Abstract
Simultaneous translation is widely useful but remains one of the most difficult tasks in NLP. Previous work either uses fixed-latency policies, or train a complicated two-staged model using reinforcement learning. We propose a much simpler single model that adds a 'delay' token to the target vocabulary, and design a restricted dynamic oracle to greatly simplify training. Experiments on Chinese<->English simultaneous translation show that our work leads to flexible policies that achieve better BLEU scores and lower latencies compared to both fixed and RL-learned policies.
Introduction
Simultaneous translation (ST) aims to generate target language output in real-time while the source language input is still being received. This task is essential for live scenarios such as international conferences and negotiations. The primary difficulty lies in the structural divergence between languages (e.g., SOV vs. SVO word orders), which requires a policy to decide when to wait for more context (READ) and when to output a translated word (WRITE).
Existing approaches generally fall into two categories: fixed-latency policies (like wait-$k$), which are simple but lack adaptability to varying linguistic contexts, and reinforcement learning (RL) based policies, which are flexible but notoriously difficult to train due to their complex two-stage architectures and unstable optimization. This paper proposes a "middle way": a single, end-to-end model that is as simple as a standard translator but as flexible as an RL agent.
Methodology
This work proposes a unified end-to-end framework that simplifies ST into a standard sequence-to-sequence task with a specialized vocabulary and a novel training objective.
The Delay Token ($\langle \epsilon \rangle$)
Instead of a separate policy network, we augment the target vocabulary with a special "delay" token, $\langle \epsilon \rangle$. Predicting $\langle \epsilon \rangle$ corresponds to a READ action, while predicting any other token corresponds to a WRITE action. This allows a single Transformer model to learn both translation and timing. To prevent semantic interference, $\langle \epsilon \rangle$ tokens are excluded from the self-attention mechanism but tracked via a separate delay embedding, ensuring the model retains timing information without semantic clutter.
Restricted Imitation Learning
To train this model, we employ a "restricted dynamic oracle." Unlike a standard oracle that provides a single gold path, our dynamic oracle defines a set of permissible actions at each step based on the full source sentence. We introduce two bounds to constrain the model's behavior:
- Aggressive Bound ($\alpha$): Prevents the model from waiting excessively, ensuring low latency.
- Conservative Bound ($\beta$): Prevents the model from guessing words before sufficient source information is available, ensuring translation quality.
The model is trained to stay within this latency "band," effectively imitating an expert that balances speed and accuracy. Furthermore, we found that using a Sigmoid function—allowing READ and WRITE choices to be treated independently—outperforms Softmax when the oracle deems both actions acceptable.
At inference, a temperature parameter can be applied to the $\langle \epsilon \rangle$ token to adjust the model's patience without retraining, allowing a single model to span multiple latency requirements.
Experiments and Results
We evaluated our approach on the NIST Chinese-to-English and English-to-Chinese translation tasks using 1 million sentence pairs for training.
Performance Comparison
Our adaptive policy consistently outperforms both fixed-latency (wait-$k$) baselines and state-of-the-art RL-based models. As shown in the quality-latency trade-off curves, the proposed method achieves higher BLEU scores at comparable or lower Average Lagging (AL) values.
The results demonstrate that a single model trained with restricted imitation learning can span a wide range of latencies, offering greater flexibility and better performance than multiple separately-trained fixed-latency models. The "temperature knob" effectively allows for real-time adjustment of the latency-quality trade-off.
Conclusion
By treating the "wait" action as a vocabulary token and utilizing a restricted dynamic oracle for training, we have developed a simple yet effective framework for simultaneous translation. This approach eliminates the complexity of RL while providing the adaptability needed for real-time translation. Future work will investigate the application of this method to speech-to-text translation and language pairs with more extreme word reordering, such as Japanese-English.
Summary updated