[Paper Review] Can neural machine translation do simultaneous translation?

Kyunghyun Cho, Masha Esipova|arXiv (Cornell University)|Jun 7, 2016

Natural Language Processing Techniques18 references134 citations

TL;DR

The paper introduces simultaneous greedy decoding, a method to adapt attention-based neural machine translation to simultaneous translation by jointly deciding segmentation and translation, enabling control over quality-delay trade-offs across language pairs En-Cs, En-De, and En-Ru.

ABSTRACT

We investigate the potential of attention-based neural machine translation in simultaneous translation. We introduce a novel decoding algorithm, called simultaneous greedy decoding, that allows an existing neural machine translation model to begin translating before a full source sentence is received. This approach is unique from previous works on simultaneous translation in that segmentation and translation are done jointly to maximize the translation quality and that translating each segment is strongly conditioned on all the previous segments. This paper presents a first step toward building a full simultaneous translation system based on neural machine translation.

Motivation & Objective

• Motivate the study of simultaneous translation within neural machine translation to balance quality and delay.
• Propose a decoding algorithm that allows translating while the full source sentence is still being received.
• Demonstrate joint segmentation and translation conditioned on past segments using a trained NMT model.
• Evaluate how waiting criteria impact the quality-delay trade-off across multiple language pairs.

Proposed method

• Introduce simultaneous greedy decoding that reads source in batches and greedily outputs target tokens when confident, while optionally waiting for more source material (controlled by delta and s0).
• Maintain encoder context as a dynamic set of source representations; update with new chunks only when a waiting criterion deems it beneficial.
• Use two waiting criteria (Wait-If-Worse and Wait-If-Diff) to decide whether to wait for additional source context versus translating immediately.
• Train standard attention-based NMT models (encoder-decoder with attention) on full sentences; apply decoding algorithm without retraining the model for simultaneous translation.
• Quantify delay with a normalized metric tau(X,Y) and analyze trade-offs between BLEU and tau across language pairs En-Cs, En-De, En-Ru.
• Compare simultaneous decoding performance to conventional consecutive greedy decoding and beam search baselines.

Experimental results

Research questions

• RQ1 Can an attention-based NMT model, trained for consecutive translation, be used for simultaneous translation via a new decoding strategy?
• RQ2 How do different waiting criteria (Wait-If-Worse vs Wait-If-Diff) affect the quality-delay trade-off in simultaneous translation?
• RQ3 What is the impact of decoding parameters delta and s0 on translation quality and delay across multiple language directions?
• RQ4 What qualitative behaviors (e.g., phrase repetition, premature commitment) emerge from simultaneous greedy decoding in various language pairs?

Key findings

• Simultaneous greedy decoding enables simultaneous translation with controllable trade-offs between translation quality and delay.
• Wait-If-Worse generally yields higher quality but larger delay, while Wait-If-Diff offers broader delay-quality trade-offs and may cause more repetitions.
• Translation quality and delay patterns vary across language pairs (En-Cs, En-De, En-Ru) and directions, influenced by morphological richness and syntax.
• There is a measurable quality-delay relationship: increasing information (larger delta or s0) changes when the model waits versus translates, affecting BLEU and tau.
• In some cases, Russian-English translations show premature commitment and phrase repetition under certain criteria, highlighting limitations of the current waiting strategies.
• BLEU scores (as a reference) indicate competitive performance relative to baselines when using the authors’ models: En->Cs 15.2 (Ours) / 13.84 (Star), En->De 19.5 / 21.75, En->Ru 17.77 / 19.54; In reverse directions: Cs->En 20.47 / 20.32, De->En 23.96 / 24, Ru->En 22.27 / 22.44.
• The approach demonstrates that end-to-end NMT systems can be repurposed for simultaneous translation without retraining for alignment or timing.