BLEU, WER, and KSMR for the EMEA test corpora using conventional (batch learning without retraining), online, and online from scratch SMT systems, with models estimated by means of the Europarl training corpora. Log-linear weights were trained via MERT. The average online learning time (LT) in seconds is shown for the online system.
| Corpus . | SMT system . | BLEU . | WER . | KSMR . | LT (sec) . |
|---|---|---|---|---|---|
| Eng-Spa | conventional | 24.0±0.8 | 64.0±0.9 | 49.1±0.6 | - |
| online | 45.6±1.2 | 49.6±1.4 | 32.3±0.8 | 0.1 | |
| online scratch | 36.0±1.2 | 52.2±1.0 | 35.6±0.8 | 0.1 | |
| Eng-Fre | conventional | 18.3±0.7 | 69.4±0.9 | 52.3±0.6 | - |
| online | 39.7±1.3 | 53.9±1.3 | 33.7±0.8 | 0.1 | |
| online scratch | 30.0±0.2 | 59.5±1.1 | 41.2±0.8 | 0.1 | |
| Eng-Ger | conventional | 17.7±0.8 | 77.8±1.3 | 55.3±0.6 | - |
| online | 35.7±1.3 | 61.8±1.5 | 37.2±0.8 | 0.1 | |
| online scratch | 27.4±1.2 | 65.5±1.2 | 43.8±0.9 | 0.1 |
| Corpus . | SMT system . | BLEU . | WER . | KSMR . | LT (sec) . |
|---|---|---|---|---|---|
| Eng-Spa | conventional | 24.0±0.8 | 64.0±0.9 | 49.1±0.6 | - |
| online | 45.6±1.2 | 49.6±1.4 | 32.3±0.8 | 0.1 | |
| online scratch | 36.0±1.2 | 52.2±1.0 | 35.6±0.8 | 0.1 | |
| Eng-Fre | conventional | 18.3±0.7 | 69.4±0.9 | 52.3±0.6 | - |
| online | 39.7±1.3 | 53.9±1.3 | 33.7±0.8 | 0.1 | |
| online scratch | 30.0±0.2 | 59.5±1.1 | 41.2±0.8 | 0.1 | |
| Eng-Ger | conventional | 17.7±0.8 | 77.8±1.3 | 55.3±0.6 | - |
| online | 35.7±1.3 | 61.8±1.5 | 37.2±0.8 | 0.1 | |
| online scratch | 27.4±1.2 | 65.5±1.2 | 43.8±0.9 | 0.1 |