Gene mention normalization and interaction extraction with context models and sentence motifs

Table 4 Performance of PPI extraction on the Spies corpus

Corpus	Short description of the submitted run	Precision	Recall	F measure
Spies	Initial pattern set	85.8	15.2	25.8
Spies	CP, single layer (POS tag including entity)	76.6	47.1	58.3
Spies	CP, multilayer (token, POS tag, stem, entity)	78.7	51.9	62.6
Spies	CP, optimized for precision	+1	-4	60.1
Spies	CP, optimized for recall	-5	+5	63.9

Performance of our approach to protein-protein interaction (PPI) extraction on other external corpora. We also show the influence of using part-of-speech (POS) tags only compared with multilayer alignments, and results for optimization towards a single metric (precision or recall). Note that these evaluations do not require the identification of proteins, as in BioCreative II, so figures are higher in general. CP, consensus patterns resulting from clustering and multiple sentence alignment.

ISSN: 1474-760X