Navigation – Plan du site

AccueilNuméros38PerspectivesAssessing human translation style...

Perspectives

Assessing human translation style with the help of NMT: a case study of French-language comic book translators

Bartholomew Hulley

Résumés

Cet article présente une méthode novatrice pour évaluer la traduction humaine en utilisant la traduction automatique neuronale (TAN) comme référence. En abordant les défis liés à l’étude des styles de traduction, la méthode offre des analyses impartiales de multiples traducteurs. Une étude de cas sur les traducteurs de bandes dessinées révèle son potentiel et fournit de nouvelles perspectives sur les pratiques de traduction des bandes dessinées en Amérique du Nord.

Haut de page

Texte intégral

  • 1 WER = Word Error Rate; TER = Translation Error Rate; HTER = Human-targeted Translation Error Rate (...)
  • 2 For example: keeping the syntactical arrangement of the ST, maintaining the ST’s structure and le (...)

1So far, texts output by machine translation engines have been evaluated by comparing them with “gold standard” human translations. This practice has led to the formulation of numerous metrics that purport to reflect the accuracy of the machine-translated text, by quantifying the number of observed disparities. Acronyms such as WER, TER, HTER, BLEU and NIST1 have therefore become familiar to those researching these rapidly evolving tools as metrics providing insight into translation engine behaviour for any given language pair. In this article, I propose a similar kind of metric conceived to reflect human translation output using the very same tools—but in reverse. That is, instead of using a human translation to measure a machine translation, to do precisely the opposite and use a machine translation to measure a human generated target text (TT). However, rather than supply a measure of the translator’s accuracy, I propose that this approach offers a way to quickly profile multiple professional literary translators according to the degree of formalism in their work—the extent to which they preserve the form of a source text (ST) in their translations2—for the benefit of publishers and editors, while also overcoming the methodological impasse encountered in translator stylistics.

1. The Enigma of Translator Style

2As Baker (2000), Saldanha (2011) and Munday (2016) have shown, research into translator style has been plagued by methodological problems. While scholars might agree with Baker’s definition of “style” as a kind of ‘thumb-print’ that embodies the factors that contribute to the translator’s voice in the TT (their manner of expression, typical use of language, and other linguistic habits), there appears to be little consensus as to how the researcher might identify decisions that are attributable to straightforward formal equivalence versus ones that are down to the translator’s individual preferences and disposition, not to mention decisions which might be ascribed to an editor. One approach, suggested by both Baker and Munday, is to discursively study surface aspects of the TT that call out for attention by their strangeness, repeated use, or inconsistency. Saldanha, however, points out that investigations of this kind risk compromising the results, because the very act of identifying notable features is necessarily based upon the expectations of the researcher. Drawing conclusions from selected linguistic or lexical features also risks overlooking any number of variables that might contribute to or explain what is observed. Indeed, she maintains that the linguistic form of a TT should not be regarded as equal to translator style, because the function of the whole text has a role to play. Interpreting a habit as evidence of a conscious, meaningful decision by the translator is unlikely to be reliable without accounting for their intentions or subconscious reflexes which, as Munday observes, can only be achieved by analysing the translators themselves and/or observing them during the act of translation.

3To add to the problem, observations made in any scientific study will always be relative to both the observer and a fixed point of reference chosen by that observer. Historically, translation studies seeking to understand, or model, any number of translation decisions have therefore usually sought to draw comparison between a target text and either i) the original source text, or ii) another target text (an earlier translation, for example). The rationale being that such comparison will reveal certain characteristics that are relatively unique to the text being studied. What is problematic about this approach is that neither i) nor ii), nor the text being studied, can be judged altogether objectively due to the fixed point of reference being the researcher themselves. That is, each observation will be influenced by the scholar’s values, knowledge, and experience unless they can find a way to measure each text without personal intervention.

1.1. Automated, formally equivalent, reference translations

  • 1 It should be noted that NMT output is sensitive to the full input string, therefore any observed (...)

4Evidently, what is required is a reference translation (RT) for each source text (ST) to facilitate the evaluation of its corresponding TT. However, in order for the resulting data to give any indication of thumb-print, to borrow Baker’s term, this reference translation needs to: exhibit zero creativity and inventiveness, lack preference for any particular vocabulary or syntax, consistently tackle the same linguistic problems in exactly the same manner, be ignorant of prior works by both the author and translator, and overlook any extra-linguistic aspects such as the purpose or function envisaged by the author. Furthermore, to facilitate the analysis of multiple translations and multiple translators, whoever supplies this reference translation will need to deliver it in exactly the same manner for every ST in any given corpus. Clearly, the only person, or thing, able to supply a consistently impervious approach to multiple tasks is a machine. Now, considering that Neural Machine Translation (NMT) engines translate according to paradigms acquired through training, with complete disregard for any extra-linguistic context (such as social space, style, purpose, and function), and resolutely tackle every source text “string” in precisely the same manner, they thereby offer a baseline from which to evaluate one or many human generated TTs, or one or many translators. Therefore, given that any proprietary trained NMT engine will translate identical input strings, at the same moment in time, in exactly the same manner,1 it would seem that NMT is the ideal candidate to supply that reference translation.

5Current NMT systems are comprised of ethereal “neurons” possessing multi-dimensional profiles of links to other neurons within a neural network. Each neuron itself might represent a single character, a string of characters, a part-word, a word, or word-sequence that will correlate with a neuron or network of neurons in the target language. For example, the figure 2 in a French ST would possess links to 2, two, pair, couple, to, or too in English. The level of activation, or strength, of each link between each neuron is established through “training” which involves automatic comparison of MT outputs with referential texts, such as “gold-standard” human-translated parallel corpora at word, phrase, paragraph, and even complete text levels (Forcada, 2017: 295). What makes modern NMT a breed apart from its predecessors is its inherent modelling of textual patterns. That is, any word in any grammatical position is assigned a probability that it will be followed by, or preceded by, a certain other word, or group of words. These probabilities are established through training with vast monolingual corpora. Thus, the advent of neural machine translation has coincided with a leap-forward in output fluency, and confidence in output is now so high that machine translation “post-editing” (MTPE) has become widespread. However, as noted by Kenny (2022), high output fluency does not necessarily equate to a highly accurate translation, with omissions and semantic errors often appearing wherever proprietary NMT services lack the necessary training to deal with an uncommon input sequence.

