Using BERT to Process Chinese Ellipsis and Coreference in Clinic Dialogues

This paper focuses on ellipsis and coreference resolution in Chinese dialogue. The experimental data are real medical diagnosis dialogues. New features for machine learning, as well as deep learning approach such as BERT, were used to develop classifiers to detect, classify, and resolve ellipsis and coreference. The experimental results show that rule-based systems, BERT, and neural networks outperform one another in different tasks. The best F-scores of ellipsis and coreference detection were 70.61% and 89.09%, respectively. The best accuracy of ellipsis and coreference classification were 77.96% and 83.09%, respectively. The best accuracy of ellipsis and coreference detection were 80.06% and 82.69%, respectively.