Summary of findings

As explained in section 4.2.3, when significant interactions are found between input variables, statistical tests don’t investigate the effect of any one input variable acting on its own, as the results of doing so could be misleading. Especially if there are not just two but three interacting input variables, as in this study, trying to interpret the effect of just one of those variables without considering the other two could lead to meaningless results. Strictly speaking, reliable conclusions can only be drawn from analyzing the effects of all three input variables in interaction. But the formulas created and applied to the linear regression model in this study did yield results on the effects of each pair of input variables, in addition to the combined effects of all three. For completeness, all those results were presented above and are reflected in the discussion below.

A statistical analysis of our corpus data involving English and five target languages from different families has shown that observed rates for all three features identified as indicators of difficulty in translation or interpretation are closely associated with structural difference in the language pair, in interaction with sentence complexity. The effect of sentence complexity on the indicators of difficulty is unsurprising. All three indicators involve changes in the linear or hierarchical arrangement of propositions. The more subordinate propositions there are in a sentence, there more scope for changes in that arrangement there will be. The more important finding is that structural difference between languages also seems to have a major effect on the indicators of difficulty – an effect which becomes increasingly pronounced in increasingly complex sentences.

At each level of sentence complexity, by far the highest rates for each indicator of difficulty were observed in Turkish and Japanese. Of the target languages considered, those were the ones with the greatest structural difference from English. Also, for all language pairs, reordering and nesting changes were much more frequent in legal translation than in the other two modes. Changes in semantic relations were especially characteristic of simultaneous interpretation.

Based on the corpus data, the formulas produced by the linear regression models also predicted a number of associations which may be generalized to similar texts, talks and speeches. The rates for each indicator of difficulty were predicted to increase greatly with increasing structural difference in the language pair, in combination with sentence complexity. In legal translation, this was especially true for reordering and nesting changes, and in simultaneous interpretation for changes in semantic relations.

The predicted effects of structural difference and sentence complexity on the three indicators of difficulty have some general implications which apply to all three modes of language transfer considered in this study. Especially for very complex sentences, those predicted effects suggest that, the more a language pair differs in structure: (a) the more the linear order of propositions is likely to change in translation or interpretation in that pair, (b) the more nested structures are likely to be created or eliminated in translation or interpretation in that pair, and (c) the more the semantic relations between propositions are likely to be changed in translation or interpretation in that pair. Taken together, these findings suggest that, across all three modes of transfer, the more a language pair differs in structure, the more difficult it’s likely to be to translate or interpret a complex sentence – especially a very complex sentence – in that pair, and the more the original meaning is likely to be changed.

Determining the precise causes for these associations is beyond the scope of this study. Below are some speculative explanations as to why structural difference in a language pair may make it more difficult to translate or interpret complex sentences in that pair, along with some related observations.

5.1.1 Structural difference and reordering

It’s been shown (Hawkins 2014) that having to keep several logical processing windows open at a time – as a translator or interpreter needs to do if juggling propositions around in different order – is more difficult than opening one logical processing window at a time, closing it, then moving on to the next one – as a translator or interpreter can do if transferring propositions in parallel order. It’s also been shown (Donolato, Giofrè and Mammarella 2017) that recalling verbal information in reverse order is harder than recalling it in the order it was received.

Studies by Hartsuiker et al. (2004, 2016) and others provide evidence that bilingual people have a single mental representation of structures which are similarly ordered in their two languages. This suggests that, besides problems involving the order of recalled information, a translator or interpreter working between languages where subordinate clauses branch in opposite directions has to deal with the added challenge of managing two different mental constructs when processing relations between propositions. In contrast, a translator or interpreter working between structurally similar languages can process those relations using a single mental construct.

Reordering is also worth recording to quantify it in its own right. It’s informative to see, for each language pair, how far away the average proposition needs to be shifted in different modes of language transfer. In legal translation of a complex sentence between languages with very different structure, like English-Turkish and English-Japanese, the total number of place shifts recorded in this study sometimes reached triple digits for a given sentence, while the same sentence had been translated with little or no reordering between structurally similar languages.

5.1.2 Structural difference and nesting changes

The findings of this study suggest that translation or interpretation of a complex sentence from a right-branching language like English into a left-branching language like Turkish, Japanese or to some extent Mandarin is likely to lead to many more nestings in the translated or interpreted version of the sentence than in the original version, and therefore to be more difficult to produce and to process. This may be partly due to a basic asymmetry between the two directions in which a subordinate proposition can branch from its parent.

