神經(jīng)機器翻譯的聯(lián)合訓練（英文版）

Contents
1 Neural Machine Translation 1
1.1 Introduction 1
1.2 Neural Machine Translation 4
References 8
2 Agreement-Based Joint Training for Bidirectional Attention-Based
Neural Machine Translation 11
2.1 Introduction 11
2.2 Agreement-Based Joint Training 12
2.3 Experiments 16
2.3.1 Setup 16
2.3.2 Comparison of Loss Functions 17
2.3.3 Results on Chinese-English Translation 18
2.3.4 Results on Chinese-English Alignment 18
2.3.5 Analysis of Alignment Matrices 19
2.3.6 Results on English-to-French Translation 21
2.4 Summary 22
References 22
3 Semi-supervised Learning for Neural Machine Translation 25
3.1 Introduction 25
3.2 Semi-supervised Learning for Neural Machine Translation 27
3.2.1 Supervised Learning 27
3.2.2 Autoencoders on Monolingual Corpora 27
3.2.3 Semi-supervised Learning 29
3.2.4 Training 30
3.3 Experiments 31
3.3.1 Setup 31
3.3.2 Effect of Sample Size k 32
3.3.3 Effect of OOV Ratio 34
3.3.4 Comparison with SMT . . . . . . . . . . . . . . . . . . . . . . . . . . .35
3.3.5 Comparison with Previous Work . . . . . . . . . . . . . . . . . . .36
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .394 Joint Training for Pivot-Based Neural Machine Translation . . . . . . 41
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414.2 Pivot-Based NMT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.3 Joint Training for Pivot-Based NMT . . . . . . . . . . . . . . . . . . . . . . 45
4.3.1 Training Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.2 Connection Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.3 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .46
.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .48
4.4.2 Results on the Europarl Corpus . . . . . . . . . . . . . . . . . . . . 49
4.4.3 Results on the WMT Corpus . . . . . . . . . . . . . . . . . . . . . . 50
4.4.4 Effect of Bridging Corpora . . . . . . . . . . . . . . . . . . . . . . . . 52
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .53
5 Joint Modeling for Bidirectional Neural Machine Translation with Contrastive Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Unidirectional Neural Machine Translation. . . . . . . . . . . . . . . . . . 57
5.3 Bidirectional Neural Machine Translation. . . . . . . . . . . . . . . . . . .57
5.4 Decoding Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.5.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .61
5.5.2 Effect of Translation Strategies . . . . . . . . . . . . . . . . . . . . .62
5.5.3 Comparison with SMT and Standard NMT . . . . . . . . . . . . 63
5.5.4 BLEU Scores Over Sentence Length . . . . . . . . . . . . . . . . 64
5.5.5 Comparison of Learning Curves . . . . . . . . . . . . . . . . . . . . 65
5.5.6 Analysis of Expected Embeddings . . . . . . . . . . . . . . . . . . 66
5.5.7 Results on English-German Translation . . . . . . . . . . . . . . .66
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .67
6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.1 Attentional Mechanisms in Neural Machine Translation . . . . . . . . 69
6.2 Capturing Bidirectional Dependencies . . . . . . . . . . . . . . . . . . . . .70
6.2.1 Capturing Bidirectional Dependencies . . . . . . . . . . . . . . . . 70
6.2.2 Agreement-Based Learning . . . . . . . . . . . . . . . . . . . . . . .70
6.3 Incorporating Additional Data Resources 71
6.3.1 Exploiting Monolingual Corpora for Machine
Translation 71
6.3.2 Autoencoders in Unsupervised and Semi-supervised
Learning 71
6.3.3 Machine Translation with Pivot Languages 72
6.4 Contrastive Learning 72
References 72
7 Conclusion 75
7.1 Conclusion 75
7.2 Future Directions 76
7.2.1 Joint Modeling 76
7.2.2 Joint Training 77
7.2.3 More Tasks 78
References 78