An Efficient Active Learning Framework for New Relation Types

Supervised training of models for semantic relation extraction has yielded good performance, but at substantial cost for the annotation of large training corpora. Active learning strategies can greatly reduce this annotation cost. We present an efficient active learning framework that starts from a better balance between positive and negative samples, and boosts training efficiency by interleaving self-training and co-testing. We also studied the reduction of annotation cost by enforcing argument type constraints. Experiments show a substantial speed-up by comparison to the previous state-of-the-art pure co-testing active learning framework. We obtain reasonable performance with only 150 labels for individual ACE 2004 relation types.