  • 2 See Guerberof-Arenas and Toral, 2020.

6Such flaws belie the nature of machine translation itself. Although NMT is characterised as using “Artificial Intelligence” (AI) there is nothing “intelligent” about it. AI is not a sentient, thinking being which means its translations are based on algorithms rather than consideration, voice, and creativity. Output is therefore optimised for the most common equivalent sequence in the target language—what Eugene Nida (1964) might refer to as formal equivalence—and exhibit many of the linguistic features of the ST. This is what makes machine translation particularly suited to technical translation tasks.2

7All this being said, the purpose of the RT in our methodology is to provide a consistent “baseline” translation across all texts in order to model the TT output by the human translators, therefore its perceived quality is of marginal concern. Moreover, any attempt to introduce translation quality assessment (TQA) into the process will simply compromise objectivity; as Pym rightly points out: “all evaluations of quality are based on human values […] and are built on a fundamental indeterminacy” (Pym, 2020: 448). To be clear, should a RT contain occasional inaccuracies, provided the studied text is of sufficient length, these would be statistically unlikely to have much influence over the collected data.

8At the opposite end of Nida’s binomial theory lies dynamic equivalence, which arguably still remains the preserve of human translators. Although NMT can be trained for specific functions, as yet it cannot take into account the extra-linguistic contexts of a ST, nor emulate the same wide range of creative approaches as an experienced professional literary translator is capable of.

1.2. Approximate string matching

  • 3 See Thomas H. Cormen (ed.), Introduction to Algorithms, MIT Press, 2001.

9In order to measure the studied TTs relative to their corresponding baseline RTs, I submit that a straightforward approximate “fuzzy” string matching algorithm3 will supply a figure that represents the similarity, in percentage terms, of each TT to its RT counterpart. This figure therefore would represent a measure of the approximate linguistic similarity between the two texts. The theoretical polar extremes would be: 100 % for a TT that exhibits machine-like formal equivalence and 0 % representing total dynamic equivalence.

  • 4 Statistically, shorter strings of characters are susceptible to return figures proportionately cl (...)

10Of course, depending on the length of the segment of text,4 high originality, or adaptation, may also equate to high dissimilarity, therefore instances of low figures in the results will need human verification to ensure the accuracy of the human TT. It is for this reason that the researcher cannot overlook the aspect of quality entirely. Certainly, the TTs will need to exhibit a certain degree of accuracy, as there would be little scholarly value in investigating the style of erroneous or amateur translations.

11Admittedly, the figure output by this methodology will not provide much insight of itself, nor will it directly reflect any preconceived notion of translation quality or style. In other words, any observed similarity between a human TT with its corresponding RT cannot be reasonably judged by statistics alone. This is why a previously studied human translation is required to facilitate the calibration of the collected data, to situate each studied text and translator relative to a known “control” translation or translator. That is, only by including a control text or translator that has been previously discursively studied in detail in their corpus, can the researcher judge the relative “style” of each subsequently studied translation or translator. For example, if a translator exhibits less formal equivalence in their texts compared to a control translator known for producing highly formal TTs, it can be deduced that they approach their texts with a greater degree of dynamic equivalence. Of course, multiple translations will be required to ascertain a translator’s average profile, or thumb-print, and several translators will need to be studied to ascertain where each is situated relative to each one of their peers.

2. Case Study: Translators of comics

12The lack of research into the vast and varied world of translated comics and graphic novels is a common lament among scholars of text-image translation; and, given the growing interest in the field, it is a complaint not without grounds. Even today, despite efforts to highlight this deficiency by Kaindl (1999) and Zanettin (2008), the greater part of research remains based upon the Astérix and Tintin series. The predominance of studies into the translations by Bell and Hockridge (Astérix) or Lonsdale-Cooper and Turner (Tintin) for the British market also draws attention to the lack of erudite knowledge about translations for the North American market, particularly since neither series enjoyed much commercial success in the USA. Indeed, not since Franco-Belgian bande dessinée first broke into the mainstream transatlantic book market has there been any indication as to whether comics translators employed strategies comparable to those used for British readers. Anecdotally however, the late Kim Thompson, translator and editor at Fantagraphics USA, believed that “anyone who’s ever read the English translations of Asterix, realizes that those are the gold standard” (Thompson et al., 2010). However, no investigations into the prolific translation work of Thompson currently exist, so it is impossible to say whether he, or indeed any other translators to the North American market, managed to emulate this coveted gold standard.

13My objective then, was to ascertain which US-based translators of bande dessinée, if any, had managed to match Thompson’s perceived gold standard, by measuring their average degree of formalism using the previously described methodology and comparing it to the figure obtained by the translators of Asterix using the same method.

  • 1 Patrick O’Beirne,. “VBA Code for Microsoft Text Translator API”, 20 January 2017. https://sysmod. (...)
  • 2 Sorin Sion, 2011. “Fuzzy Matching Scoring Algorithm.” https://forum.ozgrid.com/forum/index.php?th (...)
  • 3 I ran the same scripts on the same texts in October 2022 and the results were only marginally dif (...)

14Thus, I selected a corpus of nine translated bandes dessinées that have won either a Harvey or an Eisner comic-book award for best “foreign” material, as voted for by the comic book industry in North America, thereby assuring perceived TT quality. I then entered the text from the first 100 frames of both the ST and TT into an Excel spreadsheet to compare with a RT, facilitated by open-source visual basic (VBA) code for Microsoft’s publicly accessible general-purpose NMT.1 I then automated the analysis using further open-source VBA modules to return fuzzy matching2 scores to quantify each translation. I ran these scripts within a 3-hour window on 11 February 2019 in order to ensure the results would not be affected by any scheduled update to Microsoft’s service.3 Once the statistical analysis was complete, I created distribution histograms to help visualise each translator’s profile and then extracted example texts from the data set to check the style of translation.

  • 4 See Rothwell, 2023.

