Simplified DOM Trees for Transferable Attribute Extraction from the Web

Abstract

There has been a steady need to precisely extract structured knowledge from the web (i.e. HTML documents). Given a web page, extracting a structured object along with various attributes of interest (e.g. price, publisher, author, and genre for a book) can facilitate a variety of downstream applications such as large-scale knowledge base construction, e-commerce product search, and personalized recommendation. Considering each web page is rendered from an HTML DOM tree, existing approaches formulate the problem as a DOM tree node tagging task. However, they either rely on computationally expensive visual feature engineering or are incapable of modeling the relationship among the tree nodes. In this paper, we propose a novel transferable method, Simplified DOM Trees for Attribute Extraction (SimpDOM), to tackle the problem by efficiently retrieving useful context for each node by leveraging the tree structure. We study two challenging experimental settings: (i) intra-vertical few-shot extraction, and (ii) cross-vertical fewshot extraction with out-of-domain knowledge, to evaluate our approach. Extensive experiments on the SWDE public dataset show that SimpDOM outperforms the state-of-the-art (SOTA) method by 1.44% on the F1 score. We also find that utilizing knowledge from a different vertical (cross-vertical extraction) is surprisingly useful and helps beat the SOTA by a further 1.37%.

DOM-LM: Learning Generalizable Representations for HTML Documents

“…We use their code released on GitHub 2 and experiment with our settings. Also note that for all models, we do not apply extra post-processing such as site-level voting in original SimpDOM [29].…”

“…Unlike the page-level F1 scores used in previous work [17,29] that scores the extraction on a page as success as long as one of the predictions is correct, attribute value-level F1 considers the case where there are multiple values for the same attribute in a webpage, and penalizes the false positive predictions made by the model 3 .…”

“…For the zero-shot setting, we split the clusters into training, development and testing follow a 60:20:20 ratio. We also include SimpDOM [29] as a baseline, which is a state-of-the-art extraction model. We use their code released on GitHub 2 and experiment with our settings.…”

“…BERT utilizes transformer [26] to encode the input sentences, and pre-trains the model with masked language model (MLM) objective that predicts randomly masked tokens in the input. However, directly applying BERT to HTML documents may lead to unsatisfying results [29]. The transformer encoder in BERT is designed for unstructured text and cannot handle structured input like HTML documents.…”

“…However, not only do these visual features incur additional computational cost to the model, but also the rendering process is time-consuming and requires extra space to store the necessary images, CSS, JavaScript files, making it hard to be applied in practice. More recently, Sim-pDOM [29] directly utilizes the underlying DOM tree structure, while avoiding rendering-based features. It augments the node representation with relevant context retrieved from the DOM tree and has shown some promising results capturing website and domain invariant features.…”