Language & AI
Translationese Explained as Cognitive Response to Task Difficulty
A new study attributes the distinctive features of translated text to the inherent cognitive strain of the translation process itself.
Translated texts consistently differ from comparable texts originally authored in the target language—a well-known pattern linguists call translationese. What drives these systematic differences has puzzled researchers for years, with explanations ranging from translator interference and text simplification to cultural factors and language-pair variations.
A New Framework#
Researchers investigating translationese through computational analysis now propose that the phenomenon emerges as a straightforward adaptation to translation task difficulty. Rather than viewing these patterns as byproducts of translator behavior alone, the team suggests they reflect how humans naturally respond to the cognitive demands of translation work.
Methodology and Findings#
The investigation employed a English-German corpus encompassing both written and spoken language to test this hypothesis. Translationese was quantified using a "translatedness score"—a classifier-based probability indicating whether a text segment originated as translation. Task difficulty was measured using multiple approaches: information-theoretic metrics derived from large language model surprisal, alongside traditional syntactic and semantic complexity measures.
Key discoveries include:
- Cross-lingual transfer difficulty proved a stronger predictor than source-text complexity across most scenarios
- Source-text syntactic complexity and translation-solution entropy emerged as the strongest individual predictors
- Information-theoretic features performed comparably to or better than conventional measures in written text, yet showed no advantage for spoken language
Limitations and Scope#
The analysis revealed that translation task difficulty accounts for only part of observable translationese, particularly in the English-to-German direction. This partial explanatory power suggests additional factors beyond cognitive load shape how translators produce target-language text.