Left-branching languages tend to be “head-final,” with a phrase head typically coming at the end of its phrase in most phrase types. So a syntactic phrase expressing a proposition in such a language will generally have its predicate at the end of the phrase, preceded by any other constituents. One of those constituents is likely to be the subject of the predicate, which, even in a left-branching language, tends to be near the beginning of the phrase. This can make for lots of long‑distance attachments, with several constituents sandwiched between the subject and predicate of a long proposition. One or more of those intervening constituents can themselves be propositions. And any of those nested propositions can in turn be split, leading to multiple layers of nesting.

In my experience, a long, complex sentence in a left‑branching language tends to have many more nestings than a comparable sentence in a right‑branching language. A Turkish or Japanese legal text, for example, is likely to have sentences with many more nestings and long-distance attachments than a comparable text drafted in English, German or Russian. This makes the phrasal combination domains in those sentences more difficult to establish and to process.

Also in my experience, this nesting tendency of left-branching languages is liable to be compounded in translation from a right-branching language. A sentence in a European language can be long and complex, but have few or no nestings, consisting of a series of unbroken propositions, with each subordinate proposition linked to the end of its parent. But transferred with structural accuracy into a language like Turkish or Japanese, that same sentence can often end up with multiple nestings. This typically happens in written translation, in a genre characterized by long, complex sentences – unless the translator makes a special effort to change the hierarchical structure of the original so as to avoid nested structures in translation. They may be reluctant to do that, especially in a text which has legal consequences, or even in an article they wish to translate faithfully. Changing the hierarchical structure of an original sentence in translation also risks distorting its meaning.

I submit that these two tendencies – the inherent nesting tendency of left-branching languages, compounded by the extra nesting effect of translation from a right-branching language into a left-branching one – may go a long way towards explaining the strong preliminary association found between structural difference and recorded values for nesting changes in the language pairs considered in this study.

5.1.3 Structural difference and changes in semantic relations

A change in semantic relations in a written translation of a complex sentence compared to the original version of that sentence presumably means that the translator has: (a) had trouble understanding the structure of the source language, (b) had trouble reproducing that structure in the target language, or (c) chosen to depart from the original structure for some reason. Such a change can also result from any combination of the above factors. As explained in section 3.3.4 on changes in semantic relations, this study doesn’t attempt to determine whether difficulty as reflected in such changes results from necessity or from choice.

A translator working between languages with largely parallel structure, like two European languages, can generally reformulate the propositions of a complex sentence one by one and in order. This includes the subordinating links between propositions – like subordinating conjunctions, relative pronouns and prepositions – which they can transfer directly from one language to the other, without having to take apart and reconstruct the logical relation established by each one.

On the other hand, if a translator working in a language pair with very different structure chooses or needs to change the hierarchical relations between propositions in the translation of a complex sentence compared to the original version, that may be because faithfully reproducing those relations would result in a translation that’s hard to process or that sounds awkward. Such restructuring is likely to involve extra effort. Changing the hierarchical relations between propositions also creates a greater risk of distorting the original meaning.

In simultaneous interpretation between languages with very different structure, a change in semantic relations in an interpreted version of a sentence compared to the original version of that sentence can be due to any of the reasons described above for written translation. It’s also likely to be due to the fact that too great a burden has been placed on the interpreter’s working memory.

For these reasons, I speculate that changes in semantic relations between propositions, whether deliberate or not, are a sign of extra effort expended by the translator or interpreter in processing those relations, reproducing them or both. This may help explain the associations found in this study between structural difference in a language pair and recorded rates for changes in semantic relations.

5.1.4 Differences between modes

This study has established preliminary associations between structural difference in a language pair, in interaction with sentence complexity, and rates for indicators of difficulty in translating or interpreting complex sentences in that pair, with some differences according to the mode of language transfer. The higher rates observed and predicted for reordering and nesting changes in language pairs with very different structure appear to be especially characteristic of legal translation. This may be due to the greater sentence complexity typical of legal texts, and to the fact that standard written translation is largely free from the time, space or working memory constraints inherent to the other two modes. On the other hand, higher rates for changes in semantic relations appear to be especially characteristic of simultaneous interpretation.

Among the language pairs considered, the observed and predicted rates for reordering and for nesting changes in language pairs with very different structure are relatively low in subtitle translation and simultaneous interpretation, and appear less strongly associated with structural difference in those modes than in legal translation. For subtitle translation, this more parallel order with respect to the original may be partly due to the need or desire for subtitle segments to run parallel to the video image. For simultaneous interpretation, it may well be partly due to the constraint on interpreters’ working memory.

