English Slot Filling with the Knowledge Resolver System.

This paper describes the Knowledge Resolver system (KRes) and its performance on the TAC-KBP 2013 English Slot Filling task. KRes is a logic-based inference system aimed at improving statistical relation extraction by deduction and abduction inference towards the best document-level interpretation. For the 2013 evaluation we developed an initial KRes system that extracts a subset of seven TACKBP relations using manually constructed dependency patterns in concert with entity type and name-linking rules. For our baseline extraction engine we used the Blender Lab’s KBP-Toolkit 1.5, which was also exploited at the front-end of KRes for its document indexing, selection and name expansion capabilities. Instead of trying to improve upon KBPToolkit results using inference, for this year we simply combined its results with those of KRes for our best system which landed us in the middle of the pack (only addressing 13 out of the 40 KBP slot types). We also report results for KRes relativized to the seven slot types it addressed which shows promise for future evaluations.