PARTICULAR ANNOTATION SCHEME FOR ACTIVE LEARNING ON RECOGNITION TASKS FOR NAMED ENTITIES

Abstract

One interesting method for reducing the high annotation costs associated with creating training data
for named entity recognition (NER) tasks is active learning. The efficiency of the data instance selection is
constrained, nonetheless, because current active learning techniques on NER tasks implicitly assume the whole
annotation scheme, where the entire phrase serves as the unit of an annotation request. In this research, we offer
a novel partial annotation scheme-based active learning approach that asks human annotators to identify a
specific portion of the target phrases after choosing a portion of the sentences to be annotated. In contrast to the
current active learning techniques on NER problems, we demonstrate in the experiment that the partial
annotation strategy can train the suggested point-wise prediction model rapidly