UBC at Slot Filling TAC-KBP 2010

The system is based on the distant supervision strategy presented by Mintz et al., 2009, trained using snippets of the document collection containing both entity and filler from the KB provided by the organizers (a subset of Wikipedia infoboxes).

Abstract

This paper describes our submissions for the slot filling and surprise slot filling tasks of TAC-KBP. The system is based on the distant supervision strategy presented by (Mintz et al., 2009). We did a straightforward implementation, trained using snippets of the document collection containing both entity and filler from the KB provided by the organizers (a subset of Wikipedia infoboxes). Our system does not use any other external knowledge source, with the exception of closed lists of words for religion, causes of death, charges and religious/political affiliation, plus the use of Geonames to distinguish between cities, countries, states and provinces. We submitted three runs based on different post-processing options of the output of our classifiers, with results below the median. We did expect low results, as our system is still under development, and we still have plenty of room for improvement. 1