Zero-shot Entity Extraction from Web Pages

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 391–401, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Zero-shot Entity Extraction from Web Pages Panupong Pasupat Computer Science Department Stanford University ppasupat@cs.stanford.edu Percy Liang Computer Science Department Stanford University pliang@cs.stanford.edu Abstract In order to extract entities of a fine-grained category from semi-structured data in web pages, existing information extraction sys- tems rely on seed examples or redundancy across multiple web pages. In this paper, we consider a new zero-shot learning task of extracting entities specified by a natural language query (in place of seeds) given only a single web page. Our approach de- fines a log-linear model over latent extrac- tion predicates, which select lists of enti- ties from the web page. The main chal- lenge is to define features on widely vary- ing candidate entity lists. We tackle this by abstracting list elements and using aggre- gate statistics to define features. Finally, we created a new dataset of diverse queries and web pages, and show that our system achieves significantly better accuracy than a natural baseline. 1 Introduction We consider the task of extracting entities of a given category (e.g., hiking trails) from web pages. Previous approaches either (i) assume that the same entities appear on multiple web pages, or (ii) require information such as seed examples (Etzioni et al., 2005; Wang and Cohen, 2009; Dalvi et al., 2012). These approaches work well for common categories but encounter data sparsity problems for more specific categories, such as the products of a small company or the dishes at a lo- cal restaurant. In this context, we may have only a single web page that contains the information we need and no seed examples. In this paper, we propose a novel task, zero- shot entity extraction, where the specification of the desired entities is provided as a natural language query. Given a query (e.g., hiking seeds Avalon Super Loop Hilton Area traditional answers Avalon Super Loop Hilton Area Wildlands Loop ... query hiking trails near Baltimore our system answers Avalon Super Loop Hilton Area Wildlands Loop ... web pages web pages web pages web page Figure 1: Entity extraction typically requires ad- ditional knowledge such as a small set of seed ex- amples or depends on multiple web pages. In our setting, we take as input a natural language query and extract entities from a single web page. trails near Baltimore) and a web page (e.g., http://www.everytrail.com/best/ hiking-baltimore-maryland), the goal is to extract all entities corresponding to the query on that page (e.g., Avalon Super Loop, etc.). Figure 1 summarizes the task setup. The task introduces two challenges. Given a single web page to extract entities from, we can no longer rely on the redundancy of entities across multiple web pages. Furthermore, in the zero-shot learning paradigm (Larochelle et al., 2008), where entire categories might be unseen during training, the system must generalize to new queries and web pages without the additional aid of …

Content truncated to 500 words