Table 6.

Combinations of Ogmios (Og) with Caramba (Ca). 10-fold cross validation on the training corpus (except for Ogmios, first row), then application to the test corpus. Pairs of numbers (−n, +m) in the rest of this caption indicate the range of relative positions of n-grams of attributes. All feature sets in the CRF include bigrams of classes (B feature).

	Training						Test
	Conlleval			i2b2Evaluation			Conlleval			i2b2Evaluation
	P	R	F	P	R	F	P	R	F	P	R	F
Og	0.7079	0.8229	0.7611	0.8281	0.9602	0.8893	0.5681	0.6419	0.6027	0.7839	0.8852	0.8315
Ca	0.7761	0.6957	0.7337	0.9322	0.8336	0.8801	0.7541	0.6787	0.7144	0.9210	0.8282	0.8721
OgF	0.7581	0.8091	0.7828	0.8648	0.9206	0.8918	0.7469	0.6758	0.7096	0.9183	0.8303	0.8721
OgT	0.8483	0.8370	0.8426	0.9292	0.9144	0.9217	0.7443	0.6746	0.7077	0.9163	0.8299	0.8709
OgCa	0.8613	0.8477	0.8545	0.9362	0.9192	0.9276	0.7472	0.6795	0.7117	0.9159	0.8324	0.8721

Notes: Og: Ogmios alone, as is; Ca: Caramba alone; OgF: Ogmios output as only attributes: unigrams and bigrams of Ogmios attributes (−1, +1); OgT: Ogmios + normalized token: unigrams and bigrams of Ogmios attributes (−1, +1), with unigrams (−5, +3) and bigrams (−2, +1) of tokens, and one of the previous three tokens; OgCa: Ogmios as feature added to Caramba: unigrams and bigrams of Ogmios attributes (−1, +1), and above subset of Caramba features. Bold shows the (set of) best results per column; italics shows the lowest results when they are notable.