Active Learning Based Semi-automatic Annotation of Event Corpus

In the area of Natural Language Processing, building corpus by hand was a hard and time-consuming task. Active learning promised to reduce the cost of annotating dataset for it was allowed to choose the data from which it learned. This study presented a semi-automatic annotation method based on active learning for labeling events in Chinese text. Particularly, it focused on uncertainty-based sampling and query-by-committee based sampling algorithm to evaluate which instance was informative and could be labeled by hand in the unlabeled dataset. The selected informative instances were labeled manually for obtaining a more effective classifier. Experimental results not only demonstrated that active learning improved the accuracy of Chinese event annotation, but also showed that it reduced the number of labeling actions dramatically.