Bootstrapping Knowledge Base Acceleration

The Streaming Slot Filler (SSF) task in TREC Knowledge Base Acceleration track involves detecting changes to slot values (relations) over time. To handle this task, the system needs to extract relations to identify slot-filler values and detect novel values. Being the first attempt at KBA, the biggest challenge that we faced was the scale of the data. We present the approach used by University of Wisconsin for the SSF task and the large scale challenge. We used Elementary, a scalable statistical inference and learning system, developed in University of Wisconsin as our core system. We used Stanford NLP Toolkit to generate parse trees, dependency graphs and named-entity recognition information. These were then converted to features for the logistic regression learner of Elementary. To handle the lack of early SSF training data, we used our existing Knowledge Base Population system to bootstrap a logistic regression model and added rules to handle the new relations.