15It is worth noting that comic books are particularly well-suited to a segmented analysis, thanks to the familiar speech balloons and caption boxes which enable the researcher to clearly delineate one enunciation from the next. Prose texts, by contrast, would require a specific segmentation strategy reminiscent of that used in computer aided translation (CAT) tools like CafeTran, MemoQ, and Trados.4

2.1. The “Gold Standard” control

Figure 1. distribution of segments (y) by RT similarity (x) for Bell & Hockridge.

Figure 1. distribution of segments (y) by RT similarity (x) for Bell & Hockridge.

Given the objective of this study is to ascertain which North American translators approached their award-winning texts in a similar fashion to Bell and Hockridge, I have selected their translation of Astérix et les Normands as the control. This particular title represents a time when Bell and Hockridge’s approach to translating comics had fully matured, it being their 20th translated album from the series. It is a text that has been examined in depth by Delesse and Richet in Le Coq Gaulois à l’heure anglaise (2009), perhaps the most comprehensive study into the mechanisms of comics translation ever undertaken. In short, it is a TT packed with creative solutions to onomastic problems, puns, cultural references, and oral markers. To which end, it is characteristic of a dynamically equivalent translation that prioritises the “spirit” of the ST, with regular instances of adaptation, addition, and compensation.

  • 5 The height of the ten “bins” in the histogram indicates the number of TT segments falling within (...)

16This observation is backed up by the resulting histogram pictured in Figure 2-1 which groups the TT segments by their similarity to the RT.5 It suggests a TT that required frequent deviation from a formal approach, with 98 out of 187 segments (52 % of the total) exhibiting less than 40 % similarity to the RT. This implies that Goscinny’s text presented the translators with a variety of problems which could only be solved by target segments that linguistically diverged from the ST. Much of the complexity, no doubt, can be attributed to Goscinny’s love of puns, wordplay and jokes, that are a priori untranslatable.

Table 1. Selected translations from Astérix et les Normands.

Segment

Page/Panel

Character

ST Goscinny

RT
11 Feb 2020

TT Bell & Hockridge

% match

a

1/5

Astérix

Rien de grave, dans ce qui est gravé, j’espère ?

Nothing serious, in what is engraved, I hope?

Nothing grave engraved there, I hope?

48%

b

7/2

Obélix

Pas du tout ! Dans le midi de la Gaule, il pleut. Ici, c’est tout juste un peu humide. Vivifiant. Pas vrai, Astérix ?

Not at all! In the south of Gaul, it rains. Here, it’s just a bit damp. Invigorating. Right, Asterix?

It might be rain in the south… here it’s just a little bracing dampness in the air…

43%

c

4/11

Assurance-tourix

Vous croyez ? À Lutèce ?

Do you think so? Lutetia?

You think I’d make a hit in Lutetia?

40%

d

8/8

Obélix

Si on ne peut pas en profiter, ce n’est pas la peine de vivre au bord de la mer !

If you can’t enjoy it, you don’t have to live by the sea!

What’s the good of living by the seaside if you never get any fun out of it?

25%

Table 1 shows some examples characteristic of the approach taken by Bell and Hockridge for Asterix and the Normans. Clearly, the machine-generated RT is unable to account for voice or pictorial context, as well as being devoid of all notions of creativity, so offers a text that is firmly ST orientated, exhibiting classical formal equivalence. Comparing this reference text to the more creative TT by Bell and Hockridge demonstrates how a few, arguably subtle, changes can make the TT more fluid and therefore pleasurable to read. Evidence perhaps that a formal approach to translating comics literature is inappropriate, certainly in this instance.

17In segment a (page 1, panel 5) the reference text does not recognise the polysemy of “grave,” and its use in Goscinny’s wordplay, instead outputting “serious” as the most probable “best” translation, according to its training corpus. This demonstrates how a rigid formal approach can lead to the destruction of the humour and spirit of the original. In segment b (page 7, panel 2) the five short sentences of the ST are supplanted by a single enunciation in the TT that expresses the same sentiment but with a sense of character voice, despite sharing little in common with the syntax and vocabulary of the RT. The TT of segment c (page 4, panel 11) merges the two interrogative phrases of the ST to overcome the French juxtaposition. In segment d (page 8, panel 8) Bell and Hockridge have deviated even further from the ST but respected the sentiment expressed by the character, Obélix. This embodies their ear for natural dialogue and voice which is apparent throughout the TT. Not only do they find fluency within the confines of each speech balloon (which were constrained by hand-lettering in 1978) but also manage to maintain semantic proximity to the ST in the spirit of the original.

18On average, Bell and Hockridge’s translation exhibits 43 % similarity to the baseline reference translation. Theoretically, therefore, this figure represents a dynamic, creative approach to the task of translation.

2.2. Jean-Marc and Randy Lofficier

Figure 2. distribution of segments (y) by RT similarity (x) for JM&R Lofficier.

Figure 2. distribution of segments (y) by RT similarity (x) for JM&R Lofficier.
  • 6 For a full list of winners see https://www.harveyawards.com/en-us/winners/previous-winners.html

The Lofficiers have won three awards for their translations of titles illustrated by the late Jean “Moebius” Giraud. In 1988 their translation of Sur l’étoile (Upon a Star), a surreal dream-like, science fiction tale originally conceived for the Citroën motor company, was the inaugural winner of the prize for Best American Edition of Foreign Material at the Harvey Awards6 for which their TT exhibits a 47 % similarity with a baseline RT. In 1989, their translation of l’Incal (The Incal), a dystopian science-fiction saga written by Alejandro Jodorowsky, won the Harvey Award for Best American Edition of Foreign Material with a TT 52 % similar to a RT. Then in 1991, the Lofficiers won the Harvey Award for Best American Edition of Foreign Material for the Blueberry series, set in the American wild west and written by Jean-Michel Charlier, with their translation of La Mine de l’Allemand perdu (The Lost Dutchman’s Mine) approaching 41 % similarity to a RT.

19The distribution histogram (Figure 2) of their combined translations reflects a broadly similar degree of departure from formal equivalence as demonstrated by Bell and Hockridge. Both charts peak in the 30-40 % range which suggests frequent departure in order to optimise many TT segments for the target reader.

Table 2. Selected translations from L’Incal noir.

