Fixed and Adaptive Simultaneous Machine Translation Strategies Using Adapters

TL;DR

The paper tackles the latency-quality trade-off in simultaneous machine translation by introducing lightweight adapters into the Transformer decoder to support multiple fixed wait-$k$ latencies within a single model. It combines multi-path training with wait-$k$ adapters and adds an adaptive inference strategy (Adaptive Adapters) that selects reads or writes based on token probability and lag $k = |x| - |y|$, controlled by thresholds. The method achieves competitive or superior results versus strong baselines on En-Vi and De-En while reducing the need for training and maintaining efficiency, especially at low latency. This approach offers a practical, parameter-efficient way to deploy latency-aware SiMT systems with flexible latency control in real-time translation scenarios.

Abstract

Simultaneous machine translation aims at solving the task of real-time translation by starting to translate before consuming the full input, which poses challenges in terms of balancing quality and latency of the translation. The wait-$k$ policy offers a solution by starting to translate after consuming $k$ words, where the choice of the number $k$ directly affects the latency and quality. In applications where we seek to keep the choice over latency and quality at inference, the wait-$k$ policy obliges us to train more than one model. In this paper, we address the challenge of building one model that can fulfil multiple latency levels and we achieve this by introducing lightweight adapter modules into the decoder. The adapters are trained to be specialized for different wait-$k$ values and compared to other techniques they offer more flexibility to allow for reaping the benefits of parameter sharing and minimizing interference. Additionally, we show that by combining with an adaptive strategy, we can further improve the results. Experiments on two language directions show that our method outperforms or competes with other strong baselines on most latency values.

Figures (8)

• Figure 1: Transformer Decoder with Adapters Wait-$k$, we illustrate an example where 8 adapters are inserted with $K_A=\{1,3,5,7,9,11,13,15\}$, the generation step is $t=0$, and $A_3$ is activated because $k=3$.
• Figure 2: Translation quality (BLEU) against latency (AL) of our methods (Adaptive Adapters, Adapters Wait-$k$) and previous adaptive (MMA, Adaptive Wait-$k$) and fixed (Wait-$k$, MoE Wait-$k$, Multi-path Wait-k) strategies on En-Vi and De-En.
• Figure 3: Results of increasing the value of $k_{max}$ on De-En. Lower $k_{max}$ values achieve better BLEU score in low latency, but it is necessary to increase the value of $k_{max}$ in order to support high latency.
• Figure 4: Wall-clock time comparison between Adapters Wait-$k$ and Multi-path Wait-$k$ averaged over 5 runs on En-De.
• Figure 5: Results of varying the window sizes of the adapter lagging between 1 and 5 on En-Vi.