Positive patterns presented in this subsection are grouped according to the marker mainly subject to biased translation, i.e., an evidential verb or a modal auxiliary. Modal clusters as well as framing structures are examined separately.
3.1.1. Patterns Built around Evidential Verbs
The corpus was first searched for all tokens with a POS-tag corresponding to a verb form (“VV.*”). A total of 12,647 tokens were retrieved by the search, corresponding to 1449 different word forms. The list of verb word forms was then screened for evidential verbs, and 14 potential candidates were found (appear, confirm, demonstrate, favour, find, indicate, note, observe, reveal, see, seem, show, suggest, support
). After exclusion of negative patterns through specific markers, results of corpus searches were manually refined in order to retain only occurrences of these evidential verbs within positive lexico-grammatical patterns signaling potential beneficial effects of interventions (see Section 2
, patterns a, b, and d). Verbs occurring less than five times within such a pattern were excluded from further study (7 in total). For the remaining seven evidential verbs, aligned concordances were examined for occurrences of biased machine translation, and the corresponding post-edited versions were checked for modifications. Results are presented in Table 2
Most of these evidential verbs are machine-translated by their closest equivalents in the target language which presumably correspond to similar levels of (un)certainty, such as sembler
, or suggérer
, and thus can be considered functionally equivalent translations. Two evidential verbs, find
, which both represent factual and neutral choices in the source language, are however more frequently subject to shifts in MT output towards more affirmative, biased lexical choices implying a higher level of proof, as discussed in Section 1.4
. In the case of find
, neutral translation options are represented for instance by the verbs constater
, whereas the typical biased translation is révéler
(see Example 1 below). In MT output, 13/55 (~24%) occurrences of find
in the context of positive lexico-grammatical patterns are subject to biased translation. (For all the examples presented here, the contextual window necessary for the MT engine to produce the output that appeared in the actual production context is quoted in total, since NMT is subject to important variation with even minor changes to the segment. A simplified back-translation is given to explain the origin of the bias.)
Biased translation of find within a positive pattern.
One trial found large benefits for SBP and DBP (SBP…).
MT: Un essai a révélé des avantages importants pour la SBP et la DBP (SBP…).
[= trial REVEALED important advantages for outcomes]
As discussed in Section 1.4
., for the evidential verb show
, a neutral and functionally equivalent translation choice is montrer
, while the typical biased translation implying a greater degree of certainty is démontrer
(see Example 2 below). In the corpus sample, show
is subject to biased machine translation in 12/55 (~22%) of its occurrences in the context of a positive lexico-grammatical pattern signaling potential benefits of interventions.
Biased translation of show within a positive pattern.
Data on dropouts from two studies with 95 participants showed a clear advantage for couple therapy (RR 0.31…).
MT: Les données sur les abandons de deux études menées auprès de 95 participants ontdémontré un net avantage pour la thérapie de couple (RR 0,31…).
[= data DEMONSTRATED/PROVED a distinct advantage for the intervention]
The three biased translation choices (démontrer, (se) révéler, s’avérer
) observed in the context of positive patterns represent approximately 19% (28/146) of total occurrences of show
in the corpus sample. For comparison purposes, on a larger corpus of Cochrane abstracts compiled for the aforementioned PhD project [8
], those three translation choices represent 25% of total occurrences of show
in human-translated texts, versus 12% in statistical MT post-edited by professionals. This would seem to suggest that, in terms of evidential modal verbs, neural MT output tends more towards human translation than to statistical MT.
Finally, as can be seen in Table 2
, it appears that student post-editors tended to accept the formulations suggested by the MT engine as such without any modifications. Indeed, for evidential verbs within positive lexico-grammatical patterns, the biased translation suggested by DeepL was edited in only one single instance, where the verb show
pre-translated by démontrer
was replaced in the post-editing stage by the more neutral montrer
3.1.2. Patterns Built around Modal Auxiliaries
A total of 894 occurrences of modal auxiliaries were retrieved by the search (POS-tag “MD”) and correspond to eight different modal verbs (in order of frequency in the corpus sample, may, can, could, should, might, would, will, must
). Four of them mainly occur in epistemic use in the context of the positive lexico-grammatical pattern under study related to effectiveness of interventions (see Section 2
, pattern c), wherein their function is to signal different degrees of certainty. The total frequencies of these auxiliary verbs in the corpus sample, occurrences within positive lexico-grammatical patterns, and occurrences of biased translation among the latter are presented in Table 3
The auxiliaries can and could are used respectively within these patterns to signal capacity and potential capacity of interventions to obtain a desired effect. Both are translated in all occurrences by their closest direct equivalent, the verb pouvoir in the indicative mood for can and in the conditional mood for could, and are therefore not subject to bias in translation. For the auxiliary may, which is typically used as a hedging device for mitigation of claims, biased machine-translation by the indicative mood of the verb pouvoir (see 1.4. and Example 3 below) accounts for 142/155 (~92%) of its occurrences within a positive lexico-grammatical pattern signaling potential effectiveness of interventions.
Biased translation of may within a positive pattern.
Tocolysis may improve blood flow and therefore improve the baby’s well-being.
MT: La tocolyse peut améliorer la circulation sanguine et doncle bien-être du bébé.
