Combining joint models for biomedical event extraction

BMC Bioinformatics

Table 6 Incorporating Genia data with training data for ID track

Model	Recall	Precision	F₁
ID	38.0	59.3	46.3
ID (×1) + GE	40.2	52.0	45.3
ID (×2) + GE	41.7	52.4	46.4
ID (×3) + GE	45.0	54.8	49.4
ID (×4) + GE	43.8	55.2	48.9
ID (×5) + GE	44.7	55.1	49.4

Impact of merging several copies of ID training set with GE training and development sets for the Stanford model. F₁ scores are on ID development data (2N parser only, Core metric). Best performance is achieved when using three to five copies of ID to allow GE data to be used without overwhelming the ID data.

ISSN: 1471-2105