Chinese NER Using ALBERT and Multi-word Information

Abstract

Recently, many Chinese Named Entity Recognition (NER) problems which utilize the character-based approach have a good performance, the character-based approach becomes the dominant current of the approaches, but there are two issues in the existing models. Firstly, the traditional character vector representation is too single to express the polysemia of characters. Secondly, words contain more information than characters, but the existing models cannot use the word information sufficiently. To address the above issues, the AM-BiLSTM model is provided in this work. With the introduction of ALBERT pre-training language model and Multi-word Information(MWI), the enhanced character embedding is composed. We perform experiments on two datasets and the results demonstrate that the capability of our method is improved outstandingly in comparison with the existing NER models.