NER - VLSP 2021: A Span-Based Model for Named Entity Recognition Task with Co-teaching+ Training Strategy

Abstract

Named entities containing other named entities inside are referred to as nested entities, which commonly exist in news articles and other documents. However, most studies in the field of Vietnamese named entity recognition entirely ignore nested entities. In this report, we describe our system at VLSP 2021 evaluation campaign, adopting the technique from dependency parsing to tackle the problem of nested entities. We also apply Coteaching+ technique to enhance the overall performance and propose an ensemble algorithm to combine predictions. Experimental results show that the ensemble method achieves the best F1 score on the test set at VLSP 2021.