实时翻译研究综述

Summary

Conventional machine translation waits for the whole source input before producing output. That delay makes it unsuitable for live settings such as conference interpreting, travel, or streaming. Simultaneous translation (SiMT) instead outputs target text or speech while the source is still arriving. Zhang and Feng (ICT/CAS) note that the field lacked a dedicated systematic review and set out to fill that gap.

They first define the task formally. The model must decide when to emit each target token, based only on the source prefix seen so far. Systems are scored on two axes: translation quality and latency. The survey then groups existing methods by application scenario into four progressively more end-to-end families:

• text-to-text
• speech-to-text
• speech-to-speech
• LLM-based approaches

It names three core difficulties: choosing when to start translating, translating from incomplete input, and building end-to-end architectures. It also discusses future research directions.

The authors released a public literature collection organized by year, method, and venue. Together with the survey, it gives newcomers a structured map of a fast-growing research area.

Key points

• Origins and evaluation: The SiMT task is traced to Grissom II et al. (2014). Quality is measured with offline MT metrics (BLEU, COMET). Latency is usually measured with Average Lagging (AL) and its length-adaptive variant LAAL, which corrects AL's handling of over-generation.
• Text-to-text methods fall into four research lines: fixed policies, adaptive policies, training methods, and data augmentation.
• Fixed policies start from Wait-k (Ma et al.), where the decoder trails the source by k tokens, so g(i)=i+k−1. Several extensions address its weaknesses:
  • Adaptive Wait-k improves use of context.
  • Multipath Wait-k and MoE Wait-k let a single model serve many latency levels.
  • Variants at character, word, and information granularity target specific languages such as Chinese, Japanese, and German.
• Adaptive policies come in four types: agent-based (READ/WRITE decisions trained via reinforcement, supervised, or imitation learning), special-token-based, attention-based, and Transducer-based.
• Resource: a curated paper list is available at github.com/ictnlp/Simultaneous-Translation-Survey. The work was funded by NSFC grant 62376260.