While the paper is the result of co-writing, the authors would like to point out that the sections titled “Responsible MT”, the introductory part of “The revision experiments” and “ST-2: Silent Spring” were written by Kristiina Taivalkoski-Shilov and “Translation revision and post-editing” and ST-1: The Job” were written by Paola Brusasco. The authors would like to thank the academics, the editor, the publishers, and the translators who took part in this research.
1Neural networks and deep learning applied to machine translation (MT) have been yielding more and more adequate outputs, characterised by a marked improvement on the clumsy machine translations of previous decades. As a result, even literary texts are being investigated as potentially machine translatable, with or without human post-editing (PE) (see Besacier, 2014; Toral and Way, 2015, 2018; Toral et al.,, 2020, among others). Fascinating as this technological advancement is, the possibility of literary texts translated integrally by MT does not sit comfortably because of the threat it poses to translators and to one of the most typically human characteristics—creativity. Rossi (2018: 50) has pointed out that literary translators both read and write creatively. These two dimensions of what Rossi calls “literary creativity” are hindered by the MT process, also when combined with human PE. With respect to creative reading, the post-editor (whether a professional literary translator or not) who is reading segments of raw machine translations instead of the source text is not absorbed by the literary work and becomes an evaluator rather than a creator (Guerberof-Arenas and Toral, 2022: 207; Taivalkoski-Shilov and Koponen, forthcoming). Also creativity in writing seems to be affected: research on MT and PE of literary texts has so far indicated that even when done by professional literary translators, it leads to overly literal versions because the post-editors are primed by the MT suggestions and do not come up with other solutions (see e.g. Moorkens et al.,, 2018; Guerberof-Arenas and Toral, 2020, 2022). It seems reasonable, therefore, to leave the human in control of the literary translation process. Instead, one could try and avail oneself of the possibilities offered by MT and translation technology in other tasks related to the publication of translated literature, which do not require a comprehensive creative process. We envisaged that revising a translation for republication, which means editing, correcting and/or modernising a previously published translation (Koskinen and Paloposki, 2016 [2010]) could be such a task, at least when it is done by comparing the target text to the source text. Revision for republication can vary from light copy-editing to extensive rewriting, which is in fact retranslating (Vanderschelden, 2000: 2; Koskinen and Paloposki, 2016 [2010]; Koskinen, 2018: 316). In the latter case it can be time-consuming, even painstaking: an extreme case is Terence Kilmartin’s revision of Scott-Moncrieff’s translation of À la recherche du temps perdu by Proust (Remembrance of Things Past), which took more than ten years, and Kilmartin died before he could complete the task (Vanderschelden, 2000: 2). Revisers also tend to get little credit for their work and usually remain invisible to the readers (ibid.: 3; Koskinen, 2018: 316-317, 319). Given the laboriousness and rather low status of revision, we decided to test whether translation technology could support humans in this specific way of reprocessing translations.
2Our article is based on a study where we investigated the possibilities offered by technological tools in revising two texts originally written in English, Sinclair Lewis’ The Job (1917) and Rachel Carson’s Silent Spring (1962)—the former translated into Italian in 1955 and republished unchanged but for the title in 1970, and the latter translated into Finnish in 1963. We sought feedback from professionals working in literary translation to test the usability of our approach. Since the topics—working women’s difficulties and environmentalism—are still timely and both translations are usable mutatis mutandis, we envisaged revised versions carried out with the aid of a translation memory (TM) and MT. The TM in question is integrated in the Phrase Translation Management System (https://phrase.com/) for which the affiliation of one of the authors has a licence and the MT is the general-purpose neural MT service DeepL (https://www.deepl.com/translator), freely available online with a character limit or with subscription (DeepL Pro). Two chapters were selected from each text to be revised. Our study is an experiment to ascertain a) whether alignment with the help of TM is useful in the revision process, and b) whether selected strings of MT output can be effectively incorporated in the parts of the published versions that according to us needed revision (translation errors, omissions, outdated words or expressions etc.) or, alternatively, what kind of editing would be required. The nature of our research is applied empirical research in that it studies a practical problem related to reprocessing translations and tests whether the method in question (TM alignment and MT) could be useful in the revision process (Saldanha and O’Brien, 2013: 15). Our research methods are comparative textual analysis (between the source texts, the published human translations and DeepL versions) and data collection through interviews.
3In what follows, we first introduce the key notions related to our research: responsible MT, translation revision and PE. Then we describe both experiments and their respective results. In the concluding section we discuss our research questions in the light of both revision cases and envisage future research.
