How to make the most of NE dictionaries in statistical NER

BMC Bioinformatics

Table 1 Protein name recognition performance

Tagging		R	P	F
(a) POS/PROTEIN tagging	Full	52.91	43.85	47.96
	Left	61.48	50.95	55.72
	Right	61.38	50.87	55.63
Sequential Labelling		R	P	F
(b) Word feature	Full	63.23	70.39	66.62
	Left	68.15	75.86	71.80
	Right	69.88	77.79	73.63
(c) (b) + orthographic feature	Full	77.17	67.52	72.02
	Left	82.51	72.20	77.01
	Right	84.29	73.75	78.67
(d) (c) + POS feature	Full	76.46	68.41	72.21
	Left	81.94	73.32	77.39
	Right	83.54	74.75	78.90
(e) (d) + PROTEIN feature	Full	77.58	69.18	73.14
	Left	82.69	73.74	77.96
	Right	84.37	75.24	79.54
(f) (e) after adding protein names in the training set to the lexicon	Full	79.85	68.58	73.78
	Left	84.82	72.85	78.38
	Right	86.60	74.37	80.02

Protein name recognition performance of the proposed method, evaluated by recall (R), precision (P), and F-measure (F). The left boundary (Left), the right boundary (Right), and both boundary (Full) recognition performance were measured. (a) the performance of POS/PROTEIN tagging. (b) the performance of sequential labelling when using the word feature only. (c) the performance of sequential labelling when using the word and orthographic features. (d) the performance of sequential labelling when using the word, orthographic, and POS features. (e) the performance of sequential labelling when using the word, orthographic, POS, and PROTEIN name features. (f) the performance of sequential labelling with the features used in (e) after adding protein names appearing in the training set to the lexicon. NB: no retraining was conducted.

ISSN: 1471-2105