Character-level Analysis of Semi-Structured Documents for Set Expansion

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1503–1512, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Character-level Analysis of Semi-Structured Documents for Set Expansion Richard C. Wang Language Technologies Institute Carnegie Mellon University rcwang@cs.cmu.edu William W. Cohen Machine Learning Department Carnegie Mellon University wcohen@cs.cmu.edu Abstract Set expansion refers to expanding a par- tial set of “seed” objects into a more com- plete set. One system that does set ex- pansion is SEAL (Set Expander for Any Language), which expands entities auto- matically by utilizing resources from the Web in a language-independent fashion. In this paper, we illustrated in detail the construction of character-level wrappers for set expansion implemented in SEAL. We also evaluated several kinds of wrap- pers for set expansion and showed that character-based wrappers perform better than HTML-based wrappers. In addition, we demonstrated a technique that extends SEAL to learn binary relational concepts (e.g., “x is the mayor of the city y”) from only two seeds. We also show that the extended SEAL has good performance on our evaluation datasets, which includes English and Chinese, thus demonstrating language-independence. 1 Introduction SEAL1 (Set Expander for Any Language) is a set expansions system that accepts input ele- ments (seeds) of some target set S and automat- ically finds other probable elements of S in semi- structured documents such as web pages. SEAL is a research system that has shown good perfor- mance in previously published results (Wang and Cohen, 2007). By using only three seeds and top one hundred documents returned by Google, SEAL achieved 90% in mean average precision (MAP), averaged over 36 datasets from three lan- guages: English, Chinese, and Japanese. Un- like other published research work (Etzioni et al., 2005), SEAL focuses on finding small closed sets 1http://rcwang.com/seal of items (e.g., Disney movies) rather than large and more open sets (e.g., scientists). In this paper, we explore the impact on perfor- mance of one of the innovations in SEAL, specif- ically, the use of character-level techniques to de- tect candidate regular structures, or wrappers, in web pages. Although some early systems for web-page analysis induce rules at character-level (e.g., such as WIEN (Kushmerick et al., 1997) and DIPRE (Brin, 1998)), most recent approaches for set expansion have used either tokenized and/or parsed free-text (Carlson et al., 2009; Talukdar et al., 2006; Snow et al., 2006; Pantel and Pennac- chiotti, 2006), or have incorporated heuristics for exploiting HTML structures that are likely to en- code lists and tables (Nadeau et al., 2006; Etzioni et al., 2005). In this paper, we experimentally evaluate SEAL’s performance under two settings: 1) us- ing the character-level page analysis techniques of the original SEAL, and 2) using page analy- sis techniques constrained to identify only HTML- related wrappers. Our conjecture is that the less constrained character-level methods will produce more candidate wrappers than HTML-based tech- niques. We also conjecture that a larger number of candidate wrappers will lead to better performance overall, due to SEAL’s robust methods …

Content truncated to 500 words