  • 7 Curiously, despite Microsoft’s NMT cloud service censoring numerous other offensive terms, it dem (...)

Segment

Page/Panel

Character

ST Jodorowsky

RT

11 Feb 2020

TT Jean-Marc & Randy Lofficier

% match

a

15/2

Difool

Hé hé !!. Faut croire que ma cachette était pas si mauvaise !..

Hey hey!! I guess my hiding place wasn’t that bad!..

I guess my hiding place wasn’t so bad after all!

77%

b

6/3

Nimbéa

C’est vous John Difool, je vous attends dans une heure… pour traiter une affaire importante… hm… espace 771643

It’s you Difool John, I’ll wait for you in an hour... to deal with an important case... Hm... space 771643

Are you John Difool? I want to see you in an hour. An important matter… I’m in conapt 771643…

58%

c

13/1

Caption

Jamais je ne m’étais senti si mal dans toute ma putain de vie ! Cette lueur m’avait transpercé le crâne de part en part ! Et c’était pas tout ! Pendant ce temps là, l’épinglé s’était mis à se décomposer à toute allure !.

I’ve never felt so bad in my whole fucking life! This glow had pierced my skull from side to side! And that wasn’t all! Meanwhile, the pinned had started to decompose at full speed!7

I’d never felt so rotten in my entire, stinking life! That light had gone right through my skull! And if that wasn’t enough, the thing was now melting like cheese on a pizza!

38%

d

14/3

Difool

Merde ! Qu’est ce que c’est encore ?.

Damn!.. What is it yet?..

Shiiit! Who the hell is that?

21%

Table 2 shows selected segments from the Lofficiers’ translation of L’Incal noir from 1988 to understand their typical approach for a range of similarities to a RT. The first thing of note is segment a, on page 15 in panel 2, which lacks a translation for Difool’s exclamation (Hé hé !!) from the ST. Rather than presume this as an oversight by the Lofficiers, I would suggest that it is an intentional omission made to reinforce the overall coherence of the voice of the character. For segment b, on page 6 in panel 3, the Lofficiers have decided that a strictly formal equivalent translation would result in a strange enunciation in the target, as demonstrated by the RT. Their solution identifies an interrogative tone at the beginning of the segment as well as crafting a coherent enunciation in the context of the narrative, using the appropriate register, and reappropriating the author’s neologism “conapt” to refer to the location of the meeting. The caption segment (c), in panel 1 on page 13, carries Difool’s inner speech and therefore warrants treatment as dialogue. To this end, they use a simile at the end of the enunciation (melting like cheese on a pizza) which corresponds directly with the picture in the panel that follows it. Segment d, in panel 3 on page 14, is an example of adaptation where the translators have decided to offer some precision over what the character is referring to (the person knocking at his door) as well as adding orality by extending the phoneme in Difool’s expletive: Shiiit!

20The result of the Lofficiers approach are texts that display a great deal of coherence in which character voices are distinct and believable. Furthermore, as a bilingual couple, with Jean-Marc being French and Randy being American, they are able to demonstrate semantic and functional proximity to the ST without compromising on the fluency of the TT. In short, the Lofficiers deliver texts that prioritise the needs of the target reader. Across all three nominated texts, their approach averages out at 47 % which reflects a largely balanced approach with a marginal preference for dynamic equivalence.

2.3. Alexis Siegel

Figure 3. distribution of segments (y) by RT similarity (x) for Siegel.

Figure 3. distribution of segments (y) by RT similarity (x) for Siegel.

Alexis Siegel has also won three awards for his translations. In 2006, his TT of Le Chat du Rabbin (The Rabbi’s Cat), by Joann Sfar, won the Eisner Award for Best U.S. Edition of Foreign Material. Published by Pantheon and crediting Anjali Singh as both co-translator and editor, it tells the story of a rabbi and his talking cat exploring themes of religion, identity, and cultural coexistence in 1920s Algeria. Siegel and Singh’s translation is 67 % similar to a baseline RT. In 2009, Siegel won the Harvey Award for Best American Edition of Foreign Material with humorous cowboy western Gus and His Gang, written and illustrated by Christophe Blain, simply Gus in French, his translation presenting a 56 % similarity with a baseline RT. Then in 2010 he won the Eisner Award for Best U.S. Edition of International Material with his translation of Didier Lefèvre’s biography Le Photographe (The Photographer), written and illustrated by Emmanuel Guibert, that recounts the photojournalist’s time with Médecins Sans Frontières during the Soviet Afghan War. The translation is 57 % similar to a baseline RT.

21Compared with the histograms observed above, the segment distribution for Siegel’s combined translations immediately suggests a different approach (Figure 3). A large disparity between the 20-30 % and 30-40 % ranges perhaps suggests that while Siegel is prepared to adapt his TT to the needs of the target reader, there is a limit to how far he is prepared to deviate from the linguistic form of the ST. The histogram demonstrates a consistent preference for more a formal approach than that used by Bell and Hockridge, with 63 % of segments demonstrating a RT similarity of 50 % or more.

Table 3. Selected translations from Gus.

Segment

Page/Panel

Character

ST Blain

RT

11 Feb 2020

TT Siegel

% match

a

1/5

Gus

Je l’ai connue il y a cinq ans, à Cincinnati.

I met her five years ago, in Cincinnati.

I knew her five years ago in Cincinnati.

82%

b

4/4

Nathalie

Oh oui ! Pardonnez-moi. J’ai été très malade, je n’ai pas quitté mon lit depuis une semaine…

Oh, yes! Please forgive me. I was very sick, I haven’t left my bed for a week...

Oh, yes! Forgive me, I’ve been very ill, I’ve been bedridden for over a week.

59%

c

3/6

Nathalie

C’est l’heure de retrouver Chet. Vous allez voir, c’est un homme exceptionnel. Beau, fort et sûr de lui. Il saura me protéger en toute occasion. J’ai une totale confiance en lui, je ne peux rien lui cacher.

It’s time to find Chet. You will see, he is an exceptional man. Beautiful, strong and self-confident. He’ll protect me at any opportunity. I have total confidence in him, I can’t hide anything from him.

It’s time to meet Chet. You’ll see he’s an extraordinary man. He’s handsome, strong, and very confident. He’ll be able to protect me under any circumstances. I trust him completely. I could never hide anything from him.

