Skip to main content

Table 3 Comparison of \(F_1\) results for previously proposed parser variant (TurkuNLP-ST) and newly proposed approach initialized with various BERT models

From: Dependency parsing of biomedical text with BERT

Metric Method
TurkuNLP-ST mBERT BERT-large BioBERT BlueBERT SciBERT
UPOS 98.54 98.75 98.76 98.79 97.70 98.72
UFeats 98.63 98.76 98.77 98.79 97.54 98.75
Lemmas 99.44 99.44 99.45 99.46 99.21 99.47
UAS 91.54 92.66 93.16 93.45 91.34 92.93
LAS 90.28 91.38 91.97 92.31 89.42 91.67
CLAS 87.96 88.80 89.69 90.04 86.10 89.20
MLAS 85.93 86.88 87.77 88.22 82.76 87.22
BLEX 87.31 88.08 88.98 89.36 85.17 88.51
  1. All experiments performed using gold segmentation
  2. Maximum score in bold