Adversarial training based lattice LSTM for Chinese clinical named entity recognition

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Biomedical Informatics

Background:

• Clinical named entity recognition (CNER) is crucial for extracting information from electronic health records (EHR).
• Existing deep learning models for Chinese CNER often fail to fully utilize EHR information and lack robustness against perturbations.
• Word-based or character-based models have limitations in capturing the nuances of clinical text.

Purpose of the Study:

• To propose a novel adversarial training based lattice LSTM with a conditional random field layer (AT-lattice LSTM-CRF) for Chinese CNER.
• To enhance the utilization of information within EHR data for improved CNER.
• To increase the robustness of deep learning models in clinical text mining.

Main Methods:

• Developed an AT-lattice LSTM-CRF model integrating Lattice LSTM for richer EHR information capture.
• Employed adversarial training (AT) as a regularization technique to enhance model robustness by introducing data perturbations.
• Evaluated the model on the CCKS-2017 Task 2 dataset for Chinese CNER.

Main Results:

• The proposed AT-lattice LSTM-CRF model achieved a competitive F1 score of 89.64% on the CCKS-2017 Task 2 dataset.
• Demonstrated superior performance compared to other prevalent neural models for Chinese CNER.
• Showcased improved robustness of the neural model through adversarial training.

Conclusions:

• The AT-lattice LSTM-CRF model represents a significant advancement in Chinese CNER.
• The model effectively leverages EHR information and demonstrates enhanced robustness.
• This work provides a reinforced baseline for future research in clinical text mining and CNER.