Partial Annotation Scheme for Active Learning on Named Entity Recognition Tasks

Abstract

Active learning is a promising approach to alleviate the expensive annotation cost for making training data on named entity recognition (NER) tasks. However, since existing active learning methods on NER tasks implicitly assume the full annotation scheme of which the unit of an annotation request is the whole sentence, the efficiency of the data instance selection is limited. In this paper, we propose a new active learning method based on a partial annotation scheme, which selects a part of the sentences to be annotated and asks human annotators to label a specific part of the target sentences. In the experiment, we show that the partial annotation scheme can quickly train the proposed point-wise prediction model compared to the existing active learning methods on NER tasks.