Research Article
Named Entity Recognition for Public Interest Litigation Based on a Deep Contextualized Pretraining Approach
Table 4
Comparison of results between each of the pretrained models (NER category).
| | Methods | P | R | F1 |
| | BERT-FC [36] | 75.5 | 71.7 | 73.5 | | BERT-BiLSTM-CRF [36] | 94.0 | 94.0 | 94.0 | | BERT-WWM-FC | 77.8 | 72.6 | 75.1 | | BERT-WWM-BiLSTM-CRF | 95.2 | 95.2 | 95.2 | | BERT-WWM-EXT-FC | 75.3 | 75.9 | 75.6 | | BERT-WWM-EXT-BiLSTM-CRF (proposed method) | 96.0 | 96.0 | 96.0 | | RoBERTa-FC | 74.3 | 68.3 | 71.2 | | RoBERTa-BiLSTM-CRF | 95.2 | 94.8 | 95.0 | | RoBERTa-WWM-FC | 74.3 | 69.3 | 71.7 | | RoBERTa-WWM-BiLSTM-CRF | 95.4 | 95.0 | 95.2 | | RoBERTa-WWM-EXT-FC | 75.8 | 70.0 | 72.8 | | RoBERTa-WWM-EXT-BiLSTM-CRF | 95.5 | 94.7 | 95.1 | | ALBERT-FC | 73.8 | 64.0 | 68.6 | | ALBERT-BiLSTM-CRF | 90.0 | 90.0 | 90.0 |
|
|
The bold values mean the optimal values among all the methods.
|
