WIERT: Web Information Extraction via Render Tree

Web information extraction (WIE) is a fundamental problem in web document understanding, with significant impact on various applications. Visual plays crucial role WIE tasks as the nodes containing relevant are often visually distinct, such being larger font size or having brighter color, from other nodes. However, rendering visual of page can be computationally expensive. Previous works have mainly focused Document Object Model (DOM) tree, which lacks information. To efficiently exploit information, we propose leveraging render combines DOM tree and Cascading Style Sheets (CSSOM) contains not only content layout but also rich at little additional acquisition cost compared to tree. In this paper, present WIERT, method that effectively utilizes based pretrained language model. We evaluate WIERT Klarna product dataset, manually labeled dataset renderable e-commerce pages, demonstrating its effectiveness robustness.