Friday April 19
Ellen Riloff
University of Utah
Learning Multiple Semantic Categories Simultaneously using Collective
Evidence over Extraction Patterns

We will present a bootstrapping algorithm called Basilisk that learns
semantic lexicons for multiple categories simultaneously. Basilisk
begins with an unannotated text corpus and a small set of seed words
for each semantic category, which are then bootstrapped to learn new
words for each category. Basilisk uses two original ideas that
distinguish it from previous bootstrapping techniques for semantic
lexicon induction.  (1) Basilisk gathers collective statistical
information over a large body of extraction patterns. (2) Basilisk
learns multiple semantic classes simultaneously, which constrains the
bootstrapping process.  We evaluate Basilisk on six semantic
categories using the MUC-4 text corpus. Our results show that the
semantic lexicons produced by Basilisk have higher precision than
those produced by previous techniques, with several categories showing
substantial improvement.  We also present results demonstrating that
both of Basilisk's contributions (collective statistics over
extraction patterns, and simultaneous bootstrapping) produce
performance gains independently.