40%

d

7/1

Gus

Excusez ma tenue, Nathalie. Vous me prenez au dépourvu.

Excuse my outfit, Nathalie. You’re catching me off guard.

Pardon my appearance, Natalie, this was so unexpected.

26%

Table shows selected segments from Siegel’s translation of Gus to help get a picture of his typical approach for a range of similarities to a formally equivalent RT. The segment found in panel 5 on the first page (a) is particularly jarring from a reader’s perspective. Here, the RT has correctly identified target equivalence for “je l’ai connu” (I met her) while Siegel has clumsily retained the ST verb in the target text (I knew her) thereby, arguably, behaving more like a machine than the machine. The segments found on pages 3, 4 and 7 (b, c & d) however, demonstrate Seigel’s ability to craft a TT that is very close to the ST, that also reads fluidly in the target language while respecting the required register of the character’s voice.

22Although Siegel’s translation profile averages out at a similarity score of 60 %, which places his approach firmly in the domain of formal equivalence, he also demonstrates an ability to deviate where it is functionally expedient to do so, however he is unlikely to resort to adaptation.

2.4. Kim Thompson

Figure 4. distribution of segments (y) by RT similarity (x) for Thompson.

Figure 4. distribution of segments (y) by RT similarity (x) for Thompson.

Although Thompson has won several awards, his only prize for a translation from French was for C’était la guerre des tranchées (It Was the War of the Trenches), a series of stories set in the trenches of WWI written and illustrated by Jacques Tardi which won the 2011 Eisner award for Best U.S. Edition of International Material. Thompson’s translation is 45 % similar to the RT. Curiously, what is immediately noteworthy about Thompson’s text is that it is 3 % longer than the ST, whereas all the other TTs analysed in this case study are around 8 % shorter. This suggests that Thompson likes to use the additional space afforded by the conciseness of English to make his own additions, and indeed this supposition is supported by the observations below.

As can be seen from the histogram (Figure 4), what is notable is the relatively narrow distribution of segments for his approach, with the majority of the TT falling within the 30-50 % range, implying a high degree of departure from formal equivalence.

Table 4. Selected translations from It Was the War of the Trenches.

Segment

Page/Panel

Character

Tardi

RT

11 Feb 2020

Thompson

% match

a

15/2

Charroi

Binet ! Qu’est-ce que tu fiches ici ?

Binet! What are you doing here?

Binet what the fuck’re you doin’ here?

68%

b

9/1

Caption

Au second, les Magnin et leurs sales gosses insolents.

In the second, the Magnin and their dirty, cheeky kids.

On the third, the Magnins and their awful, obnoxious kids.

43%

c

14/4

French Soldier

Ta gueule ! …gros con !

Shut up! ... Big jerk!

Ah. Shut yer stinkin’ hole!

32%

d

20/2

German Soldier

Mon dieu !... C’est un gosse ! J’ai abattu un gosse habillé en soldat … mais comment a-t-il pu arriver jusqu’ici et se balader entre les lignes avec un fusil ?

My god!... He’s a kid! I shot a kid dressed as a soldier ... but how did he get here and walk between the lines with a gun?

Mensch, das ist ein Bengel! Ich habe ein als Soldat verleidetes Kind erschossen! Was machte der denn hier zwischen den Schützengraben?

24%

  • 8 I am unable to comment on Thompson’s use of German.

Table 4 shows selected segments from Thompson’s award-wining translation of C’était la guerre des tranchées by Tardi. There are few dialogues, so the majority of the narrative is carried by a single voice in the captions. In the first example (segment a), on page 15 in panel 2, despite reflecting a 68 % similarity Thompson has kept the low register of the ST and adapted the TT to the voice of the surprised soldier, involving two oral contractions: doin’ and fuck’re. For segment b, on page 9 in panel 1, he has identified the equivalent floor of the building for the target culture, as well as selecting adjectives that reflect the paradigmatic tone of the narrator. On page 14, in panel 4, (segment c) he has combined the sentiment of the two ST statements into a single phrase, using orality markers to adapt to the imagined voice of the character: an orthographic variant of your (yer) and a contraction (stinkin’). However, the most striking decision by Thompson is reflected by segment d in panel 2 on page 20, which is part of a short sequence written entirely in German. This is creative adaptation in its purest form. Rather than adapt the TT to reflect character voice in the target language, Thompson has opted to break with convention entirely and directly challenge the reader to either deduce what the character is saying or to seek out their own translation of the TT from this third language. At the time of publication (2010), online machine translation tools had already been in service for a number of years, so it is probable that Thompson was aware how easy it had become for motivated readers to translate short passages of text as desired. Although some would argue that this strategy is a highly unfaithful approach to take, it offers something that the ST did not: a more faithful rendering of the linguistic otherness across no-man’s land. In this sense it might be said that Thompson has succeeded in improving upon the original, or indeed has shown greater fidelity towards the “spirit” of the original.8

23Thompson’s approach can perhaps be characterised by a lack of concern for formal equivalence with the ST, but a “faithful” approach to the purpose and function of the text, which, for the reader, results in supremely coherent TTs. Furthermore, closer subsequent inspection also reveals that Thompson’s translations convey a certain fondness for creating voices and adapting the text accordingly.

2.5. Helge Dascher

Figure 5. distribution of segments (y) by RT similarity (x) for Dascher.

Figure 5. distribution of segments (y) by RT similarity (x) for Dascher.

In 2014, Dascher won the Eisner award for Best U.S. Edition of International Material with Putain de guerre! (Goddamn this War!), a historical account of WW I illustrated by Jacques Tardi and written by Jean-Pierre Varney, with a translation 41 % similar to the RT. It is worth noting that the texts found in the ST demonstrate very little semantic interplay with the pictures, the words often conveying a narrative below the illustrations. Unlike the other texts looked at in this case study the ST features no characters, no dialogue, and no speech balloons. Instead, its third-person narrative is contained entirely within captions.

24The unambiguously narrow distribution of segments shown in Figure 5 for Goddamn this War! is notable, with the vast majority of the TT falling within the 30-50 % range. The implicit departure from strict formal equivalence is therefore stark.

