New York University 2012 System for KBP Slot Filling

摘要

This paper describes the New York University 2012 system for the KBP regular slot filling (SF) task. The NYU 2012 SF system has a similar architecture to the NYU 2011 system. We improved our distant-supervision based slot-filling component with a few techniques including filtering errors by statistical measures collected from the source corpus, and relabeling erroneous training examples with a set of maximum entropy models and by applying bootstrapped/hand-coded patterns. We also augmented our query expansion procedure. After the formal evaluation we experimented with models for estimating slot confidence. We report on the impact of these changes.