Dual Reconstruction: a Unifying Objective for Semi-Supervised Neural Machine Translation

Weijia Xu, Xing Niu, Marine Carpuat

Abstract

While Iterative Back-Translation and Dual Learning effectively incorporate monolingual training data in neural machine translation, they use different objectives and heuristic gradient approximation strategies, and have not been extensively compared. We introduce a novel dual reconstruction objective that provides a unified view of Iterative Back-Translation and Dual Learning. It motivates a theoretical analysis and controlled empirical study on German-English and Turkish-English tasks, which both suggest that Iterative Back-Translation is more effective than Dual Learning despite its relative simplicity.

Introduction

Neural Machine Translation (NMT) has achieved remarkable success, primarily driven by large-scale parallel corpora. However, obtaining high-quality sentence pairs is expensive and often unfeasible for low-resource language pairs or specific domains. To address this, researchers leverage monolingual data through two prominent semi-supervised techniques: Iterative Back-Translation (IBT) and Dual Learning.

IBT uses a model to generate synthetic parallel data from monolingual text, which is then used for further training. Dual Learning, conversely, employs a game-theoretic framework where two models collaborate to reconstruct sentences, often utilizing reinforcement learning and auxiliary language models. Despite their popularity, the theoretical relationship between these methods and their relative performance has remained under-explored.

The Dual Reconstruction Objective

This research introduces the Dual Reconstruction Objective, a unified framework that encompasses both IBT and Dual Learning. This objective views semi-supervised training as a bidirectional reconstruction process:

  1. Target-Source-Target: A target sentence is translated into a "pseudo-source" and then reconstructed back to the original target.
  2. Source-Target-Source: A source sentence is translated into a "pseudo-target" and then reconstructed back to the original source.
The Dual Reconstruction framework illustrating the inference and reconstruction paths for both target-source-target and source-target-source cycles.

The authors demonstrate that maximizing this objective, while maintaining high mutual information between source and target, leads to the same global optimum as the ideal marginal likelihood objective. Theoretically, IBT and Dual Learning are shown to be different approximations of this unified objective. IBT simplifies the process by "blocking" certain gradients (treating synthetic data as ground truth), while Dual Learning attempts to estimate them using noisy policy gradients.

Experimental Evaluation

The framework was evaluated across three scenarios using WMT18 and IWSLT17 benchmarks:

  • High-Resource: German-English (de-en) with 4.5M parallel sentences.
  • Low-Resource: Turkish-English (tr-en) with 0.2M parallel sentences.
  • Cross-Domain: Training on News data and testing on TED talks (de-en).

Key Results and Insights

The empirical study yielded several critical findings:

IBT Superiority and Efficiency

In nearly all test cases, IBT outperformed Dual Learning in terms of BLEU scores. Notably, batch-level IBT was 30–50% faster than Dual Learning, as it avoids the computational overhead of policy gradient estimation required by the latter.

BLEU score comparison on the Turkish-English low-resource task, showing that IBT variants generally outperform Dual Learning.

Redundancy of Language Model Loss

Dual Learning often incorporates a separate Language Model (LM) loss to ensure fluency. However, the theoretical analysis proves that the reconstruction objective implicitly handles this. Adding an explicit LM loss does not shift the global optimum and is largely redundant.

Stability of Heuristic Gradients

While Dual Learning's policy gradients are theoretically more rigorous, they are often noisy and unstable in practice. The "heuristic" gradient blocking used in IBT provides a more stable optimization path for the dual reconstruction objective, leading to better final performance.

Conclusion

The dual reconstruction objective provides a powerful lens for understanding semi-supervised NMT. By unifying IBT and Dual Learning, this work clarifies why the simpler IBT method often yields superior results. Future work may extend this objective to fully unsupervised translation and other sequence-to-sequence tasks like summarization or style transfer.

Summary updated

Abstract

While Iterative Back-Translation and Dual Learning effectively incorporate monolingual training data in neural machine translation, they use different objectives and heuristic gradient approximation strategies, and have not been extensively compared. We introduce a novel dual reconstruction objective that provides a unified view of Iterative Back-Translation and Dual Learning. It motivates a theoretical analysis and controlled empirical study on German-English and Turkish-English tasks, which both suggest that Iterative Back-Translation is more effective than Dual Learning despite its relative simplicity.
BibTeX
@inproceedings{xu2020dual,
  author = {Weijia Xu and Xing Niu and Marine Carpuat},
  title = {{Dual Reconstruction: a Unifying Objective for Semi-Supervised Neural Machine Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=QhnXRUdhkEY},
}
Figure