Reducing semantic drift in bootstrapping for entity relation extraction

Abstract

This paper presents a novel bootstrapping algorithm for entity relation extraction. Shortest dependency patterns connecting entity pairs in sentences are captured initially and in turn applied to extract new binary relationships. The patterns are evaluated through correlation detection. In addition, we effectively prevent semantic drift by co-training with trigger words. Experiments for slot filling on the Knowledge Base Population (KBP) newspaper corpora show that our enhanced bootstrapping system achieves an 11% F1-score improvement over traditional bootstrapping algorithm.