Adaptive Policy with Wait-k Model for Simultaneous Translation

Libo Zhao, Kai Fan, Wei Luo, Wu Jing, Shushu Wang, Ziqian Zeng, Zhongqiang Huang

Abstract

Simultaneous Machine Translation (SiMT) requires a delicate balance between translation quality and latency. Traditional fixed policies lack flexibility, while existing adaptive policies are often tightly coupled with the translation model, making them difficult to optimize and adapt. We present Divergence-based Adaptive Policy (DaP), a novel approach that decouples the read/write decision from the translation model. By training a lightweight policy network to predict the divergence between partial and full-context distributions, DaP-SiMT achieves a superior quality-latency trade-off compared to current state-of-the-art systems.

Introduction

Simultaneous Machine Translation (SiMT) is the task of translating a source sentence into a target language while the source is still being generated. The core challenge lies in the "Read/Write (R/W) policy": deciding at each step whether to wait for more source information (Read) or output a translated word (Write).

Historically, systems have relied on two types of policies:

  1. Fixed Policies (Wait-$k$): These systems wait for a fixed number of words ($k$) before starting. While simple, they fail to adapt to varying sentence complexities.
  2. Adaptive Policies: These learn to make dynamic decisions but are typically "tightly coupled" with the translation model. This joint training is computationally expensive and makes it difficult to leverage existing high-quality, pre-trained translation models.

DaP-SiMT addresses these issues by introducing a smart, adaptive "navigator" that operates on top of a frozen, pre-existing translation model.

The DaP-SiMT Framework

The proposed method, DaP-SiMT, consists of two primary components designed to work in tandem without requiring joint retraining of the translation weights.

Multi-path Wait-$k$ Model

Instead of a model optimized for a single $k$, the authors utilize a multi-path model trained to handle various wait lengths. Crucially, this model remains frozen during the policy training phase, preserving its translation capabilities while the policy learns to drive it.

Divergence-based Adaptive Policy (DaP)

The policy is supervised by the concept of statistical divergence. It essentially asks: "How much would my prediction for the next word change if I had the full sentence?" By calculating measures like KL-divergence or Cosine distance between the probability distribution of the next word given a partial sentence versus the full sentence, the system gains a signal for confidence. Low divergence suggests the model is ready to Write, while high divergence indicates a need to Read more.

The DaP-SiMT architecture features a lightweight R/W Policy Net trained on top of a frozen Multi-path Wait-k translation model.

To implement this in real-time, where the full sentence is unavailable, a Lightweight Policy Network is trained. This network, consisting of a single Transformer decoder layer and a regression head, learns to predict what that divergence would be based solely on the available context.

Experimental Results

The approach was validated on major benchmarks including WMT22 Chinese $\rightarrow$ English, WMT15 German $\rightarrow$ English, and IWSLT15 English $\rightarrow$ Vietnamese.

Performance Trade-offs

DaP-SiMT consistently outperformed both fixed wait-$k$ strategies and the previous state-of-the-art adaptive policy, ITST. As shown in the results for German $\rightarrow$ English translation, DaP-SiMT (represented by the red line) achieves higher BLEU scores at comparable or lower Average Lagging (AL) levels.

Performance comparison on the De-En task showing DaP-SiMT achieving superior BLEU scores across various latency (AL) levels compared to Mp Wait-k and ITST.

Key Insights

  • Effective Decoupling: The results prove that joint training is not a prerequisite for high performance. A well-trained multi-path model contains sufficient internal information for an external policy to navigate effectively.
  • Efficiency: The lightweight nature of the policy network (adding only one decoder layer) makes it highly memory and computationally efficient compared to complex jointly-trained models.
  • Robustness: Ablation studies indicated that the model is not overly sensitive to the specific type of divergence measure used, simplifying implementation.

Conclusion and Future Work

DaP-SiMT demonstrates that decoupling the R/W policy from the translation model is a viable and efficient path for SiMT. While the system currently relies on a threshold ($\lambda$) to balance speed and quality, which can be non-linear to tune, it represents a significant step toward more modular and adaptable simultaneous translation systems. Future work may explore more principled ways to handle "exposure bias" and test the divergence-based approach on newer architectures like Large Language Models (LLMs).

Summary updated

Abstract

Simultaneous machine translation (SiMT) requires a robust read/write policy in conjunction with a high-quality translation model. Traditional methods rely on either a fixed wait-k policy coupled with a standalone wait-k translation model, or an adaptive policy jointly trained with the translation model. In this study, we propose a more flexible approach by decoupling the adaptive policy model from the translation model. Our motivation stems from the observation that a standalone multi-path wait-k model performs competitively with adaptive policies utilized in state-of-the-art SiMT approaches. Specifically, we introduce DaP, a divergence-based adaptive policy, that makes read/write decisions for any translation model based on the potential divergence in translation distributions resulting from future information. DaP extends a frozen wait-k model with lightweight parameters, and is both memory and computation efficient. Experimental results across various benchmarks demonstrate that our approach offers an improved trade-off between translation accuracy and latency, outperforming strong baselines.
BibTeX
@inproceedings{zhao2023adaptive,
  author = {Libo Zhao and Kai Fan and Wei Luo and Wu Jing and Shushu Wang and Ziqian Zeng and Zhongqiang Huang},
  title = {{Adaptive Policy with Wait-k Model for Simultaneous Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=i0vMIpaEn4},
}
Figure