Sieve-based relation extraction of gene regulatory networks from biological literature

BMC Bioinformatics

Table 8 Relations extracted by different subsets of feature functions on a development dataset.

Subset of feature functions	S	D	Dev I	M	SER
A	0	67	1	0	1.01
A - B	1	64	2	2	1.00
A - C	1	64	2	2	1.00
A - D	0	52	7	15	0.88
A - E	0	41	12	26	0.79
A - F	1	38	12	28	0.76
A - G	0	37	12	30	0.73
A - H	0	37	12	30	0.73

The table shows the number of substitutions (S), deletions (D), insertions (I), matches (M) and slot error rate (SER) metric. The results are measured on the development dataset using CRF-based sieves only. Best results per metric are highlighted in bold. The feature function subsets are selected as follows: (A) target label distribution, starts upper, starts upper twice, Hearst co-occurence, mention token distance, (B) parse tree mention depth, parse tree parent value, parse tree path, (C) BSubtilis, IsBSubtilis, IsBSubtilisPair, (D) prefix value, suffix value, (E) consequent value, current value, (F) context value, (G) previous/next value combination, left/right/between value and (H) split to values. For their detailed descriptions see Table 1 and Table 2.

ISSN: 1471-2105