Research presented at the Seventh Conference on Machine Translation (WMT22) highlights major advances in reference-free Quality Estimation (QE) and document-level neural translation that resolves cross-sentence coreference and pronoun agreement.

Moving Beyond Sentence-by-Sentence Isolation in Neural MT

For decades, both statistical and neural machine translation systems translated text one isolated sentence at a time. While sentence-level translation works adequately for standalone strings, it fails whenever grammatical accuracy depends on surrounding discourse context—such as resolving whether the English pronoun 'it' or 'they' refers to a masculine, feminine, or neuter noun in the preceding sentence, or maintaining consistent lexical choices and formal/informal register across an entire document. At WMT22 in late 2022, document-level NMT architectures demonstrated substantial gains in discourse coherence.

Reference-Free Quality Estimation (MTQE) in Production Workflows

Alongside document-level context, WMT22 showcased rapid maturation in reference-free Quality Estimation (QE) models such as CometKiwi. Unlike evaluation metrics that require a pre-existing human reference translation, QE models predict translation quality and highlight word-level error spans using only the source segment and the MT output. In enterprise TMS pipelines, QE scores help project managers identify segments that require intensive human revision versus high-confidence segments.

Need certified multilingual translations, sworn legal asseverations, or ISO-compliant localization for your international projects? Request a free customized quotation from Alphatrad and receive a detailed turnaround plan within 24 hours.