Tagging schemes can do more in Named Entity Recognition: Take Chinese as an example

Abstract

Tagging schemes are essential for sequence labeling tasks like chunking, named entity recognition (NER), etc. However, they are often specified beforehand and receive much less attention in NER now. We conduct research on existing commonly used tagging schemes and form a unified framework to describe existing tagging schemes. Taking Chinese NER (CNER) as an example, we find that tagging schemes can do more in NER. Motivated by the observation that existing tagging schemes tag non-entity tokens with the same label ‘O’, we propose a BIO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup> ES tagging scheme to tag non-entity tokens surrounding entities with ‘O-’ and ‘O+’ labels, thus utilizing the boundary information of entities. We proposed a tagging scheme with stacked labels to recognize nested entities. Experimental results demonstrate that the proposed BIO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup> ES tagging scheme can improve the performance without modifying the model. And the stacked labels can help recognize nested entities. The results show that we can carry out more targeted research on the sequence labeling task transferred from NER by adjusting the tagging scheme reasonably.