Table 3.
Official evaluation of NER system on the test data
| Entity | Combined entity model | Independent entity model | CNN-BiLSTM-CRF model |
|---|---|---|---|
| Cell | 0.783 / 0.708 / 0.743 | 0.767 / 0.749 / 0.758 | 0.769 / 0.641 / 0.699 |
| Cellular | 0.673 / 0.508 / 0.579 | 0.630 / 0.571 / 0.599 | 0.634 / 0.495 / 0.556 |
| Protein | 0.729 / 0.739 / 0.734 | 0.728 / 0.745 / 0.736 | 0.764 / 0.768 / 0.766 |
| Organism | 0.860 / 0.809 / 0.834 | 0.823 / 0.852 / 0.837 | 0.789 / 0.771 / 0.780 |
| Molecule | 0.775 / 0.587 / 0.668 | 0.661 / 0.681 / 0.671 | 0.667 / 0.595 / 0.629 |
| Tissue | 0.727 / 0.575 / 0.642 | 0.650 / 0.700 / 0.674 | 0.646 / 0.622 / 0.634 |
| All | 0.747 / 0.694 / 0.720 | 0.719 / 0.730 / 0.725 | 0.739 / 0.702 / 0.720 |
Evaluation of the original model submitted to the shared task (combined), improved CRF model (independent) and the neural character level model (CNN-BiLSTM-CRF) based on the official evaluation script with strict entity span matching on the test set. Numbers within cells are precision/recall/F-score.