25Table 5 shows selected segments from Dascher’s 2013 translation of Putain de guerre! to help get a picture of her typical approach. For segment a, on page 3 in panel 2, although showing strong similarity to the RT, Dascher adds a contraction (would’ve) and an idiom (in their shoes) to make the TT reflect an oral paradigm. In segment b, on page 1 in panel 2, Dascher again emulates speech in the TT with the aid of an oral marker gonna although no equivalent marker exists in the ST. In panel 1, on page 6, segment c features an addition (Europe’s masters) and the removal of the pronoun. Perhaps Dascher felt this offered greater precision over which masters the narrator is referring to, given the intended target reader is North American. Segment d, on page 20 in panel 2, however, although the figure of 38 % suggests it exhibits some degree of adaptation (which the inclusion of the idiom scared shitless partially confirms) Dascher’s text still feels slightly awkward due to her respect for the negation in the ST. It would seem more natural to express the sentiment in a positive clause; for example: “We were scared shitless that we might die on the front line that very fucking day.”

Table 5. Selected translations from Goddamn this War!

Segment

Page/Panel

Character

ST Verney

RT

11 Feb 2020

TT Dascher

% match

a

3/2

Caption

C’étaient des prisonniers et j’aurais donné beaucoup pour être à leur place. Pour eux, la guerre était finie.

They were prisoners and I would have given a lot to be in their place. For them, the war was over.

They were prisoners, and I would’ve given a lot to be in their shoes. Their war was over.

76%

b

1/2

Caption

Mais cette fois, nous étions prêts, on allait leur faire bouffer leurs casques à pointe en cuir bouilli !

But this time we were ready, we were going to make them eat their helmets with boiled leather tip!

But this time we were ready. We were gonna stuff their boiled leather helmets down their throats, spikes and all.

58%

c

6/1

Caption

C’est qu’il en fallait de la viande humaine pour rassasier les appétits insatiables de nos maitres !!

It was that it needed human meat to satisfy the insatiable appetites of our masters!

So much human meat was needed to satisfy the insatiable appetites of Europe’s masters.

47%

d

20/2

Caption

Nous avions tous très peur de ne pas la terminer vivants, cette nouvelle putain de journée de guerre commencée en première ligne !

We were all very afraid not to end it alive, this new fucking day of war started on the front line!

We were scared shitless that we might not see the end of that fucking day on the front line.

38%

  • 9 See Hulley, 2022

Dascher’s decisions appear to reinforce the observation that, while she certainly does adapt, there is little evidence to suggest that she is motivated by a desire to emulate character voice in the TT. Indeed, statements made by Dascher herself9 suggest a more formal approach to translation than the figure determined by the methodology (41 %) might represent. This perhaps reinforces the view that it is only studying the translator themselves that will afford a fuller understanding of the decisions made throughout the process of translation. Evidently, it would be necessary to study more texts by Dascher in order to get a clearer picture of her approach.

Mattias Ripa

Figure 6. distribution of segments (y) by RT similarity (x) for Ripa.

Figure 6. distribution of segments (y) by RT similarity (x) for Ripa.

Autobiography Persepolis, set against the backdrop of the Islamic revolution in Iran, won the Harvey Award for Best American Edition of Foreign Material in 2004. Satrapi’s husband Matthias Ripa is credited with the translation, and it remains his one and only published translation to date. Pantheon’s Anjali Singh, who also worked with Siegel on The Rabbi’s Cat, is credited as the book’s editor. The TT is 64 % similar to a RT, making it the most formal of texts studied here. The histogram shown in Figure 6 reflects a marked preference for linguistic proximity with 70 % of segments exhibiting similarity to the ST of over 50 %.

Table 6. Selected translations from Persepolis.

Segment

Page/Frame

Character

Satrapi

RT

11 Feb 2020

Ripa

Fuzzy

a

14/1

Father

Demain il y aura une manifestation

Tomorrow there will be a demonstration

Tomorrow there will be another demonstration

83%

b

15/3

Marjane

Pour qu’une révolution réussisse, tout le peuple doit s’y mettre.

For a revolution to succeed, all the people must do it.

For a revolution to succeed, the entire population must support it.

62%

c

1/5

Caption

Nous n’aimions pas beaucoup porter le foulard, surtout qu’on ne savait pas pourquoi.

We didn’t much like wearing the scarf, especially since we didn’t know why.

We didn’t really like to wear the veil, especially since we didn’t understand why we had to.

41%

d

5/5

Marjane

Ben, c’est simple, ça sera interdit

Well, it’s simple, it will be prohibited

It will simply be forbidden.

35%

Table 6 shows selected segments from Ripa’s translation to help get a picture of his typical approach. In segment a, on page 14 in panel 1, the use of the modal will instead of “going to” seems odd as it implies that Marjane’s father organises the demonstrations, which he patently does not. The addition of another also alters the semantics because the singular demonstration evoked in the ST is specifically organised in response to the Tehran Rex cinema fire of 19 August 1978. In segment b, on page 15 in panel 3, the use of the modal must instead of the less formal “has to” is a claque that perhaps betrays Ripa’s non-native approach. In segment c, panel 5 on page 1, although Ripa has correctly identified it as a veil rather than a scarf, his decision to use an infinitive to wear instead of a progressive “wearing” after like counters the expected spoken paradigm. In segment d, on page 5 in panel 5, Ripa has dispensed with the parenthesis, however, while the segment reads well, he has overlooked the interjection Ben.

26Ripa’s similarity score of 64 % identifies his approach as firmly in the domain of formal equivalence and is perhaps indicative of his status as a non-native literary translator. Moreover, closer inspection reveals that he has a reluctance to deviate from the linguistic form of the ST to account for character register and voice.

2.7. Analysis

Table 7. Average RT similarity by translator.

