Models of interpreting

In considering prediction during simultaneous interpreting, it is important to consider the theoretical context in which this prediction takes place, in other words, the overall process of simultaneous interpreting, and in particular how comprehension and production, and source and target language, might interact. This section reviews models of the

interpreting process.

Gerver (1976) and Moser (1978) made the two earliest attempts to model the (simultaneous) interpreting process. Both models portray a mainly bottom-up view of language comprehension during simultaneous interpreting, in which the interpreter waits for chunks of information before beginning production in the source language. The models thus highlight the role of working memory (Gerver, 1976) and long-term memory (Moser, 1978) in allowing the formulation of these chunks. Moser’s (1978) model shows the language comprehension process as almost bottom-up, and taking place before the production process, portrayed as almost fully top-down (broadly in line with theories of comprehension and production at the time). The one exception to this “bottom-up” idea is the “prediction possible?” step in Moser’s model – however, this occurs very late in the process; once the target language has been “activated” and once the interpreter has

succeeded in paraphrasing, in the target language, information from the source speech. This

reflects the view that the two language systems are stored separately (something which is shown explicitly in Gerver’s 1976 model), as well as the implicit assumption that, when carried out by experts at least, translation into and production in the target language only occur when the source language meaning has been fully understood (Christoffels & De Groot, 2005). This is also in line with Kroll and Stewart’s (1994) model, which supposes that as bilinguals become more proficient in their L2, their two languages function in an

increasingly independent manner. Moser’s (1978) model also contains a step described as

“activation of target-language elements residing at the nodes of the conceptual network”

which appears to make direct reference to Seleskovitch’s (1976) theory that once the meaning of an interpreted message is understood in the source language, it is stored non-verbally, in a process called “déverbalisation” before being re-expressed in the target language (but see Isham, 1994 for evidence against déverbalisation). This is also in accord with models of language production which view the conceptual stage as non-language specific (e.g. De Bot, 1992). De Bot (2000) also assumes that during simultaneous

interpreting, the message is first non-verbally conceptualized before the production system starts working. However, importantly, Seleskovitch (1976), and De Bot (1992, 2000) fail to take into account the ongoing activation of the source language (as it is being

comprehended), and any resultant cross-language activation of translation equivalents (Francis, 2005) or homophones (Dijkstra & van Heuven, 2002).

De Groot (1997), and Macizo and Bajo (2004, 2006) have referred to these sequential process models of interpreting (in which the meaning of an utterance is fully understood before that utterance is translated and produced) as vertical translation⁵. They

5 If we take this definition of vertical translation literally, and assume that an utterance is equal to a sentence,

contrast this with a horizontal translation process, in which the target language is activated in parallel, and the message is represented and comprehended in the target language. This contrast of horizontal with vertical translation is also present in Paradis’ (1994) model, which again presents a sequential view of the interpreting process: from comprehension in the source language, to production in the target language, to monitoring in the target language. His model therefore recognizes the parallel nature of comprehension in

simultaneous interpreting (for both SL comprehension and error monitoring in the TL). He also proposes that translation takes place via one of two strategies – either by conceptual mediation (vertical translation) or structurally, by directly transcoding at the levels of phonology, morphology, syntax and semantics (horizontal translation). Table 1 provides an overview of models of interpreting; specifically, whether or not they consider interpreting as a sequential process, and whether or not they posit separate lexicons for source and target language.

Table 1. An overview of key interpreting models and whether they assume vertical translation and/or separate lexicons for source and target language

None of these models are compatible with the more recent models of bilingual and monolingual comprehension we have reviewed. On the one hand, the vertical approach to interpretation is not compatible with the findings showing that both of a bilinguals’

languages are activated in a bilingual setting. Daro (1994) has described the vertical approach as a “consolidated professional ideology” (Daro, 1994: 265). In other words, professional interpreters may prefer to assume that they always fully understand the meaning of the speaker before beginning their interpretation. On the other hand, the horizontal approach, which considers that comprehension takes place from the bottom up, is not coherent with findings that prediction routinely takes place in monolingual and bilingual comprehension. It seems more likely that top-down and bottom-up processes interact both within and between comprehension and production.

Christoffels and De Groot (2005) suppose that during simultaneous interpreting, multiple comprehension and production processes are ongoing: specifically that the interpreter comprehends new input, while memorizing prior input, mentally reformulating earlier segments into the target language and articulating even earlier segments. They propose different alternatives in an attempt to explain how, despite concurrent activation of both languages, only one language is chosen for production (with the other language inhibited). In their most parsimonious model, they suggest a shared lexicon for production and comprehension, containing separate lexicons for the source and target language, with the source language production system inhibited. The second model proposes an “input”

lexicon, linked to the comprehension system, and an “output” lexicon, linked to the production system. Each lexicon comprises the source and target language, with both source and target languages being activated in the input lexicon, and only the target

language being activated (and the source language being inhibited) in the output lexicon.

The third model is similar to the second, but there is no global activation/inhibition of source or target language. Instead, a subset of lexical items is activated in each language for input or output, to a greater or lesser extent. In all models, the language production and language comprehension systems are viewed as separate, and the lexicons are linked at the semantic level, which is considered to be non-language specific. These models thus account for the parallel nature of source and target language processing. However, they do not consider how comprehension and production, or source and target language, may interact at different levels of representation.

The models reviewed above, which have been central to interpreting theory and pedagogy, have not been empirically tested. However, more recently, two models focusing on individual aspects of the interpreting process have been the object of empirical study.

Seeber (2011) proposes that word order differences between source and target language constitute a source of cognitive load in simultaneous interpreting. He posits that interpreting sentences with the same syntactic structure in source and target language generates a baseline level of cognitive load. Where the syntactic structures differ between source and the target language (e.g., when a Subject-Object-Verb sentence in German must be translated with a Subject-Verb-Object sentence in English), he posits that cognitive load will increase, unless the interpreter predicts, and predictively produces the verb in the SVO sentence.

Seeber and Kerzel (2012) tested Seeber’s (2011) model⁶ of cognitive load in simultaneous interpreting using pupillometry. They found that cognitive load was indeed greater when sentence structure was not matched between source and target language.

However, the study did not establish whether prediction reduced this cognitive load. A potential confound of the study is that the mismatched German sentences had longer dependencies (the verb was further away from its object) than the matched German sentences (where the verb and object were adjoining). This could have created additional load regardless of the interpreting task (Lewis, Vasishth, & Van Dyke, 2006).

Meanwhile, Van Paridon, Roelofs and Meyer (2019) modelled the concurrent comprehension and production processes taking place in interpreting and shadowing. They compared interpreting and shadowing performance at different speeds, and found that interpreting performance was more error prone at all speech rates, and that increasing speech rate had a greater negative effect on interpreting performance than on shadowing performance. By modelling these findings, they found that they could be explained by a bottleneck account, according to which a lemma may not be selected for production and comprehension at the same time.

The early models of simultaneous interpreting are useful for describing the basic processes involved in simultaneous interpreting. However, as we have seen, they are not compatible with current evidence about how bilingual comprehension and production take place, nor is their aim to provide clearly defined hypotheses upon which to base empirical studies. More recently, Seeber and Kerzel (2012) and van Paridon et al. (2019) empirically tested the hypotheses of their models in controlled studies. These later models (Seeber,

2011, 2017; Van Paridon et al., 2019) thus provide empirically tested insights into key parts of the process. In Chapter 2 we present a model of prediction in simultaneous interpreting which can be tested empirically.