Summary
Simultaneous Machine Translation (SiMT) requires balancing translation quality with latency. This paper introduces GSiMT, a generative framework that treats SiMT as a structural sequence-to-sequence learning problem. By using latent variables for "read/translate" actions and a Poisson prior for regularization, GSiMT optimizes the quality-latency tradeoff more effectively than traditional heuristic or reinforcement learning methods.
1. Introduction
Simultaneous Machine Translation (SiMT) is the task of generating a translation while the source sentence is still being produced. The core challenge lies in the quality-latency tradeoff: waiting for more context improves translation quality but increases delay, while translating early reduces latency at the risk of errors due to insufficient information.
Existing approaches typically fall into two categories: fixed Wait-$k$ policies, which are simple but lack flexibility for complex language pairs, and Reinforcement Learning (RL) methods, which attempt to learn dynamic policies but often suffer from training instability and high gradient variance.
2. The GSiMT Framework
This paper introduces GSiMT, a generative framework that models SiMT as a Neural Transducer problem. Instead of training a separate policy agent, the model treats the decision to "read" or "translate" as a latent alignment problem.
2.1 Key Technical Components
- Latent Action Variables: The model incorporates a binary latent variable $a_{i,j}$ at each step to represent the action (0 for reading a source word, 1 for generating a target word).
- Dynamic Programming (DP): During training, GSiMT uses DP to integrate over all possible sequences of actions. This approach significantly reduces variance compared to RL-based sampling, leading to more stable and efficient training.
- Poisson Prior: To regulate latency, the authors introduce a re-parameterized Poisson distribution. This acts as a "soft boundary" that encourages the model to stay close to the diagonal of the source-target alignment matrix. Researchers can tune the desired latency by adjusting the Poisson parameters ($\lambda$ and $\zeta$).
- Transformer Backbone: The system utilizes uni-directional Transformer encoders and decoders, ensuring the model only accesses the available prefix of the source and target streams.
3. Experimental Results
The framework was evaluated on the WMT15 German-to-English (DE→EN) benchmark and the multimodal Multi30K dataset.
- Superior Tradeoffs: On WMT15 DE→EN, GSiMT consistently outperformed Wait-$k$ and RL baselines. At a low latency level (Average Lagging ≈ 3.6), GSiMT achieved 28.82 BLEU, a substantial improvement over Wait-$k$'s 25.22.
- Multimodal Integration: In Multi30K experiments, the generative framework effectively leveraged visual features to resolve ambiguities, surpassing existing multimodal SiMT models.
- Test-time Flexibility: A key advantage of GSiMT is that a single pretrained model can be adjusted for different latency requirements at inference time by simply modifying the Poisson offset, eliminating the need for retraining.
4. Conclusion
GSiMT offers a robust alternative to traditional SiMT methods by reframing policy learning as a latent structure problem. By considering the entire distribution of possible translation paths and employing a flexible Poisson prior, the framework achieves state-of-the-art performance while maintaining the stability of maximum likelihood estimation.
Summary updated