Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition

Table 4

Example of a feature vector in RNNLM experiments for word sam (Eng. am).

Index	Feature type	Feature	Value	Remarks
0	Constant	Constant	0.01	For math reasons
1–5	Special	Special word feat.	0	For bos/eos/unk/brk/silence
6	Unigram	Unigram prob.	0.00788	Scaled unigram log-prob.
7	Length	Word length	0.00186	Scaled word length
8–36	Word	1-hot vect. elem.	0	—
37	Word	1-hot vect. elem.	0.21	Scale based on unigram prob.
38–97644	Word	1-hot vect. elem.	0	—
97645–97657	Final	Lett. n-gram prob.	0	—
97658	Final	3-gram -am$ pr.	0.12	Scaled letter 3-gram prob.
97659–97758	Final	Lett. n-gram prob.	0	—
97759	Final	2-gram -m$ pr.	0.047	Scaled letter 2-gram prob.
97760–97869	Final	Lett. n-gram prob.	0	—
97870–98050	Initial	Lett. n-gram prob.	0	—
98051	Initial	2-gram ^s- pr.	0.03	Scaled letter 2-gram prob.
98052	Initial	3-gram ^sa- pr.	0.069	Scaled letter 3-gram prob.
98053–98144	Initial	Lett. n-gram prob.	0	—
98145–98300	Match	Lett. n-gram prob.	0	—
98301	Match	2-gram -am- pr.	0.064	Scaled letter 2-gram prob.
98302–100451	Match	Lett. n-gram prob.	0	—
100452	Match	2-gram -sa- pr.	0.057	Scaled letter 2-gram prob.
100453–100459	Match	Lett. n-gram prob.	0	—
100460	Match	3-gram -sam- pr.	0.11	Scaled letter 3-gram prob.
100461–101306	Match	Lett. n-gram prob.	0	—