Table 3.
Performance of the CRF-based NER systems on Chinese admission and discharge notes when different features were used
| Feature | Admission notes | Discharge summaries | ||
|---|---|---|---|---|
| Exact-match | Inexact-match | Exact-match | Inexact-match F-measure (R/P) | |
| BOC | 93.18 (93.70/92.66) | 94.32 (94.85/93.80) | 88.89 (89.80/87.99) | 90.75 (91.68/89.83) |
| BOC+BOW-STAN | 93.19 (93.59/92.79) | 94.40 (94.81/94.00) | 89.01 (89.87/88.16) | 90.95 (91.83/90.08) |
| BOC+BOW-STAN+POS | 93.14 (93.46/92.81) | 94.37 (94.70/94.04) | 88.89 (89.59/88.21) | 90.86 (91.57/90.16) |
| BOC+BOW-DICT | 93.30 (93.66/92.94) | 94.50 (94.87/94.13) | 89.19 (90.16/88.24) | 90.97 (91.96/90.00) |
| BOC+SECTION | 93.28 (93.63/92.93) | 94.40 (94.76/94.05) | 88.95 (89.96/87.96) | 90.71 (91.74/89.70) |
| BOC+BOW-STAN+SECTION | 93.22 (93.61/92.83) | 94.45 (94.85/94.06) | 89.02 (89.95/88.12) | 90.89 (91.83/89.96) |
| BOC+BOW-DICT+SECTION | 93.52 (93.77/93.26) | 94.69 (94.95/94.43) | 89.23 (90.29/88.20) | 91.00 (92.08/89.94) |
Values are F-measure (recall/precision) (%).
BOC, bag-of-characters; BOW-DICT, bag-of-words from dictionary lookup; BOW-STAN, bag-of-words from the Stanford Word Segmenter; CRF, conditional random fields; NER, named entity recognition; POS, part-of-speech information from Stanford Word Segmenter; SECTION, section information.