Similarity to RT

Dascher

41 %

Bell & Hockridge (Control)

43 %

Thompson

45 %

The Lofficiers

47 %

Siegel

60 %

Ripa

64 %

Table 7 summarizes the average similarity to the NMT-generated reference translations, in order, by translator, for the case study. With similarity scores of over 50 %, the thumb-print figures for Siegel and Ripa suggest their translations are source-text orientated, and closer inspection of selected segments from both translators’ TTs confirms this. Implicitly, these figures enable us to imagine a scale of equivalence that ranges from the low forties, where the dynamic approach of Bell and Hockridge’s gold standard is situated, to the low sixties, and the formal approach used by Ripa. Consequently, it could be deduced that the figures obtained for Dascher, Thompson, and the Lofficiers all indicate a preference for a more dynamic approach to translation; and indeed, closer inspection of the TT by Thompson reveals an inclination to adapt for the target reader, an approach that is also employed to a lesser extent by the Lofficiers. However, while the 41 % similarity figure for Dascher’s TT implies that her approach is even more creative than that used by the translators of Asterix, closer inspection reveals that this is not the case. Although this might call into question the validity of the proposed methodology, it should be acknowledged that the ST tackled by Dascher was exceptional within the confines of the case study, being the only one devoid of dialogue and exhibiting limited verbal-visual semantic interplay. In other words, the ST tackled by Dascher was enough of an anomaly to render the resulting observations inconclusive. It should also be acknowledged that the figures obtained for the Lofficiers and Siegel are based on multiple texts and therefore are more likely to be representative of the respective approaches employed. By contrast, the other translators featured in this case study are represented by solitary translations, therefore additional TTs will need to be found in order obtain figures that more accurately reflect their preferred approach.

27Interestingly, the results also suggest that the methodology could be used to reveal the broad editorial preferences of publishing houses and editors. For example, Singh’s involvement in the two titles published by Pantheon (Persepolis and The Rabbi’s Cat) appears to reflect a formal editorial policy—with an average of 62 % similarity for the two TTs.

Conclusion

28The goal of this paper has been to demonstrate how leveraging NMT as a baseline reference translator can help translation scholars broaden the scope of their research, as well as offer publishers and editors a way to profile, and select, literary translators, using metrics. Conceivably, it could also lend itself to demonstrating the added value of human translation for teaching purposes. The methodology I have proposed offers a way to evaluate a translator’s approach more objectively than any discursive study into surface aspects might otherwise afford. Furthermore, by quantifying a translator’s approach for a complete body of text, it facilitates direct comparison with other translators regardless of the text type or genre of the ST. However, the case study has revealed a necessity to verify the validity of results, as well as ensure the studied corpus enables a like-for-like comparison. While no method could ever replace detailed linguistic analysis for accuracy and insight, I have shown that this methodology presents scholars with the means to quickly extend their study horizon beyond the limits set by time and resource constraints. In the process, despite the case study offering but a limited view of translations for the North American comic book market, I have been able show that some TTs are functionally comparable to those done by Bell and Hockridge on the Asterix series; thereby shedding new light on historical comics translation practices on the other side of the Atlantic, and revealing paths for further investigation.

Haut de page

Bibliographie

Corpus Texts

Blain, Christophe, Gus: tome 1, Nathalie, Dargaud, 2007.

———, Gus & His Gang, trans. Alexis Siegel, First Second, 2008.

Giraud, Jean, Le monde d’Edena, Sur l’étoile, Les Humanoïdes, 1983.

———, Moebius, Upon a Star, trans. Jean-Marc Lofficier and Randy Lofficier, Marvel, 1987.

Giraud, Jean and Charlier, Jean-Michel, Marshal Blueberry, La Mine de l’Allemand perdu, Dargaud, 1972.

———, Marshal Blueberry, The Lost Dutchman’s Mine, trans. Jean-Marc Lofficier and Randy Lofficier, Epic, 1991.

Giraud, Jean and Jodorowsky, Alexandro, L’Incal, L’Incal noir, Les Humanoïdes, 1981.

———, The Incal: The Dark Incal and the Bright Incal, trans. Jean-Marc Lofficier and Randy Lofficier, Epic, 1988.

Goscinny, René and Uderzo, Albert, Astérix et Les Normands, Une aventure d’Astérix, Dargaud, 1966.

———, Asterix and the Normans, trans. Anthea Bell. Hodder Dargaud, 1978.

Guibert, Emmanuel and Lefèvre, Didier, Le Photographe, Dupuis, 2003.

———, The Photographer, trans. Alexis Siegel, First Second, 2009.

Satrapi, Marjane, Persepolis, L’Association, 2000

———, Persepolis, The Story of a Childhood, trans. Mattias Ripa, Pantheon, 2003

Sfar, Joann, Le Chat du rabbin, La Bar-Mitsva, Dargaud, 2001.

———, The Rabbi’s Cat, trans. Anjali Singh and Alexis Siegel, Pantheon, 2005.

Tardi, Jacques, C’était la guerre des tranchées : 1914-1918, Casterman, 1993.

———, It Was the War of the Trenches, trans. Kim Thompson, Fantagraphics, 2010.

Tardi, Jacques, and Verney, Jean-Pierre, Putain de guerre !, Casterman, 2008.

———, Goddamn This War!, trans. Helge Dascher, Fantagraphics, 2013.

Works Cited

Baker, Mona, “Towards a Methodology for Investigating the Style of a Literary Translator”, Target. International Journal of Translation Studies 12, no. 2 (2000), p. 241-66.

Delesse, Catherine and Richet, Bertrand, Le coq gaulois à l’heure anglaise analyse de la traduction anglaise d’Astérix, Artois presses université, 2009.

Forcada, Mikel L., “Making Sense of Neural Machine Translation”, Translation Spaces 6, no. 2 (4 December 2017), p. 291-309.

Guerberof-Arenas, Ana and Toral, Antonio, “The Impact of Post-editing and Machine Translation on Creativity and Reading Experience”, Translation Spaces 9, no. 2 (December 31, 2020), p. 255-82.

Hulley, Bartholomew, “French Comics in English: A Descriptive Translation Study”, 2022. http://www.theses.fr/2022LORR0171.

Kaindl, Klaus, “Thump, Whizz, Poom: A Framework for the Study of Comics under Translation”, Target 11, 1999, p. 263-88.

Kenny, Dorothy, “Human and Machine Translation”, in Machine Translation for Everyone: Empowering Users in the Age of Artificial Intelligence, 23-49, Zenodo, 2022.

