Information and Entropy Measures of Rendered Literal Translation

The definition and effect of translation literality (word-for-word translation, as opposed to sense-for-sense translation) have been a topic in translation studies for very many years. Schaeffer and Carl (2014) introduced measures to quantify the literality of observed translations based on their distance to a hypothetical absolutely literal translation assuming monotonicity (identical word order), compositionality (one-to-one translation), and entrenchment (lack of translation variation). This chapter introduces measures to assess the rendered translation literality, which reformulates the three literality criteria in an information-theoretic framework. We introduce rendered literality measures on the word and segment levels as the joint self-information and the joint entropy of the three literality criteria. These literality measures respect the additional requirement for translations to be grammatical according to target language (TL) constraints (i.e., rendered) and rank sets of alternative translations according to their rendered literality. We evaluate the predictive potential of rendered translation literality on translation duration within the multiLing dataset for different languages and different translation production modes. Our data show that the most probable translations respecting TL constraints are also most monotone, compositional, and entrenched translations. We conclude that the information-theoretic literality measures are powerful new predictors for behavioral measures, such as word and segment translation duration.