Sgardine writesup Wang Cohen 2009

This is a review of Wang_2009_automatic_set_instance_extraction_using_the_web by user:Sgardine

Summary

The presented system ASIA makes use of the previous SEAL system to expand a set of seeds into a target set; the seeds are automatically detected by hyponym extraction from a seed semantic class name. Given a class name and a set of hyponym extraction patterns (language-dependent but class-independent) the Noisy Instance Provider finds several candidate seeds. The set is expanded using a variant of SEAL which tolerates noise by demanding only that two or more seeds cooccur, rather than all seeds. The discovered instances are then bootstrapped using iSEAL. The system is evaluated on the lists from the SEAL papers in three languages, and is found to outperform previous methods.

Commentary

I thought the hypernym hierarchy discussion was interesting -- it seems like the "true" hierarchy as represented on the web is very noisy also because people can conceptualize things differently in different contexts. Interesting that ASIA picks out one layer down in the hierarchy rather than finding leaves