Natural Language Processing Algorithms for Normalizing Expressions of Synonymous Symptoms in Traditional Chinese Medicine

Table 2

Model parameters and development set results on HFDS.

Model	LR	DR	MC	Accuracy	Precision	Recall	F1-score
Encoder (Char)-Decoder (Char)	0.0003	0.5	512	0.8631 ± 0.0042	0.8637 ± 0.0091	0.8587 ± 0.0038	0.8611 ± 0.0053
Encoder (Char)-Decoder (Label)	0.0005	0.5	256	0.8688 ± 0.0046	0.8812 ± 0.0070	0.8623 ± 0.0044	0.8716 ± 0.0048
Encoder (Word)-Decoder (Label)	0.0005	0.3	512	0.8631 ± 0.0042	0.8637 ± 0.0091	0.8587 ± 0.0038	0.8611 ± 0.0053
Encoder (Word)-Decoder (Word)	0.0005	0.3	512	0.8549 ± 0.0055	0.8596 ± 0.0047	0.8468 ± 0.0065	0.8531 ± 0.0052
Encoder (Char)-Classification	0.005	0.3	512	0.8377 ± 0.0060	0.9020 ± 0.0109	0.8414 ± 0.0062	0.8706 ± 0.0058
Encoder (Word)-Classification	0.005	0.5	512	0.8326 ± 0.0061	0.8978 ± 0.0068	0.8335 ± 0.0056	0.8645 ± 0.0043
BERT-UniLM (Char)	0.00003	0.1	N/A	0.8966 ± 0.0027	0.9013 ± 0.0064	0.8920 ± 0.0041	0.8966 ± 0.0025
BERT-UniLM (Label)	0.00003	0.1	N/A	0.8957 ± 0.0042	0.8996 ± 0.0063	0.8895 ± 0.0038	0.8945 ± 0.0039
BERT-Classification	0.00003	0.1	N/A	0.9087 ± 0.0029	0.9216 ± 0.0027	0.9084 ± 0.0034	0.9150 ± 0.0018

Note. LR: learning rate; DR: dropout rate; MC: number of memory cells of RNN; N/A: not applicable.