The source code repository for the paper "Towards Precise Lexicon Integration in Neural Machine Translation"
Providing Mullti-choice Lexical Constraints
- Instead of creating:
{ 'text': 'This is a test .', 'constraints': ['constr@@ aint',
'multi@@ word constr@@ aint',
] }- Create:
{ 'text': 'This is a test .', 'constraints': [ ['constr@@ aint', 'constr@@ ain@@ ts', 'Cons@@ tr@@ aint'],
['multi@@ word constr@@ aint', 'Multi@@ word constr@@ aint'],
] }- And let your trained NMT model to decide for the best fitting constraint options
| Model | Term. Rate | Term. Prec. | Term. Recall | Term. F1 | BLEU (Δ) |
|---|---|---|---|---|---|
| baseline (no constraints) | 57.43 | 78.20 | 81.16 | 79.65 | 33.2 |
| Dinu et al., 2019 (with all lemmata matching factors) | 81.22 | 62.76 | 95.54 | 75.76 | 30.2 (-3.0) |
| Dinu et al., 2019 (with BERT selected factors) | 57.17 | 79.13 | 81.45 | 80.27 | 31.8 (-1.4) |
| Post and Vilar, 2018 (with all lemmata matching constraints) | 99.88 | 49.04 | 99.23 | 65.64 | 26.0 (-7.2) |
| Multi-Choice Lexical Constraints (with alll lemmata matching constraints) | 99.68 | 50.82 | 99.54 | 67.29 | 28.2 (-5.0) |
| Post and Vilar, 2018 (with BERT selected constraints) | 61.67 | 75.02 | 87.30 | 80.69 | 31.1 (-2.1) |
| Multi-Choice Lexical Constraints (with random subset of lemmata matching constraints) | 70.71 | 66.92 | 86.55 | 75.48 | 31.7 (-1.5) |
| Multi-Choice Lexical Constraints* (with BERT selected constraints) | 61.62 | 77.35 | 87.30 | 82.03 | 32.5 (-0.7) |
| Model | Term. Rate | Term. Prec. | Term. Recall | Term. F1 | BLEU (Δ) |
|---|---|---|---|---|---|
| baseline (no constraints) | 57.33 | 77.19 | 75.01 | 76.08 | 28.8 |
| Dinu et al., 2019 (with all lemmata matching factors) | 81.42 | 64.72 | 92.72 | 76.23 | 26.4 (-2.4) |
| Dinu et al., 2019 (with BERT selected factors) | 58.27 | 79.09 | 77.88 | 78.48 | 27.8 (-1.0) |
| Post and Vilar, 2018 (with all lemmata matching constraints) | 99.79 | 51.13 | 99.32 | 67.51 | 24.6 (-4.2) |
| Multi-Choice Lexical Constraints (with alll lemmata matching constraints) | 99.51 | 52.46 | 99.15 | 68.62 | 24.9 (-3.9) |
| Post and Vilar, 2018 (with BERT selected constraints) | 63.90 | 74.35 | 84.73 | 79.20 | 27.4 (-1.4) |
| Multi-Choice Lexical Constraints (with random subset of lemmata matching constraints) | 72.31 | 65.17 | 82.54 | 72.83 | 27.3 (-1.5) |
| Multi-Choice Lexical Constraints* (with BERT selected constraints) | 63.84 | 75.84 | 84.52 | 79.94 | 28.1 (-0.7) |
- Update sockeye.lexical_constraints script to produce MLC compatible constraints.