Investigation of Deep Active Self-learning Algorithms Applied to Named Entity Recognition

• José Reinaldo Cunha Santos A. V. Silva Neto Osaka University
• Thiago de Paulo Faleiros UnB

Resumo

Active Self-Learning algorithms reduce the labeled data required to train a Machine Learning model through supervised training. This paper explores various Active Self-Learning algorithms for named entity recognition tasks. Firstly, we investigate the impact of different self-training techniques on Active Self-Learning algorithms. Secondly, we propose a novel token-level Active Self-Learning algorithm that achieves near-peak performance using fewer hand-annotated tokens compared to existing works. Through numerous experiments, we found that the sentence-level Active Self-Learning algorithm did not consistently yield significant results compared to pure active learning. However, our proposed token-level Active Self-Learning algorithm showed promising performance, training a neural model to nearly peak accuracy with fewer human-annotated tokens compared to state-of-the-art active learning baseline algorithms. The experimental results are presented and discussed, demonstrating the superior performance of the token-level Active Self-Learning algorithm