Explicit Anticipation for Simultaneous Machine Translation

This direction is merged into another direction: latency-quality trade-off. Its last version stays readable and citable here.

Version 1 of 125 papers26 conceptsfrozen 2026-10-04

In simultaneous machine translation (SiMT), a model must begin translating before the speaker has finished their sentence. This creates a fundamental conflict: to achieve low latency, the model must act on incomplete information; to achieve high quality, it needs the full context.

When a model is forced to generate a word before its source counterpart has appeared—a situation common when translating between languages with different word orders, like German (verb-final) to English—it often falls into the "hallucination trap." Rather than waiting, the model generates a plausible but factually incorrect guess to satisfy the timing constraint. Recent research has shifted from treating this as an implicit learning problem to a challenge of explicit anticipation, where models proactively predict the future of a sentence to bridge the information gap.

The Hallucination Trap of Forced Anticipation

Traditional SiMT models use a "policy" to decide whether to Read (wait for more source tokens) or Write (generate a target token). As noted by Shoutao Guo and Yang Feng in Simultaneous Machine Translation with Tailored Reference, standard training data assumes full-sentence context. When these models are deployed in streaming settings, they are often "forced" to anticipate future words.

Without a mechanism to handle this, models hallucinate. Hexuan Deng and Liang Ding observe in Improving Simultaneous Machine Translation with Monolingual Data that these hallucinations occur because the model is forced to commit to a translation path before the underlying facts are established. This is particularly problematic in speech-to-speech systems like Seamless, where autoregressive decoders struggle with fragmented inputs.

Peeking into the Future with LLMs

To mimic human interpreters who use linguistic patterns to guess upcoming phrases, researchers have introduced Translation by Anticipating Future (TAF). Zhehuai Chen and Siqi Ouyang demonstrate in Anticipating Future with Large Language Model for Simultaneous Machine Translation that Large Language Models (LLMs) can be used to explicitly "guess" the remainder of a source sentence.

By predicting multiple possible continuations and using a majority-voting mechanism to find a consensus, the system can generate a translation grounded in a "pseudo-future." This allows the model to produce initial translations that are logically reconcilable with the eventual conclusion of the sentence, a capability Hao Xiong and the Baidu team sought with their DuTongChuan system.

Quantifying Information Sufficiency

If we cannot guess the future, we must decide exactly when we have "enough" information to proceed. Shaolei Zhang proposed the Information-Transport-based Policy (ITST) in Information-Transport-based Policy for Simultaneous Translation, which treats translation as a process of transporting information units. By quantifying how much information has been received for each target word, the system makes more informed decisions about when to wait.

A similar intuition drives PsFuture, developed by Libo Zhao and Jing Li. In PsFuture: A Pseudo-Future-based Zero-Shot Adaptive Policy for Simultaneous Machine Translation, the model compares its prediction based on current input against a prediction "padded" with a generic future suffix. If the two predictions match, the model concludes it has sufficient context and proceeds to write; if they differ, it reads more.

Realigning the Training Signal

Another approach involves changing the data itself to match the constraints of simultaneity. Mingbo Ma and Liang Huang introduced pseudo-references in Improving Simultaneous Translation with Pseudo References. By rewriting training datasets into a more monotonic word order that follows the source stream, they reduce the need for the model to "wait" for distant verbs, thereby naturally reducing hallucinations.

Beyond Text: Multimodal and Document Context

When the immediate text stream is insufficient, models can look to other dimensions. Aizhan Imankulova and Masahiro Kaneko explored using visual cues in Towards Multimodal Simultaneous Neural Machine Translation. If a model sees an image of a "cake" while translating a partial sentence about baking, it can resolve the object ambiguity before the word "cake" is even spoken. Similarly, Studying The Impact Of Document-level Context suggests that information from previous sentences can help "sift" through the ambiguity of the current partial sentence, though effectively utilizing this context remains a significant challenge.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_explicit_anticipation_for_simultaneous_m_d2a6d331,
  title = {Explicit Anticipation for Simultaneous Machine Translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explicit-anticipation-for-simultaneous-machine-translation/rdr_d2a6d331f6f944a389abbab2ef2201d3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_explicit_anticipation_for_simultaneous_m_d2a6d331_v1,
  title = {Explicit Anticipation for Simultaneous Machine Translation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explicit-anticipation-for-simultaneous-machine-translation/rdr_d2a6d331f6f944a389abbab2ef2201d3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}