, and integrating and processing both in the same inference process by the generation AI to generate a product description in natural language; a specifying means for searching an external product database based on the product information generated by the generating means, specifying category information, and specifying attribute items associated with the category information; an attribute information generating means for generating attribute information consisting of text to be input into the attribute items using the generation AI based on the category information and the attribute items identified by the identifying means; A product information generation system comprising:

Images