Table 4 Performances of LSTM and CRF-based models for the three tasks (F1-score %)
From: Entity recognition from clinical texts via recurrent neural network
Model | 2010 i2b2 challenge (Concept Extraction) | 2012 i2b2 challenge (Event Detection) | 2014 i2b2 challenge (De-Identification) | |||
|---|---|---|---|---|---|---|
Exact | Inexact | Span | Type | Exact | Token | |
CRF | 83.69 | 91.39 | 90.82 | 83.72 | 92.58 | 95.37 |
LSTM-BASELINE | 85.36 | 92.58 | 92.20 | 87.74 | 93.30 | 96.05 |
LSTM + char-LSTM | 85.81 | 92.91 | 92.29 | 86.94 | 94.29 | 96.54 |
LSTM + char-CNN | 85.65 | 92.77 | 92.25 | 87.66 | 94.37 | 96.67 |
LSTM + char-LSTM + CNN | 85.78 | 92.76 | 92.28 | 87.80 | 94.16 | 96.44 |