Is interactive translation the same as automatic post-editing?

No. Automatic post-editing revises output algorithmically; interactive translation uses a person's feedback in an ongoing prediction loop.

Must users correct from left to right?

That is a feature of prefix-based protocols. Other research designs allow corrections and validation in separate parts of the sentence.

Does each correction update model parameters?

Not necessarily. The decoder may simply condition the current prediction on the correction. Persistent adaptation requires a separate learning mechanism.

AI Glossary / Natural language and speech / Intermediate

Interactive machine translation

Definition

Interactive machine translation is a translation process in which a person and a machine translation system repeatedly exchange suggestions and corrections. The system uses the person's feedback to generate a revised proposal during the current task. A common design preserves an accepted prefix and predicts the remaining text, while other designs support corrections in multiple segments.

What it means

The central feature is a feedback loop. A translator examines a proposal, accepts or changes part of it, and receives a new proposal that takes that intervention into account.
This differs from simply editing a fixed machine translation after it has been produced. In an interactive system, the machine actively updates its suggestions as the work proceeds.
The term does not require one particular translation model or input method. It also does not mean that every correction permanently retrains the model. Using feedback to constrain the current sentence and learning from feedback for future sentences are separate capabilities.

How it works

The system first receives a source sentence and proposes a target-language translation. The person checks the suggestion against the source and identifies what to accept or correct.
In a prefix-based protocol, the person validates an initial portion and supplies a correction at its end. The system then searches for a completion compatible with that accepted prefix.
Other protocols allow selected, potentially disjoint segments to be validated. The decoder must respect those constraints when generating the remaining material.
Repeat until the person accepts the result. A usable interface also needs clear handling of revisions, undo actions, and changes to previously accepted text. The model's output remains a proposal until the workflow's acceptance criteria are satisfied.

Three practical examples

A translator corrects a technical term near the start of a sentence and receives a newly predicted continuation consistent with that choice.
A research interface lets a reviewer retain several correct fragments while requesting new wording for the gaps between them.
An evaluation compares interactive completion with ordinary post-editing, measuring final quality and actual completion time as well as keystrokes.

Think of it this way

Think of a writing partner who revises a draft after each correction while keeping the parts you have explicitly approved. The analogy is bounded because a translation system follows constraints and learned patterns; it may not understand why the correction was made and can still propose a misleading continuation.

Worked example

Consider a constructed French-to-English task. The source is “La réunion commence à neuf heures demain.” An initial proposal is “The meeting starts at ten tomorrow.”
The translator validates “The meeting starts at” and corrects the next word to “nine.” A prefix-based system receives the accepted prefix “The meeting starts at nine” and proposes the remaining word “tomorrow.” The combined result is “The meeting starts at nine tomorrow.”
The important operation is not merely replacing ten with nine. The system is asked to produce a continuation under the newly accepted prefix. In a longer sentence, that continuation may differ substantially from the original suffix.
The person still checks the full result against the source. If an interface allows disjoint validation, the translator might preserve tomorrow separately as well, but a strict prefix-only protocol does not provide that same interaction.
These sentences are teaching examples, not outputs from a named model. The example demonstrates feedback constraints rather than a measured reduction in translator effort.

Illustrative case study

In a fictional project, Alder Languages tests an interactive interface with professional translators working on short product descriptions. Users can accept a prefix and request a new continuation after a correction.
Some users make fewer keystrokes but spend extra time rereading changing suggestions. The team therefore measures completion time and final translation quality instead of assuming that fewer edits prove higher productivity.
The imagined study also asks whether the interface preserves approved terminology and makes updates easy to follow. It reports no actual benchmark result and does not assume all translators prefer the same correction order.

Common misconceptions and mistakes

Interactive machine translation is not simply a chatbot capable of translating when asked. The relevant feature is how feedback constrains and updates the translation process.
Human involvement does not guarantee an error-free final text; reviewers can miss errors or accept misleading suggestions.
A system that honors a correction in the current sentence has not necessarily learned it for future use.
Prefix-based interaction is common but not the only protocol.
Lower edit counts or simulated keystroke savings do not automatically establish lower cognitive effort or faster work by real translators.

Business use

For a business translation workflow, interactive prediction may be useful when users want machine suggestions while retaining close control of wording. Its benefit depends on translation quality, response latency, interface stability, and how well the interaction fits the translators' work.
Evaluate with representative users and material. Include terminology preservation, rereading effort, total task time, and final quality. Make it clear which text is accepted and which text is newly proposed, especially when suggestions change after each intervention.

Check your understanding

A user validates an initial phrase, and the system produces a new suffix compatible with it. Does that prove the system will remember the user's preferred term next week?

Answer: No. It shows that the current prediction used the feedback. Future behavior depends on whether the application stores preferences or updates a model through an explicit adaptation process.

Related terms

Post-editing
Machine translation
Online learning
computer-assisted translation
translation prediction
constrained decoding
terminology management
human-in-the-loop AI

Learn next

compare one prefix-based interaction with one segment-based interaction on the same sentence. Record not only edits but also how much text the reviewer must reread after each change, then distinguish sentence-level constraints from persistent model adaptation.

Sources

González-Rubio, Ortiz-Martínez, Benedí, and Casacuberta, “Beyond Prefix-Based Interactive Translation Prediction,” original 2016 paper: https://aclanthology.org/K16-1020.pdf. The paper explains prefix and segment-based feedback and discusses limitations of simulated effort measures. The meeting example and fictional interface evaluation above are original teaching material. Independent source review: The original paper’s prefix and disjoint-segment protocols and cautions about simulated effort versus actual human evaluation were read.

Previous: Interactive evolutionary computation Next: Intercept