CMUML System for KBP 2013 Slot Filling

In this paper, we present an overview of the CMUML system for KBP 2013 English Slot Filling (SF) task. The system used a combination of distant supervision, stacked generalization and CRF-based structured prediction. Recently available anchor text data was also used for better entity matching. The system takes a modular approach so that independently developed semantic annotators can be effectively integrated without needing target ontology-specific retraining. While precision can of course be improved, the system turned out to be particularly conservative in its predictions resulting in lower recall. In addition to the main submission, we also made publicly available1 automatically tagged semantic categories of about 13 million noun phrases extracted from the KBP 2013 source corpus.