Normalized log-likelihood of the EM algorithm for a different number of training epochs and different implementations of the update rule for the translation model, including batch, incremental, and interlaced versions. Results executed on the XRCE and Europarl English–Spanish training corpora are shown.
| . | . | Norm. Log-likelihood . | |
|---|---|---|---|
| . | #Epochs . | XRCE . | Europarl . |
| Batch EM | 5 | -41.8 | -90.9 |
| Incr. EM | 1 | -41.8 | -91.6 |
| Incr. EM (Interlaced, R=10) | 5 | -40.3 | -87.1 |
| Incr. EM (Interlaced, R=100) | 5 | -40.0 | -85.6 |
| Incr. EM (Interlaced, R=1000) | 5 | -39.6 | -84.5 |
| . | . | Norm. Log-likelihood . | |
|---|---|---|---|
| . | #Epochs . | XRCE . | Europarl . |
| Batch EM | 5 | -41.8 | -90.9 |
| Incr. EM | 1 | -41.8 | -91.6 |
| Incr. EM (Interlaced, R=10) | 5 | -40.3 | -87.1 |
| Incr. EM (Interlaced, R=100) | 5 | -40.0 | -85.6 |
| Incr. EM (Interlaced, R=1000) | 5 | -39.6 | -84.5 |