Translationese or “third code”

The features which characterise languages in contact have been well doc-umented in the literature on second language acquisition and have been termedinterference orinterlanguage (Selinker 1972), quasi-correctness (Papp 1984), lack of naturalness (Sinclair 1984), while in the context of translation a similar phenomenon has come to be known as thethird code (Frawley 1984) ortranslationese (Toury 1979, Gellerstam 1986):

Theoretical considerations highly verify it, and even lead to hy-pothesising that the language used in translation tends to be interlanguage (sometimes designated “translationese”), or that a translation is, as it were, an “inter-text” by definition. (Toury 1979: 227)

To our knowledge the term translationese was first used by Toury (1979) to refer to the supposedly peculiar language used in translations: in his work the term was strictly linked to the notion of interlanguage that had been developed in the field of second language learning to denote a “separate linguistic system [resulting] from a learner’s attempted production of a TL norm”, a system which “enjoys an intermediate status between SL and TL, and [...] reflects the interference of these two codes in the performance of the learner” (Toury 1979: 223).

Toury argued that since interlanguage forms are likely to occur whenever one language is used in contact with another, and more precisely, as a con-sequence of this contact, they are also unavoidable in translation, irrespective of the translator’s mastery of the two languages involved and even when translation is performed into the translator’s native language. In his view,

“the occurrence of interlanguage forms in translation follows from the very definition of this type of activity/product, thus being a formal ‘translation universal’” whose analysis “should form an integral part of any systematic descriptive study of translation as an empirical phenomenon” (ibid.: 224-225).

The idea that translation could be seen as a kind of separate sub-language drawing from both the SL and TL was similarly expounded by Frawley (1984), who argued that

the translation [...] is essentially a third code which arises out of the bilateral consideration of the matrix and target codes [...] it

emerges as a code in its own right, setting its own standards and structural presuppositions and entailments, though they are neces-sarily derivative of the matrix information and target parameters.

(1984: 168-169)

From this perspective and mainly seen as re-codification, translation is a new code which receives input from both the matrix (the SL/ST) and the target language (TL) codes: “the matrix code provides the essential information to be re-codified, and the target code provides the parameters for the re-rendering of that information” (ibid.: 161). However, “insofar as the third code supersedes its matrix information and target parameters, it differentiates itself” (ibid.: 169), representing “either a moderate innovation or a radical innovation with respect to the codes that contribute to its genesis”

(ibid.: 173, original emphasis), but in any case establishing itself as a new code with its own features.

Yet translationese is a complex phenomenon which manifests itself on each and every level of the linguistic system, and even beyond, it has often been used to refer to the difference in the distribution of particular lexical items or to the translated version of the target language (Gellerstam 1986).

The language of translations does not bear only features of SL, however. Like interlanguage in second language acquisition, translations also show universal traits: “features which typically occur in translated text rather than original utterances and which are not the result of interference from specific linguistic systems” (Baker 1993: 243). Such features are present even in translations which House (1977) has termed covert. In contrast to overt translations, which can easily be identified as such, covert translations sound like natural target language texts. Research (e.g. Laviosa 1997) has proven that even these texts carry features which differ from those of genuine texts in the same language.

Various attempts have been made to assess whether and how translated texts are really identifiable if compared with original texts. The typical methodology adopted in studies focusing on translationese is based on the construction of monolingual comparable corpora, with a view to identifying features which might be indicative of the distinction between translated and non-translated texts.

A first study carried out by Tirkkonen-Condit (2002) based on a Finnish monolingual comparable corpus aimed at investigating the human perception of translationese, i.e. whether it is possible for people to tell originals and translations apart, and at “identify[ing] the linguistic features shared by texts assumed to be translations, as well as those shared by texts assumed to be originally produced” (ibid.: 207). The conclusion drawn by Tirkkonen-Condit

on the basis of the quantitative results of her study was that translations could not be readily distinguished from original writing according to their linguistic features (ibid.: 216), and that evaluators tended to look for deviance to identify translated texts, whereas normalcy and idiomaticity were considered as signals of originality.

On the basis of experiments using Support Vector Machines (SVM) to automatically recognise translated texts within an Italian monolingual com-parable corpora (MCC), Baroni and Bernardini (2006) reported that features of translationese are “robust enough to be successfully used for the automated detection of translated text” (ibid.: 260). The perspective taken by them is quite novel: reporting a machine learning approach for the task of classifying Italian texts as translated or originals. Several features were employed in the feature vector, including unigrams, bigrams, trigrams, word forms, lemmas, and part-of-speech tags in order to prove that shallow data representations can be sufficient to automatically distinguish professional translations from non-translated texts with an accuracy above the chance level and hypothesise that this representation captures the distinguishing features of translationese.

The authors found that “an ensemble of SVMs reaches 86.7% accuracy with 89.3% precision and 83.3% recall [...] well above the average performance of ten human subjects, including five professional translators, on the same task”

(thus confirming Tirkkonen-Condit’s finding about the difficulty for humans to discriminate originals and translations). A particularly interesting aspect of Baroni and Bernardini’s study is that the SVM approach does not require the preliminary manual selection of features expected to be relevant for the translated/original distinction, thus reducing the risk of introducing a bias in the investigation.

That translation is qualitatively different from authentic text production

—although they both belong to the same linguistic code— is an idea which has received prominence in recent TS. More and more studies support the view that there is only partial overlap between the two modes of text production.

Correspondence is found where translated texts read like original texts, and what lies beyond the common area belongs to the third code. Used in a non-evaluative way, translationese refers to all those features, overt or covert, at each level of the linguistic structure which distinguish translated text from original, genuine language. The third code is used as a synonym while others likeinterlanguage or quasi-correctness are ignored because of their implicit evaluative stance. In line with Toury (1980) and Frawley (1984) who claim that translationese is the product of the translation process itself, which results from the confrontation of the SL and TL under circumstances specific to the process of translation. Translationese is regarded here as a particular code with its own characteristic features.

Translationese is made up of two sets of components: one set consists of features of TL which behave differently from what is typical of TL usage.

The second is made up of what has been referred to in DTS as universals of translation, i.e. explicitation, simplification and normalisation.