Table 1.
Quantitative evaluation via fivefold cross-validation
Source | Model | gNID | NID | ||
LI | IB | 0.18 (±0.07) | 0.18 (±0.10) | 0.31 (±0.07) | 1.03 (±0.01) |
RKK+ | 0.70 (±0.23) | 0.47 (±0.10) | 0.32 (±0.10) | ||
U | IB | 0.24 (±0.09) | 0.39 (±0.12) | 0.56 (±0.07) | 1.06 (±0.01) |
RKK+ | 0.95 (±0.22) | 0.65 (±0.08) | 0.50 (±0.10) |
Shown are averages over left-out languages ± 1 SD for the LI and uniform (U) source distributions. Lower values of ε∕ , gNID, and NID are better. Best scores are in boldface.