Iterative Learning for Semi-automatic Annotation Using User Feedback

With the advent of state-of-the-art models based on Neural Networks, the need for vast corpora of accurately labeled data has become fundamental. However, building such datasets is a very resource-consuming task that additionally requires domain expertise. The present work seeks to alleviate this limitation by proposing an interactive semi-automatic annotation tool using an incremental learning approach to reduce human effort. The automatic models used to assist the annotation are incrementally improved based on user corrections to better annotate the next data. To demonstrate the effectiveness of the proposed method, we build a dataset with named entities and relations between them related to the crime field with the help of the tool. Analysis results show that annotation effort is considerably reduced while still maintaining the annotation quality compared to fully manual labeling.