Abstract
Simultaneous Machine Translation (SiMT) requires a model to make real-time decisions on whether to read more source context or generate target words. This paper introduces the Dual-Path SiMT framework, which leverages the inherent symmetry of translation tasks. By jointly training forward (source-to-target) and backward (target-to-source) models and enforcing duality constraints through path transposition, the system learns a more precise read/write policy. Our approach significantly improves the balance between translation quality and latency, outperforming existing adaptive policies on standard benchmarks like IWSLT15 and WMT15.
Introduction
Simultaneous Machine Translation (SiMT) is critical for applications like live subtitling and real-time interpretation. Unlike traditional "offline" translation, SiMT models must operate on a stream, deciding at each step between a READ action (waiting for more context) and a WRITE action (generating a translation token). This sequence of decisions, known as the read/write path, must balance two competing goals: minimizing latency and maximizing translation quality.
Existing adaptive methods often lack direct supervision for these paths, leading to sub-optimal policies that either wait too long or guess prematurely. The challenge lies in finding the "Goldilocks" zone where the model has just enough information to translate accurately without unnecessary delay.
Methodology: Dual-Path SiMT
The core innovation of this work is the Dual-Path approach. It posits that the optimal information chunks for a source-to-target translation should mirror those of the corresponding target-to-source translation.
To implement this, the system maintains two models trained concurrently. The "Duality Constraint" is enforced by transposing the read/write path of one model to serve as a guide for the other. This transposition involves three key steps:
- Segment: Dividing the sentence pair into semantically equivalent "segment pairs" based on attention weights.
- Transpose: Swapping the source and target roles for these segments.
- Merge: Reconstructing a new read/write path for the opposite direction.
A joint optimization objective, including an $L_2$ distance penalty, forces the paths of the two models to align. Additionally, the authors introduced a "Single Path" modification to Monotonic Multi-head Attention (MMA) to ensure consistency across decoder layers, preventing "outlier heads" from causing unnecessary delays.
Results and Discussion
The Dual-Path model was evaluated on the IWSLT15 English-Vietnamese and WMT15 German-English datasets.
As shown in the results for the English-German task, the Dual-Path approach consistently outperforms strong baselines like wait-k and MMA across various latency levels. On the English-Vietnamese task, the model achieved quality levels (BLEU scores) nearly identical to offline models while maintaining low latency.
A key finding was the improvement in path precision. The model learned to start writing tokens much earlier than previous adaptive models once sufficient information was available, significantly improving the "Necessity" metric. Furthermore, the Intersection over Union (IoU) between forward and backward paths increased from 0.60 to 0.74 in the German-to-English task, validating that the duality constraints successfully aligned the two translation directions.
Conclusion
By treating SiMT as a dual task, the Dual-Path framework provides a robust mechanism for internal supervision of read/write policies without needing external word-alignment tools. This approach effectively addresses the lack of ground-truth paths in simultaneous translation, leading to state-of-the-art performance in balancing quality and latency. Future work may extend this duality principle to other streaming tasks such as speech-to-speech translation.
Summary updated