An Alternative to “Deverbalization”

Brian Mossop ©2003


In a common box-diagram depiction of the translation process, the SL linguistic ‘clothes’ of a text are removed (during a comprehension phase), revealing the contextualized Message, in what is presumably some sort of translinguistic form. This Message is then ‘redressed’ in TL clothes (during a reformulation phase).


Now this picture is certainly useful for pedagogical purposes, when beginning translation students are being taught to attend to ideas rather than words. However pedagogical utility does not imply theoretical value. That is, the mere fact that a theory may be useful does not imply or even suggest that it is true. In this paper, I propose that in fact the Text-(Deverbalized) Message-Text (TMT) theory is highly questionable as a picture of what goes on in the mind, and I put forward an alternative.


Direct SL-TL links

Simplifying somewhat, there are traditionally two models for how expressions in one language comes to be related to expressions in another language: either the forms of the SL are directly linked to forms of the TL, or else the link is mediated by a separate representation of meaning. Typically theorists ask such questions as: which of the two models reflects the way the mind works, or which of the two models reflects the way a particular translator is working (e.g. a student versus a professional), or which model is called for by the norms of the receiving society.


My proposal is that both processes (direct links and meaning-mediated links) occur simultaneously and they do so whenever someone is translating. It is never a case of one or the other.


At the outset, as the translator begins to hear or read a chunk of ST, he/she starts to interpret ST, that is, arrive at an understanding of its meaning. On the other hand, and at the same time, the translator’s bilingual brain automatically produces TL lexical and syntactic material based on the incoming SL forms and on the connections (whatever these may be) between TL and SL items in the mental store of language knowledge. I’ll call this activity of the brain Rendering. Rendering occurs beyond all conscious control and cannot be ‘unlearned’. It happens automatically: bilingual brains render just as stomachs digest incoming food.


The input to the composition of the translation thus consists of: a) the (unobservable) output of the brain’s Rendering work,  b) the output of the ST comprehension process at that point in time, and c) the translator’s various knowledges (of the purpose of the translation, of the subject matter, of TL genres and rhetoric, and so on). I’ll call the composing work of the translator, based on these inputs, Addressing (the translator addresses the receivers of the translation, quoting the ST to them).


The translator’s Addressing work can override the ST-influenced TL material automatically produced by the brain during Rendering, subtracting some or many of the oddities that arise from this influence. The extent to which overriding occurs will depend on the translator’s experience, on the social norms of translation under which the translator is working, on workplace procedures and factors (deadlines etc) and the processing strategy for a particular passage.


Let’s call this view of things Rendering-and-Addressing (R&A), and contrast it with the traditional Text-(Deverbalized) Message-Text (TMT) theory. Both theories contain an unexplained unconscious moment—brain activity that is not directly accessible. No theory of translation can avoid referring to this unconscious moment, for it is clear that translation, like all language production, is not entirely a conscious process. The question thus becomes one of relative plausibility.



In the present state of knowledge about how the brain works, we cannot prove any theory about unconscious brain processes during translation. Still, we can ask whether there are any observations which can provide at least initial plausibility for a view.


The plausibility of TMT rests on the observation that the wording of a translation sometimes has very little lexico-syntactic resemblance to the wording of the source text (Setton 1999:xx). One can hardly conclude from this, however, that TMT ‘must’ be true. R&A has a different account: the structurally different wording of the translation arises during intralingual rewording of the output of Rendering. The source text is certainly interpreted by the translator, but it is not deverbalized into a translinguistic representation of meaning, which is then reverbalized in TL.


The initial plausibility of R&A rests on two kinds of translational phenomena manifested by aphasic bilinguals, which make it plausible to believe in the existence of direct SL-TL links in the brain that are independent of the SL comprehension and TL production processes. The evidence comes from bilinguals who have suffered brain damage and display two odd forms of translational behaviour (Fabbro 2000: xx).


The first odd behaviour is spontaneous translation. There are many reports of patients who, asked a question in one of their languages (e.g. What time is it?) will respond not by answering the question but by translating the question into their other language. (They will not say ‘three o’clock’ but rather ‘Quelle heure est-il?’). The second interesting behaviour of brain-damaged bilinguals is paradoxical translation. During their recovery period, these patients have a strong language, in which they can converse, and a weak language, in which they have great difficulty conversing. The paradox is that while they cannot produce in the weak language in response to what someone else has said in that language, they can easily translate from their strong language into their weak language. 


Neurolinguists assume that such pathological behaviours tell us something about the normal state. The theoretical basis for this is the assumption that different aspects of mental activity correspond to different areas of the brain; thus if an injury to a given part of the brain regularly corresponds to a particular pathological language behaviour, while other aspects of language are preserved, then we have identified a ‘module’—a particular aspect of mental functioning.


