Review Article
Extracting Parallel Sentences from Nonparallel Corpora Using Parallel Hierarchical Attention Network
Table 3
The precision (P), recall (R), and F1 scores of extracting parallel sentences.
| | Data | En-Fr | En-De | En-Zh | | SMT | NMT | SMT | NMT | SMT | NMT |
| | Baseline | 23.71 | 22.32 | 21.62 | 21.35 | 21.1 | 17.32 | | Top20K | 24.84 (+1.13) | 25.42 (+3.1) | 23.38 (+1.76) | 25.06 (+3.71) | 23.21 (+2.11) | 24.56 (+7.24) | | Top50K | 26.16 (+2.45) | 26.35 (+4.8) | 24.63 (+3.01) | 26.42 (+5.07) | 24.66 (+3.56) | 25.89 (+8.57) | | Top100K | 28.31 (+3.6) | 27.48 (+5.03) | 25.72 (+4.1) | 27.67 (+6.32) | 25.78 (+4.68) | 27.02 (+9.7) | | Top200K | 29.37 (+4.66) | 29.51 (+6.06) | 26.76 (+5.14) | 28.73 (+7.38) | 26.86 (+5.76) | 28.13 (+10.81) | | Top300K | 30.39 (+5.68) | 30.55 (+8.10) | 27.79 (+6.17) | 29.80 (+8.45) | 27.91 (+6.81) | 29.18 (+11.86) | | Top400K | 30.41 (+6.70) | 30.57 (+9.12) | 28.83 (+7.21) | 30.82 (+9.47) | 28.92 (+7.82) | 30.21 (+12.89) | | Top500K | 31.56 (+7.85) | 31.58 (+10.13) | 30.14 (+8.52) | 31.85 (+10.50) | 29.93 (+8.83) | 31.22 (+13.9) |
|
|
