Process 1: Reading and extracting product attribute information

In the process of reading and extracting product attribute information, product attribute information is extracted from data in various formats (Word, Excel, PDF, images, etc.). This process is handled by a combination of multiple AI models.

First, in order to detect the general information contained in the data, it is processed by (1) the area detection AI model. The area detection AI model combines the object detection AI model, decomposes the information contained in the data into "image area", "text area", "line area", "table (table) data area", etc., and divides each area into detect. In some cases, AI-OCR models are used in this process. Each detected information is organized and sent to the next process.

Next, in order to select information and associate it with product attribute information, ② attribute classification AI model processing is performed. The attribute classification model creates separate models for images and text, but combines multiple AI models for processing.

Image information is classified by attribute for each type of image using an attribute classification AI model for images. An example of a combination of image attribute classification AI models is EfficientNet and ResNet.

Text information is processed by combining an attribute classification AI model for text and regular expressions. The attribute classification AI model for text uses a natural language processing AI model to learn the relevance of the extracted text and classify which product attribute information it is. The attribute classification AI model for text is developed individually for each product attribute information. Therefore, it takes time to develop the model, but it achieves high accuracy. An example of a combination of text attribute classification models is BERT and LSTM.

Next, (3) processing is performed by the page layout analysis AI model, and the product attribute information of text and images is grouped by product. The page layout analysis AI model is developed individually for each layout, and links products and product attributes according to the page layout. For example, if the product attribute information in the text is on the right side of the image and if it is on the bottom of the image, the layout will be different, and it is necessary to apply a separate AI model, but the product attribute information in the text Both are listed on the right side of the image, and if the combination or order of product attribute information is different, the layout will be the same and processing will be performed with a common AI model.

3rd summary