Variation Impoverishment resulting from Machine Translations: Empirical Evidence from Spanish Counterfactual Predicates

Abstract

The present article compares online machine translations and human translations in the aim of describing variation effects in Spanish output segments. Our survey focuses in a counterfactual construction existing both in English and Spanish but whose uses and frequencies differ in each language. English perfective should have predicates have a formal equivalent in Spanish-i.e. Debería haber + past participle-although the functional equivalent used by Spanish-speakers authors is a free-modal subjunctive construction-i.e. Hubiera + past participle (Repiso 2018). Our survey is based on 1.7 million-word Social Sciences corpus covering 8 essays, 4 political biographies and 2 dystopian novels. In all, 95 perfective Should have predicates were elicited in English. The human and machine translations in Spanish were subsequently analyzed. Our results show that human translations significantly preserve a wider range of lexical and morphological features compared to machine translations, which in turn favor prominently the formal equivalent construction Debería haber. The effect of this is the anglizing of the target language's morphosyntactic level by a quantitative reversal in the constructions pertaining to the irrealis semantic domain. Our survey suggests that a systemic use of MT risks to impoverish the target language outputs and thus, to oversimplify human communication. In a broader sense, MT challenges two phenomena widely observed in linguistics-i.e., natural languages' richness and individual differences within groups.