4One of the starting points of our research is the notion of responsible MT that we extend to all translation technologies. As Moniz and Parra Escartín (2023: 2-3) observe, researchers in MT have only recently started to pay attention to ethical aspects related to the development and use of MT, perhaps in the wake of ethically oriented research in Natural Language Processing. This trend has probably been influenced also by the discipline of Translation Studies (see Kenny et al.,, 2020: 1-3) that has a long tradition of ethical reflection on translation and encounters with Otherness (for an outline of seminal texts, see e.g. Chesterman, 2016: 24-27), and where scholars have in recent years openly commented on ethical aspects of MT and translation technologies (see e.g. Cronin, 2017; Kenny, 2017; Taivalkoski-Shilov, 2019; Kenny et al.,, 2020). Ethically aware research on translation technologies addresses central moral issues that are common in any use of AI: human benefit, responsibility (e.g. for adequate training data and translation quality as well as data privacy and environmental effects), transparency, education, and awareness of its implications (Moniz and Parra Escartín, 2023: 1-4). Moreover, it takes into account the risks involved when using translation technologies and seeks sustainability both with regard to humans working in the translation industry and the environment (see Kenny et al.,, 2020: 4-5). It opposes extractivism (i.e., a short-sighted, dominance-based culture of exploitation) that threatens the environment and translation as “traditional knowledge and know-how” (Cronin, 2017: 29, 32-33; Talbot, 2022: 160). For Moniz and Parra Escartín (2023: 3), responsible MT can be understood “as a combination of all factors that need to be considered when developing and deploying MT systems to ensure that such systems are ethically designed, including, but not limited to, data bias, data licences and rights, ecological footprint, and intended end-users”. In the context of literary translation, where the uses of MT have also been studied from the mid-2010s, the rights of authors, translators and other agents of translated literature, readers and entire cultures have to be taken into account (see Taivalkoski-Shilov, 2019).
- 1 Sorbonne nouvelle, 20 October 2022.
5Research on literary MT has moved rapidly in the last few years after the introduction of Neural Machine Translation systems (Bahdanau et. al, 2014; Hadley et al.,, 2022: 5-7). The pioneering researchers started with experiments on how fully automated machine translation engines handled the translation of short passages from narrative texts (Toral and Way, 2015). Literature-specific MT engines in some language pairs have also been trained and this work goes on (Toral and Way, 2018; Toral et al.,, 2020; Hansen et al.,, 2022). The suitability of PE for literary translation was also studied from early on (Besacier, 2014); however, as we already mentioned, the experiments have not been encouraging since this practice seems to negatively affect the quality of translation, both as a process and a product (see Moorkens et al.,, 2018; Kenny and Winters, 2020; Guerberof-Arenas and Toral, 2020, 2022). Hence new research questions have emerged, such as whether machines can inspire literary translators, which was the theme of Antonio Toral’s plenary lecture at the conference Literary Translation and AI: assessing changes in translation theory, practice and creativity1 (see also Kolb and Miller, 2022) or whether MT can provide an on-demand translation of difficult passages for readers of e-books in a foreign language (Oliver et al.,, 2022). This trend also indicates an emerging ethical awareness in research on literary MT. Our study, which aims at finding suitable uses of MT without threatening professionals of literary translation, ascribes to this line of research.
6As a broad initial distinction, we can say that revision refers to “the process of checking human translations” (Feinauer & Lourens, 2021: 165) carried out by human revisers, while PE refers to the checks and corrections made by human post-editors to machine-translated texts—probably the oldest form of interaction between humans and machine-generated texts for translation purposes (O’Brien et al.,, 2014: 8).
7As far as non-literary texts are concerned, revision aims at “making those changes which are necessary to achieve compliance with recognised linguistic and functional criteria” (Koponen et al.,, 2021: 1), thereby assessing quality and the usability of the text for its intended purposes. Functional criteria, however, hardly apply to literary texts, whose revision is geared towards checking interpretation, lexicon, and style. Sometimes—as we are envisaging here—an already published translation is revised for republication, usually because some or most of it no longer reflects the target culture’s use of lexical items, syntactical or stylistic patterns, translational conventions, or the advancement of knowledge (see also Van Poucke, 2017). However, aged lexicon, and even syntax, would not normally be an issue because readers tend to accept it as typical of classics; therefore, revision seems instead to concern style or translational norms, and—depending on its extension—may blur the boundary with retranslation (Koskinen and Paloposki, 2016 [2010]).
8In preparation for publication, a human translation can undergo either monolingual editing or comparative revision. The former involves checking the target text mainly for surface aspects (typos, punctuation) and naturalness, only occasionally resorting to the source text, with actions that fall within the scope of “meaning-preserving changes” (Faigley and Witte, 1981: 403-404), i.e. adding connectives, glosses, or elements that clarify the sentence; deleting repetitions, redundant information and functional items that are not wrong but would not be normally used in the target language (for example possessive adjectives and subject pronouns in Italian); replacing lexical items because of register or usage; rearranging the components of a sentence. In the same category of “meaning-preserving changes” fall “distributions” and “consolidations” as ways to, respectively, expand and compress sentences by breaking them into two units or, conversely, bringing them together into one; such interventions, however, rarely apply to literary texts because they alter syntax and rhythm, which in turn would affect the style of the source text author and the first translator.
