Generalizing from Freebase and Patterns using Cluster-Based Distant Supervision for TAC KBP Slotfilling 2012.
Benjamin Roth, Grzegorz Chrupała, Michael Wiegand, Mittul Singh, Dietrich Klakow · Theory and applications of categories · 2012
For the slot filling task of TAC KBP 2012 we extended last year’s system in several respects. The core of the system is a set of semisupervised per-relation classifiers, trained by a scheme known as distant supervision. Training data are generated by using Freebase and applying patterns. Relation models rely on (1) word clusters generalizing from context surface forms and (2) additional argument-level features. For the retrieval of answer candidates, we use document retrieval in combination with an entity expansion model based on Wikipedia link texts. We do not use a separate sentence retrieval step and rely entirely on the classifier for filtering out bad candidates. Our system does not rely on any syntactic analysis or co-reference resolution. The best-ranked run of the full system achieves an F-score of 23.4% on the official test queries.