Munday, Jeremy, Introducing translation studies: theories and applications, Routledge, 2016.

Nida, Eugene Albert, “Principles of Correspondence”, in Toward a Science of Translating: With Special Reference to Principles and Procedures involved in Bible translating, E. J. Brill, 1964, p. 159-61, 165-77

Pym, Anthony, “Quality”, in The Routledge Handbook of Translation and Technology, edited by Minako O’Hagan, Taylor & Francis Group, 2020, p. 437-52.

Rothwell, Andrew, Moorkens, Joss, Fernández Parra, Maria, Drugan, Joanna and Austermuehl, Frank, Translation Tools and Technologies, First edition. Routledge Introductions to Translation and Interpreting, Routledge, 2023.

Saldanha, Gabriela, “Translator Style: Methodological Considerations”, The Translator 17, no. 1 (April 2011), p. 25-50.

Thompson, Kim, Dascher, Helge, Nieh, Camellia and Singh, Anjali, Translation Roundtable, Interview by Kristy Valenti. Site: The Comics Journal, 2 June 2010. https://web.archive.org/web/20130703035549/http://classic.tcj.com/international/translation-roundtable-part-one-of-three/.

Zanettin, Federico, “Comics in Translation: An Overview”, in Comics in Translation, St. Jerome, 2008, p. 1-32.

Haut de page

Notes

1 WER = Word Error Rate; TER = Translation Error Rate; HTER = Human-targeted Translation Error Rate; BLEU = BiLingual Evaluation Understudy; NIST = National Institute of Standards and Technology. (See Stephen Doherty, “Translation Technology Evaluation Research.”, in The Routledge Handbook of Translation and Technology, edited by Minako O’Hagan, Taylor & Francis Group, 2020, p. 339-53.)

2 For example: keeping the syntactical arrangement of the ST, maintaining the ST’s structure and length, translating words and phrases as closely as possible to their literal meaning (including idioms and cultural references), and neglecting readability or fluency in the target language in favour of linguistic proximity to the ST.

1 It should be noted that NMT output is sensitive to the full input string, therefore any observed variability of output is normally directly attributable to a variation of input. Other variations may be attributed to: scheduled training updates to the translation engine, customized training data, or a change in user preferences. (See Rothwell et al., 2023 and Hulley, 2022.)

2 See Guerberof-Arenas and Toral, 2020.

3 See Thomas H. Cormen (ed.), Introduction to Algorithms, MIT Press, 2001.

4 Statistically, shorter strings of characters are susceptible to return figures proportionately closer to either of the polar extremes than longer strings.

1 Patrick O’Beirne,. “VBA Code for Microsoft Text Translator API”, 20 January 2017. https://sysmod.wordpress.com/2017/01/20/vba-code-for-microsoft-text-translator-api/.

2 Sorin Sion, 2011. “Fuzzy Matching Scoring Algorithm.” https://forum.ozgrid.com/forum/index.php?thread/102729-fuzzy-matching-scoring-algorithm/

3 I ran the same scripts on the same texts in October 2022 and the results were only marginally different to those I had obtained 44 months previously.

4 See Rothwell, 2023.

5 The height of the ten “bins” in the histogram indicates the number of TT segments falling within that margin of similarity. Therefore, if the TT were to entirely match its corresponding RT the histogram would show all segments falling into the right-most bin [90 %, 100 %] with zero segments in the other nine bins. Similarly, for a TT that shows high dissimilarity to its corresponding RT most segments would fall into bins towards the left-hand side of the scale [0 %, 10 %]. Although zero similarity is theoretically possible, a histogram showing all segments falling into the left-most bin is improbable because a string with 26 characters or more in length is likely to match at least one letter in the alphabet.

6 For a full list of winners see https://www.harveyawards.com/en-us/winners/previous-winners.html

7 Curiously, despite Microsoft’s NMT cloud service censoring numerous other offensive terms, it demonstrates no reticence here in proposing “fucking” as the formal equivalent of “putain de.”

8 I am unable to comment on Thompson’s use of German.

9 See Hulley, 2022

Haut de page

Table des illustrations

Titre Figure 1. distribution of segments (y) by RT similarity (x) for Bell & Hockridge.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-1.png
Fichier image/png, 33k
Titre Figure 2. distribution of segments (y) by RT similarity (x) for JM&R Lofficier.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-2.png
Fichier image/png, 32k
Titre Figure 3. distribution of segments (y) by RT similarity (x) for Siegel.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-3.png
Fichier image/png, 31k
Titre Figure 4. distribution of segments (y) by RT similarity (x) for Thompson.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-4.png
Fichier image/png, 35k
Titre Figure 5. distribution of segments (y) by RT similarity (x) for Dascher.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-5.png
Fichier image/png, 32k
Titre Figure 6. distribution of segments (y) by RT similarity (x) for Ripa.
URL http://journals.openedition.org/palimpsestes/docannexe/image/9140/img-6.png
Fichier image/png, 34k
Haut de page

Pour citer cet article

Référence électronique

Bartholomew Hulley, « Assessing human translation style with the help of NMT: a case study of French-language comic book translators »Palimpsestes [En ligne], 38 | 2024, mis en ligne le 29 novembre 2024, consulté le 19 mars 2025. URL : http://journals.openedition.org/palimpsestes/9140 ; DOI : https://doi.org/10.4000/12spd

Haut de page

Auteur

Bartholomew Hulley

Bartholomew Hulley a obtenu son doctorat en traductologie à l’université de Lorraine sous la direction de Catherine Delesse. Il est actuellement directeur adjoint de la faculté satellite de l’université de Syracuse à Strasbourg où il enseigne également l’écriture académique et dispense un cours sur la bande dessinée. De plus, il travaille comme traducteur littéraire indépendant, contribue régulièrement au The Comics Journal et dirige un cours sur la traduction littéraire à l’Institut Européen des Métiers de la Traduction (IEMT) de l’université de Strasbourg. Il a présenté des communications sur la traduction des bandes dessinées aux États-Unis, en Espagne, en France et en Allemagne.

Haut de page

Droits d’auteur

Le texte et les autres éléments (illustrations, fichiers annexes importés), sont « Tous droits réservés », sauf mention contraire.

Haut de page
Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search