BERT-Based Named Entity Recognition in Chinese Twenty-Four Histories
Lists
Only you see this
Abstract
Named entity recognition in classical Chinese plays a fundamental role in improving the ability of information extraction and constructing knowledge graphs from classical Chinese. However, due to the lack of annotated data and the complexity of grammatical rules, named entity recognition in classical Chinese has made little progress. In order to solve the problem of lack of labeled data, we propose an end-to-end solution that is not based on domain knowledge, which instead is based on the pre-trained BERT-Chinese model and integrates the BiLSTM-CRF model for classical Chinese named entity recognition. The BERT-Chinese model converts the input text into a character-level embedding vector, then the BiLSTM model is used for future training, and finally CRF is able to normalize the output of BiLSTM to obtain a globally optimal labeling sequence. We conducted fine-tuning training on ChineseDailyNerCorpus. By designing a optimized fine-tuning method, we have realized the named entity recognition task in the Chinese Twenty-Four Histories. We evaluated our model on the Chinese twenty-four histories data, and finally achieved an average F1 value of about 75%. The experimental results also show that the BERT model has a strong ability in transfer learning.
BibTeX
@inproceedings{yu2020bert,
author = {Peng Yu and Xin Wang},
title = {{BERT-Based Named Entity Recognition in Chinese Twenty-Four Histories}},
booktitle = {WISA},
year = {2020},
doi = {10.1007/978-3-030-60029-7\_27},
}Loading related content...