[= intervention MAY/CAN improve outcomes]
Translation by the indicative mood of the verb pouvoir represents approximately 89% (348/391) of all occurrences of may in the corpus sample, while the remaining 11% mainly correspond to the use of the conditional mood. Again, for comparison, on the aforementioned larger corpus of Cochrane abstracts, the indicative mood was present in 54% of human-translated texts versus 45% of texts pre-translated by a statistical engine and post-edited by professionals. This seems to confirm the previous observation on neural MT tending towards human-like output in the translation of modal markers—perhaps even to the point of exaggerating typically human characteristics. Finally, for the modal auxiliary might, used for signaling higher uncertainty than may, the neutral and functionally equivalent translation choice, the verb pouvoir in the conditional mood, is chosen in most cases, biased translation by the indicative mood representing 15% (3/20) of the occurrences of might within a positive lexico-grammatical pattern.
Again, in most cases, the biased translation suggested by the MT engine for modal auxiliaries within positive patterns was accepted as such by the student post-editors, even when the shift in the level of certainty was as marked as for the auxiliary might
translated by the indicative mood (see Table 3
). Generally speaking, students tended to modify more frequently the suggested translation for modal auxiliaries than for evidential verbs in the MT output. Indeed, on a total of seven occurrences in the corpus sample, students edited MT output for the auxiliary may
from the indicative mood into the conditional mood. In one instance, however, a neutral and functionally equivalent translation of may
in the MT output was changed into a biased one by the post-editor (i.e., conditional to indicative mood), while in three other instances, the bias in translation was further reinforced at the post-editing stage through other choices (see Example 4 below).
Bias reinforced at the post-editing stage.
Acupuncture plus routine primary physician care may improve pain and function compared to routine primary physician care alone.
MT: L’acupuncture et les soins médicaux primaires de routine peuvent améliorer la douleur et le fonctionnement comparativement aux soins médicaux primaires de routine seuls.
PE: L’acupuncture combinée à des visites de routine chez le médecin peuvent se révéler plus efficaces pour réduire la douleur et améliorer le fonctionnement de la hanche que des visites seules.
[= intervention MAY/CAN improve outcomes vs MAY/CAN prove to be more effective]
In the example above, the student post-editor has added a positively biased evidential verb (se révéler) which reinforces the bias originating from the use of the indicative mood in the translation of the modal auxiliary may.
3.1.3. Modal Clusters
Modal clusters combine different markers, for instance evidential verbs, modal auxiliaries, and adjectives or adverbs, within a larger lexico-grammatical pattern for more precise signaling of the level of proof. Such clusters are occasionally subject to biased translation of the different markers in the MT output (see Example 5a,b below).
Biased translation of a modal cluster within a positive lexico-grammatical pattern.
The review of available evidence found that combinations of antipsychotics may be more effective (…)
MT: L’examen des données probantes disponibles a révélé que les associations d’antipsychotiques peuvent être plus efficaces (…)
[= review REVEALED that intervention MAY/CAN be more effective]
Contributing to the distortion originating from more affirmative lexical choices for modal markers, another biased translation observed in the MT output in the context of these clusters is the omission of hedges, such as the nominal group some evidence in Example 5b below.
Biased translation of an evidential verb reinforced by omission.
One well-controlled study shows some evidence of effect of two interventions for childhood apraxia of speech (…)
MT: Une étude bien contrôlée démontre l’effet de deux interventions sur l’apraxie de la parole chez l’enfant (…)
[= study DEMONSTRATES/PROVES the effect of interventions]
In previous research [8
], biased translation of modal clusters and specifically the elimination of hedges, which leads to markedly more affirmative statements, was typically observed in human translation. Although only a few occurrences (3/103) were observed in the present corpus sample, this again contributes to the previous observations on too convincing mimicking of human translation by neural MT as a potential source of distortion in the communication of specialized knowledge.
3.1.4. Existential Framing Structures
The existential framing structure (see Section 2
, pattern e) has a total of 43 occurrences in the corpus sample, 22 of which correspond to positive lexico-grammatical patterns signaling potential benefits of interventions. Although only three instances of biased translation of the existential structure framing positive patterns are observed, it can be noted that they all correspond to markedly more affirmative statements in the MT output than in the source text (see Example 6a,b below). The biased translation combines the omission of a nominal group built around evidence
, as already observed above for modal clusters, with the adding of an evidential verb signaling a high level of proof such as prouver
Biased translation of an existential framing pattern.
There is evidence that burns treated with honey heal more quickly (…)
MT: Il est prouvé que les brûlures traitées au miel guérissent plus rapidement (…)
[= it is PROVEN that intervention is more effective]
Biased translation of an existential framing pattern.
There was some evidence that trihexyphenidyl may improve individual goals set by the child and family (…)
MT: Il a été démontré que le trihexyphénidyle peut améliorer les objectifs individuels fixés par l’enfant et sa famille (…)
[= it has been DEMONSTRATED/PROVEN that intervention MAY/CAN improve outcome]
In accordance with the previous observations, this very human-like translation combining omission and addition—and resembling something like biased interpretation—could potentially lead to important distortion in the interpretation of research results and thus, the communication of specialized knowledge.