Accelerating the Sequence Annotation using Split Annotation with Active Learning

Active learning succeeds in reducing the size of labeled corpus while maintaining the high accuracy. However, active learning requires several iterations of the tagger training, which will not be practical when the training of an iteration takes long time. In this paper, we propose to simplify the all-entity labeling task by splitting the task into a set of single-entity labeling subtasks. After all entity types are labeled, we merge the data sets into an all-entity corpus and train the final tagger using the merged set. The proposed method achieved the competitive F1 to the multientity learning but required much less computational time on the CoNLL chunking and named entity recognition data sets.