Thus both spontaneous and paradoxical translation appear to suggest the existence of direct linkages in the mind between SL and TL lexicogrammatical material, independent of ‘meaning’. In addition, paradoxical translation suggests that translational language production involves a different mental process from non- translational production: the latter is defective in these patients’ weak language, but the former continues to operate.[1]


Whatever one may think of this initial evidence, arguments in favour of R&A over TMT cannot rely on empirical evidence in the present state of knowledge. One must therefore rely on other kinds of argument, such as the scope and simplicity of the two theories.



I suggest that R&A is more plausible than Text-Message-Text because it offers a unified account of both ‘bad’ and ‘good’ translation, in terms of what does or does not happen during the second, Addressing phase of the language production process. The Text-Message-Text theory has to call on a separate ‘transcoding’ path (Text-to-Text) to account for ‘bad’ translations.


R&A treats the production of unidiomatic TL expressions as normal in translation and accounts for the degree of its presence in terms of the clash  between Rendering and Addressing. TMT excludes ‘transcoding’ as non-translational. This of course in no way disqualifies TMT as a reflection of reality: it could be that TMT reflects what happens when ‘good’ translators are at work. Still, it is unsatisfying to have a theory that accounts only for a predefined notion of ‘good’ translation.


TMT seems to rest on the idea that somehow a (well-trained) translator can get the wording of ST out of his/her mind. This seems implausible, because the production of the translation typically begins within seconds of reading or hearing ST. This is always so with the oral forms of translations (4-5 seconds in the case of simultaneous); and even with written translation, translators very frequently begin composing immediately after reading (sometimes just reading the first part of the sentence), with no lengthy period of cogitation.


The short time gap means that the wording of ST does influence the translation, and indeed there is growing evidence that even experienced translators never manage to completely eliminate it. R&A accounts for several degrees of interference resulting from exactly the same mental process, whereas TMT seems to suggest a false duality: either deverbalization leading to complete elimination of interference or ‘transcoding’ leading to a bad translation filled with interference, depending on which of two different mental processes is followed.


Rendering & Addressing theory avoids one great drawback of TMT theory, namely the need to posit a representation of the message in some sort of translingual ‘language of thought’ (LOT). Invoking a LOT has the effect of introducing two new questions: what are the units and combination rules of LOT? and how is a sentence in SL turned into a LOT representation, and then a LOT representation into a sentence in TL?


In an article proposing a type of Text-Message-Text model of translational language production, specifically with regard to simultaneous interpretation, de Bot (2000:xx) admits that the evidence is “not all that clearcut”:


If speaking takes place more or less along the lines of the model, then there is simply no direct way to connect the incoming speech and the speech produced: all incoming speech will be parsed, delexicalized and turned into a conceptual code which accordingly serves as input for the production mechanism. So, there is no direct relation between the grammatical patterns of the source text and the target text. The evidence is not all that clearcut, however, There is quite some data suggesting that in simultaneous interpretation the interpreters tend to use grammatical structures that are similar in the two language, even if that leads to production problems. [my italics]


R&A theory has the additional advantage of avoiding the whole question of whether LOT exists. It does so by separating the SL-TL conversion process (Rendering) from the comprehension process (assumed to be more or less identical to non-translational understanding, and therefore not requiring an account within translation theory) and from the composition process (seen as intra-TL work on the output of Rendering). If LOT does exist, R&A theory is not invalidated: one could claim that LOT is involved somehow in the Addressing work, but this Addressing work is still a unilingual operation on the output of Rendering. Arguments about the existence of LOT are likely to remain purely deductive and argumentative in nature for some time; there is at present no empirical evidence bearing on the question.




de Bot, Kees, “Simultaneous Interpreting as Language Production” in Dimitrova & Hyltenstam (eds)               Language Processing and Simultaneous Interpretation, Amsterdam: Benjamins, 2000.

Fabbro, Franco, The Neurolinguistics of Bilingualism, Hove UK: Psychology Press, 1999.

Setton, Robin, Simultaneous Interpretation: a cognitive-pragmatic analysis, Amsterdam:Benjamins 1999.

Smith, Neil and I.M. Tsimpli, The Mind of a Savant, Oxford: Blackwell 1995.



[This article is based on passages in the manuscript of a book in preparation entitled Strange Words: Translating as Peculiar Language Production.]

[1] Also of considerable interest is the translation behaviour of a mentally impaired man described at length in Smith & Tsimpli 1995 (a book about the separability of language from other cognitive faculties). His translations are described by the authors as ‘automated’ and ‘lacking any sense of cohesion and coherence’.]