STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework

Abstract

Simultaneous translation, which translates sentences before they are finished, is useful in many scenarios but is notoriously difficult due to word-order differences. While the conventional seq-to-seq framework is only suitable for full-sentence translation, we propose a novel prefix-to-prefix framework for simultaneous translation that implicitly learns to anticipate in a single translation model. Within this framework, we present a very simple yet surprisingly effective "wait-k" policy trained to generate the target sentence concurrently with the source sentence, but always k words behind. Experiments show our strategy achieves low latency and reasonable quality (compared to full-sentence translation) on 4 directions: zh↔en and de↔en. * M.M. and L.H. contributed equally; L.H. conceived the main ideas (prefix-to-prefix and wait-k) and directed the project, while M.M. led the implementations on RNN and Transformer. See example videos, media reports, code, and data at https://simultrans-demo.github.io/. President Bush met with Putin in MoscowBùshí Bush zǒngtǒng President zài at Mòsīkē Moscow yǔ with Pǔjīng Putin huìwù meet prediction read write Source side → Target side → 2 Preliminaries: Full-Sentence NMT We first briefly review standard (full-sentence) neural translation to set up the notations.Regardless of the particular design of different seq-to-seq models, the encoder always takes

Simpler and Faster Learning of Adaptive Policies for Simultaneous Translation

“…6 Our test-time wait-results are better than those in Ma et al (2019), because the added source side <eos> token helps the full-sentence model to learn when the source sentence higher BLEU scores when latency AL is small, which we think will be the most useful scenarios of simultaneous translation. Furthermore, this figure also shows that our model can achieve good performance on different latency conditions by controlling the threshold , so we do not need to train multiple models for different latency requirements.…”

“…

Simultaneous translation is widely useful but remains challenging. Previous work falls into two main categories: (a) fixed-latency policies such as Ma et al (2019) and (b) adaptive policies such as Gu et al (2017). The former are simple and effective, but have to aggressively predict future content due to diverging source-target word order; the latter do not anticipate, but suffer from unstable and inefficient training.

…”

Monotonic Infinite Lookback Attention for Simultaneous Machine Translation

“…• The schedule is simple and fixed and can thus be easily integrated into MT training, as typified by wait-k approaches (Dalvi et al, 2018;Ma et al, 2018).…”

“…2. We extend the recently-proposed Average Lagging latency metric (Ma et al, 2018), making it differentiable and calculable in expectation, which allows it to be used as a training objective.…”

“…However, due to computational challenges, they pre-train an NMT model on full sentences and then train an agent that sees the fixed NMT model as part of its environment. Dalvi et al (2018) and Ma et al (2018) use fixed schedules and train their NMT systems accordingly. In particular, Ma et al (2018) advocate for a wait-k strategy, wherein the system always waits for exactly k tokens before beginning to translate, and then alternates between reading and writing at a constant pre-specified emission rate.…”