On the other hand, the observed and predicted rates for changes in semantic relations in language pairs with very different structure are particularly high in subtitle translation and in simultaneous interpretation, and appear more strongly associated with structural difference in those modes than in legal translation. This may be for similar reasons to those suggested for reordering and nesting changes. In subtitle translation, the timing constraint can lead a translator to prefer a more parallel order to the original than would be the case in a standard written translation, thereby distorting the relations between propositions. In simultaneous interpretation, the working memory constraint can have a similar effect.

These findings suggest that there may be a trade-off between ease of production and preservation of meaning in translation or interpretation of complex sentences between languages with very different structure. Subtitle translation and simultaneous interpretation in such pairs seem to be characterized by more parallel order to the original and more manageable structure in the output than legal translation. But that comparative ease may come at the price of greater changes to the hierarchical relations between propositions, meaning greater potential distortion of the original message.

5.1.5 Individual variation

We’ve examined the observed and predicted effects of three input variables – structural difference in a language pair, sentence complexity and mode of language transfer – on rates for indicators of difficulty in translation or interpretation. There is of course another set of factors which can have a major effect on such indicators: individual differences between translators or interpreters, who might have different levels of experience or different personal styles.

This study tries to limit the potential effect of such individual differences in a number of ways. First, to the extent possible, it considers translation and interpretation produced by several different people under similar conditions in each mode of transfer. The three legal texts were translated into each target language by three different people. For each of those texts, the translated versions were all produced by trusted professionals working under the same instructions. Likewise, the five TED talks were subtitled in each target language by five different people. For each of those talks, the subtitled versions were all produced and revised by volunteers working under the same instructions.

For reasons of availability, it wasn’t possible to ensure the same degree of variety for simultaneous interpretation. It was hard enough to get an interpreted version of the same auhentic speech in the five different languages of the study. Getting versions of several authentic speeches produced by different interpreters in all those languages was simply impossible. Two of the interpreted versions of the speech were produced by experienced UN staff interpreters working under similar conditions. The other three interpreted versions were all produced by experienced interpreters who are accredited by the EU or are members of the International Association of Conference Interpreters (AIIC). Both of those organizations apply strict conditions of training, quality and experience. All five interpreters were given a written copy of the speech shortly before starting to interpret, but weren’t looking at a written translation while working.

Also, as discussed in section 2.1.3, the product-oriented analysis chosen for this study doesn’t measure physiological indications of individual effort. Instead, it records the products of manipulations that we have reason to believe are, in general, inherently difficult. One individual may have a harder time than another in performing a particular translation or interpretation task, depending on differences in experience or in what they’re used to. A translator or interpreter who works between English and Japanese may be more used to reordering entire propositions and parts of nested structures than one who works between English and Russian. There may well be a difference in inherent difficulty between the task of the Japanese translator or interpreter and the task of their Russian colleague. But even if there is, that objective difference may not correspond directly to a difference in difficulty as perceived subjectively by each one, or as reflected in individual measurements of their physiological responses while working.

Two of the output values considered in this study involve the products of mental tasks that have been empirically shown to be difficult – reordering (Donolato et al. 2017; Vanroy 2021) and manipulation of nested structures (Hawkins 2000; Karlsson 2007). The third output value involves the product of an alteration which language professionals are advised to avoid so as not to distort meaning – changes in semantic relations (Larson 1984). A product-oriented analysis has its limitations in terms of reflecting actual cognitive effort, as discussed in section 2.1.3. But that’s an advantage for this study, which analyzes the effects of various factors on general indicators of objective difficulty, regardless of individual experience or perception. Yes, the analysis is still carried out on the product of individual effort. But in focusing on product rather than process, it minimizes the effects of differences in real or perceived effort due to individual experience. That leads to more objective results on the comparative difficulty of translation and interpretation tasks in response to the various factors considered.

This study has achieved what it set out to do. It has established preliminary associations between three input variables– structural difference in a language pair, sentence complexity and mode of language transfer – on the one hand and observed and predicted rates for indicators of difficulty in translation or interpretation on the other hand. The study has also given an initial indication of particular types of difficulty associated with translation or interpretation between languages with very different structure, which seem less likely to characterize transfer between languages with similar structure.

For written translation, the main take-away from these findings is awareness – awareness of the difficulty, its reasons, its magnitude, the problems it can create, and therefore the need for professionals working between languages with very different structure to take extra time and care in production.

All this applies to simultaneous interpretation as well, plus the great added challenge of the natural constraint on working memory. Though not much can be done to reduce that added challenge, interpreters working in language pairs with very different structure can get better at coping with it. They can do so through practice and by applying various tactics when interpreting complex sentences. Those tactics are discussed in the next section.