Flex CRFs v.s. CRF++

比較對照表

	CRF++	FlexCRFs
Configuration	* Freq = 1 * C(sigma^2) = 1.00000 * Number of features: UC (Unigram for current word)/U+BC (Unigram+Bigram for current word) = 7504/11280 * Iteration = 40/43	* f_rare_threshold = 1 * sigma_square = 1 * Number of features: 1851 * Iteration = 128
Result	* UC-recall / precision / f-score = 0.6656 / 0.6535 / 0.6595 * U+BC-recall / precision / f-score = 0.6705 / 0.6582 / 0.6643	* recall / precision / f-score = 0.6332 / 0.6362 / 0.6347

FlexCRFs V.S. CRF++

	CRF++	FlexCRFs
Feature numbers	30,986	No edge features	With edge features
		34,848	34,861
Training time	690.10 seconds	20,370.0 seconds
Num. of iterations	88	46
Performance (recall / precision / f-score)	0.6948 / 0.7002 / 0.6975	No edge features	With edge features
		0.5659 / 0.6033 / 0.5840	0.6642 / 0.6515 / 0.6578
Based on 86918 sentences C(sigma^2) = 1.0