Table 3 Regression performance of the baseline MRMR and ml-MRMR methods for the Turkish task when the minimum occurrence threshold (Nmin) is sets 3 and 5. In the underlined bold case, improvement is significant compared to the baseline system. The result with Gemaps feature set was 11.48
Dim | Baseline | Nmin(3) | Nmin(5) | Nmin(3) | Nmin(5) |
|---|---|---|---|---|---|
(Tr+Ger) | (Tr+Ger) | (Tr+Eng) | (Tr+Eng) | ||
2 | 13.30 | 10.84 | 11.98 | 10.79 | 10.88 |
3 | 12.30 | 10.51 | 11.26 | 10.40 | 11.94 |
4 | 12.45 | 10.85 | 10.74 | 9.85 | 12.68 |
5 | 12.56 | 10.58 | 11.23 | 9.36 | 13.65 |
10 | 12.45 | 10.82 | 12.13 | 11.87 | 13.93 |
15 | 12.08 | 11.12 | 12.00 | 11.99 | 13.30 |
20 | 12.87 | 11.91 | 11.46 | 10.92 | 12.00 |
40 | 13.28 | 12.67 | 11.98 | 10.93 | 10.31 |
80 | 11.58 | 12.28 | 13.06 | 10.80 | 10.50 |
100 | 11.75 | 11.95 | 13.08 | 10.88 | 10.23 |
200 | 11.32 | 11.55 | 12.14 | 11.05 | 11.06 |
400 | 11.42 | 11.72 | 12.00 | 10.99 | 11.23 |
800 | 11.31 | 11.39 | 11.35 | 11.10 | 11.08 |