MIT Press

Skip to Main Content

Table 3:

A more informed unknown word model (character n-gram) in combination with the word-level known word model consistently performs better than the alternatives for all four languages in our dataset.

Known Word Model	Unknown Word Model	% Character Error Rate	% Word Error Rate
Known Word Model	Unknown Word Model	ain	grk	ybh	kwk	ain	grk	ybh	kwk
CharLM	(not needed)	0.64	1.43	6.22	3.85	4.50	6.44	16.78	22.90
WordLM	Character uniform	0.64	1.42	6.12	3.95	4.50	6.39	16.71	23.11
Ours	Character n-gram	0.63	1.37	5.98	3.82	4.43	6.36	16.65	22.61

Known Word Model	Unknown Word Model	% Character Error Rate	% Word Error Rate
Known Word Model	Unknown Word Model	ain	grk	ybh	kwk	ain	grk	ybh	kwk
CharLM	(not needed)	0.64	1.43	6.22	3.85	4.50	6.44	16.78	22.90
WordLM	Character uniform	0.64	1.42	6.12	3.95	4.50	6.39	16.71	23.11
Ours	Character n-gram	0.63	1.37	5.98	3.82	4.43	6.36	16.65	22.61