MIT Press

Table 1:

Our semi-supervised approach improves performance over the baselines (10-fold cross-validation averaged over five randomly seeded runs). “Self-Training” and “Lexical Decoding” refer to experiments where we use these methods independently. “Both” refers to their combination. We highlight the best model for each language.

Model	% Character Error Rate				% Word Error Rate
Model	ain	grk	ybh	kwk	ain	grk	ybh	kwk
First-Pass	1.34	3.27	8.90	7.90	6.27	15.63	31.64	38.22
Base	0.80	1.70	8.44	4.97	5.19	7.51	21.33	27.65

Semi-Supervised
Self-Training	0.82	1.45	7.20	4.00	5.31	6.47	18.09	23.98
Lexical Decoding	0.81	1.51	7.56	4.28	5.18	6.60	19.13	25.09
Both	0.63	1.37	5.98	3.82	4.43	6.36	16.65	22.61

Error Reduction $(\frac{Base - Both}{Base})$	21%	19%	29%	23%	15%	15%	22%	18%

Model	% Character Error Rate				% Word Error Rate
Model	ain	grk	ybh	kwk	ain	grk	ybh	kwk
First-Pass	1.34	3.27	8.90	7.90	6.27	15.63	31.64	38.22
Base	0.80	1.70	8.44	4.97	5.19	7.51	21.33	27.65

Semi-Supervised
Self-Training	0.82	1.45	7.20	4.00	5.31	6.47	18.09	23.98
Lexical Decoding	0.81	1.51	7.56	4.28	5.18	6.60	19.13	25.09
Both	0.63	1.37	5.98	3.82	4.43	6.36	16.65	22.61

Error Reduction $(\frac{Base - Both}{Base})$	21%	19%	29%	23%	15%	15%	22%	18%

View Large

This Feature Is Available To Subscribers Only