Computer Aided Translation

Summary

The lecture examines how computer-aided translation tools can help human translators produce high-quality translations more efficiently. Its central premise is that machine translation is useful for many purposes, but publishable translation still depends heavily on people. Tools should therefore support translators rather than simply replace them.

Translation needs and workflow

The speaker distinguishes three broad uses of machine translation:

• Assimilation: Helping someone understand foreign-language material, even if the translation is imperfect.
• Communication: Supporting multilingual exchanges in which participants can clarify misunderstandings.
• Dissemination: Translating material for publication to readers who may rely on it as accurate. This is the most demanding use and the main focus of the talk.

To investigate where assistance might help, the research team recorded translators’ keystrokes as they translated French news into English. These logs made it possible to analyze typing, pauses, and revisions. The speaker describes short pauses as possible hesitation, medium pauses as time spent resolving a translation problem, and long pauses as potentially involving research or terminology checks. The initial study did not capture what translators were doing during pauses; later work added eye tracking.

The study involved university students rather than professional translators and used relatively accessible news content. The presenter cautions against treating the results as a definitive picture of professional translation. He notes that professional translators may work differently, including by reading less of a sentence before beginning to translate.

Assistance methods and prototype

The team compared several methods of assisting translators:

1. Sentence completion: As the user types, the system suggests continuations drawn from a phrase-based machine-translation model. It searches a precomputed decoder search graph, matching the user’s text to a likely path while aiming to respond quickly.
2. Translation options: The interface presents suggested word and phrase translations that users can click to insert.
3. Post-editing: The machine-translated sentence is placed in the editing area for the user to revise.
4. Combinations of these methods, alongside unaided translation.

The team demonstrated a prototype web interface using a Moses machine-translation engine. It showed the source text, the user’s draft, the raw machine translation, and—in post-editing mode—visual cues for changes made to the machine output. Users were not required to follow a particular editing sequence.

Findings from the user study

In the student study, all assistance conditions were faster on average than unaided translation. The reported average time per source word was:

Method Average time per source word Unaided 4.4 seconds Post-editing 2.7 seconds Translation options 3.7 seconds Sentence prediction 3.2 seconds Prediction and options together 3.3 seconds

The assistance mainly reduced typing effort and long pauses. However, the presenter stresses that the task was relatively easy and the participants were not professional translators. The measured speed gains therefore do not establish equivalent gains for professional work.

Quality results were difficult to interpret. Judges marked only about half of the translations as fully fluent and meaning-equivalent. The speaker considered this result surprisingly low and potentially affected by the evaluation method, so he avoided making strong claims about quality improvement.

Use of the tools also varied. Some participants made little or no use of predictions or clickable options. Participants disliked post-editing, even though it produced the fastest average result. The talk suggests that perceived effort and user acceptance matter alongside speed: an interface that feels like correcting a machine’s mistakes may be less appealing than one that lets translators actively construct a sentence.

The speaker also notes that post-editing depends on machine-translation quality. It may be useful for some language pairs and domains, but poor output can slow translators down or make the method impractical. Potential improvements discussed include better machine translation, confidence measures to avoid showing unreliable suggestions, and highlighting words likely to be wrong.

Monolingual post-editing

The lecture briefly covers experiments in which people who did not know the source language tried to improve or interpret machine translations, including Arabic material. This is not presented as a substitute for bilingual or professional translation. Results varied, and prior knowledge of the subject matter strongly affected performance.

The speaker suggests that such methods may be useful for understanding or triaging foreign-language material, where the goal is to identify relevant content rather than produce a publishable translation.

Reviews, guides, and tutorials

• Product review: None. The talk reports research studies and prototype evaluations rather than a consumer product review.
• Guide or tutorial: No step-by-step guide. The speaker demonstrates and explains a prototype for sentence prediction, translation-option selection, and post-editing.

Main speakers and sources

• Philip Koehn, a University of Edinburgh machine-translation researcher and presenter. His name is rendered as “Philip Kern” in the autogenerated subtitles.
• An unnamed Microsoft Research host, who introduces the speaker.
• Audience members, who ask questions during the lecture.