A Comparison and Evaluation of Named Entity Recognition Methods in Thai Traditional Medicine Texts

The knowledge extraction and transliteration processes for the Thai traditional medicine documents requires experts to extract information from the textbook and spend a plenty of time. To reduce the complexity of these processes, the named entity recognition (NER) from the machine learning principle is applied to extract the key information from the document. In this work, three NER models including, CRFSuite, Bi-LSTM-CRF, and pre-trained model (WangchanBERTa) are considered the accuracy of each model of NER for the Thai traditional documents. After the testing process, the experimental results show that CRFSuite model is the most computationally efficient model and deliver high precision and recall. Moreover, CRFSuite model has the highest scores of each entity. While the accuracy of the WangchanBERTa model is the highest and faster convergence in the training process. Although the results from the Bi-LSTM-CRF model provide moderate accuracy, the Bi-LSTM-CRF can detect the common entities higher than other models.