Concept recognition for extracting protein interaction relations from biomedical text

Table 6 GN results: performance of nine heuristics used to filter false-positive gene mentions or modify gene mentions to improve dictionary matching performance.

	Presence of ...	Example	P	R	F	Modified
0			0.770	0.673	0.718	0
1	Gene chromosome location	3p11-3p12.1	0.772	0.673	0.719	34
2	Single, short lowercase word	heme	0.778	0.672	0.721	112
3	Strings of only numbers &/or punct	9+/-76	0.779	0.672	0.722	206
4	Extra preceding words	protein SNF to SNF	0.790	0.681	0.731	225^a
5	Extra trailing words	SNF protein to SNF	0.812	0.723	0.765	419^a
6	Amino acids	Ser-119	0.815	0.723	0.766	460
7	Protein families	Bcl-2 family proteins	0.816	0.722	0.766	701
8	Protein domains, motifs, fusion	SNH domain	0.828	0.722	0.771	883
9	Nonhuman keywords	rat IFN gamma	0.829	0.725	0.774	1,086^a

Results depicted here are from the development dataset. Step 0 indicates performance before application of any rules. At each step, the rules of preceding steps are also applied. Modified refers to the cumulative number of gene mentions removed or altered. ^aRules 4 and 5 result in modification of gene mentions only. Rule 9 can result in either modification or removal of gene mentions. All other rules result in removal of gene mentions. GN, gene normalization.

ISSN: 1474-760X