What Is QE?
QE (Quality Estimation) is an umbrella term for any automated translation quality scoring performed by a model rather than a person. It covers classic reference-based metrics like BLEU, reference-free approaches like MTQE, and LLM-judge scoring (GEMBA, LLM-as-a-judge).
If LQA is a human review, QE is its automated counterpart: faster, cheaper, but less sensitive to subtle meaning and cultural nuance.
Two Generations of QE
Reference-Based
The first generation of metrics — BLEU, and later the more refined chrF and TER — compare a machine translation against one or more human-written reference translations. Word and phrase overlap with the reference becomes a number.
Reference-Free
The second generation — COMET in reference-free mode and LLM-based approaches — needs no reference at all. The model is trained to predict how a human would score a translation, looking only at the source-target pair.
By 2026 the industry has mostly moved to the second generation: keeping reference translations for every new document isn't practical, and reference-free models' quality has come close to reference-based metrics on common language pairs.
Why QE Matters When LQA Exists
QE doesn't replace human review where an error is costly — it solves a different problem: how much content a human can realistically get through at all. A company translating millions of lines of user content a day can't physically route it all through LQA. QE lets a team:
- Prioritize — send low-scoring segments to editors first
- Auto-accept — skip high-confidence segments where the risk profile allows it
- Monitor the system — spot an engine's quality degrading from statistics, not from customer complaints
The 2026 QE Stack
The current industry approach isn't a single metric but a combination: COMET gives a fast overall score, while an LLM-judge rubric evaluates dimensions COMET can't separate on its own — adequacy, fluency, idiom transfer, cultural register, terminology consistency. No automated method replaces MQM-compatible human review where an error is expensive.
The Limits of Any QE
Every QE method, including the most modern ones, is trained to approximate an averaged human judgment — they're statistical by nature. A model can score a fluent text with a factual error highly if it stylistically resembles good translations in its training data, and underscore an unconventional but correct one. QE is a tool for sorting a stream, not a final verdict on text where a mistake is costly.
FAQ
Are QE and MTQE the same thing?
No: QE is the general term for any automated scoring, and MTQE is specifically its reference-free subtype. In practice the two are often conflated because modern pipelines almost always run reference-free.
Can QE fully replace a human?
For low-risk content, yes. For legal, medical, and marketing text, QE is used as a filter before a human, not instead of one.
Which QE metric is "best" in 2026?
There's no single "best" metric — the current approach combines COMET with an LLM-judge rubric rather than relying on one model.
How does QE relate to ISO 5060?
ISO 5060 describes the translation quality evaluation process as a whole and isn't tied to any specific tool; QE is one way to automate part of that process.
Does a small translation team need QE?
Yes, if content volume is growing faster than the editorial team — QE helps decide where to spend a limited amount of human attention first.