A Dataset for Information Extraction from News Web Pages

In this paper we propose a new dataset for information extraction from news web pages. Accurate collection of news articles is necessary to build systems that aggregate and analyze data from a large number of news sources. However, existing news page datasets are designed for the problem of article content extraction and do not consider other metadata such as dates or authors, while modern generic approaches for structured web data extraction are evaluated on other domains, mostly e-commerce products websites. Our dataset contains 724 web pages from 114 Russian news websites. On each page we manually annotated title, text, publication date, tags and other article attributes. We describe data collection and annotation process and demonstrate evaluation results for open source tools and neural network based approaches. The dataset is publicly available on GitHub.