9Comparative revision, instead, entails checking the whole translation against the source text, a demanding activity that usually does not entitle the reviser to any mention in the published book. This operation might reveal the need for meaning changes (ibid.) at the level of microstructure by applying the same categories of actions described above to sections where misinterpretation has occurred; meaning changes at macrostructural level are not considered because it is hard to imagine interpretive errors that would generate changes in the plot. Comparative revision applied to an already published text, which is the focus of our experiment, is a twofold attempt to salvage and honour a translation that mostly works and at the same time allow the author and the text to enjoy renewed visibility. As Washbourne (2016: 152) notices, this kind of revision—which may also include retranslated segments—may generate multivocality in the text, an aspect that needs to be carefully considered whenever a change is made.
10Research on both revision and PE has been growing, focusing on the product, the process or the participants. Literary revision and PE, however, is less researched due to its peculiarities and niche status compared to the global volume of translation. Robin (2016) described the ideal steps to follow and the skills needed by revisers, underlining the need for a macro-level perspective that foregrounds the text as a whole (see also Vanderschelden, 2000: 3). Scocchera (2017) investigated the relationship between Italian literary translators and their revisers, and later (Scocchera, 2019) envisaged a course in revision alerting trainees to the risks of introducing errors, misinterpreting and rewriting. Feinauer and Lourens (2017) investigated the loyalty of revisers of a draft translation to see whether they leaned towards the source text, the author, the target text or the reader, finding that the alleged needs of prospective readers did not determine any particular strategy or choice, while loyalty was mostly reserved to source text and author. They also questioned the need to distinguish between revision and editing, given that in real practice each text may involve different editorial practices and collaborations, an aspect also discussed in Feinauer and Lourens (2021).
11Similarly to revision, PE implies human evaluations conducive to minor amendments and/or substantial changes where necessary. Toral et al., (2018) conducted the first study on professional literary translators post-editing a novel translated by phrase-based statistical MT and neural MT, in order to assess the time and cognitive effort involved: they found that productivity increased, but so did the cognitive effort required by PE compared to human translation from scratch.
12In our study, we are envisaging a hybrid process drawing on both revision and PE to be carried out by professionals who may work as translators, revisers or editors for the publishing industry.
13The texts to be revised, Sinclair Lewis’ The Job (1917, translated by Beatrice Boffito in 1955, then republished in 1970 with a different title) and Rachel Carson’s Silent Spring (1962, translated by Pertti Jotuni in 1963), were selected owing to their status as classics and the fact that both translations have ceased to be republished. Since the last editions date back to the 1970s, modifications are likely to be required if they were to be published again. We agreed on a common revision principle: to modernise the texts for contemporary readers of the 2020s while still retaining the translator’s voice. This involves comparative revision tasks such as correction of mistranslations and omissions, updating terminology and lexis, and making occasional stylistic improvements, while paying attention to overall textual cohesion, which according to Vanderschelden (2000: 3) might suffer when details are modified (see also Washbourne, 2016: 152 and Robin, 2016). We also agreed on common procedures in the revision process. Two chapters were chosen from both works. The source texts and the published translations were typed as Word files and checked. In addition, a DeepL translation in Word format was made from both source texts. The resulting text pairs (The Job and Boffito’s translation; The Job and DeepL translation in Italian—Silent Spring and Jotuni’s translation; Silent Spring and DeepL translation in Finnish) were then aligned using the TM in Phrase TMS. The alignment yielded Excel files of the text pairs. The segments in each Excel file were then checked and, if needed, re-aligned manually. Finally, the DeepL segments were copied and inserted into cells in a column parallel to the human translations in order to facilitate comparative analysis.
14Because of differences in the two source texts, target languages and circumstances, our revision projects went slightly separate ways. In the case of Boffito’s version of The Job, the study was more focused on the usability of the DeepL output to replace items or segments of the existing translation perceived as no longer usable. In the case of Silent Spring, where DeepL suggestions were less useful for the revisions needed, the experiment was rather on the usability of the process as a whole and on what other kind of editing was needed. Both revision processes will be described below.
15American author Sinclair Lewis’ The Job (1917) revolves around the character of Una Golden, a young woman who feels constrained by the prospects her native small town offers to women (marrying or finding work—as a teacher or domestic help depending on their family background) and, finding herself in need after her father’s death, moves to New York with her widowed mother in an attempt to shape a new life. She begins as a stenographer and manages to build a career in business despite the difficulties working women face in the workplace and in society.
The book was translated into Italian by Beatrice Boffito and published by Rizzoli as Le donne lavorano [Women Work] in 1955, then again in 1970 for another publisher, I Nobel letterari, with the more captivating title of Una donna emancipata [An Emancipated Woman]. The translation still reads fairly well, but there are two aspects that require revision: 1) a certain degree of ageing of the language, and 2) some mistakes, either in interpretation or, more frequently, concerning culture-specific elements which at the time were not known as well as information that in pre-internet times would be hard to check without physically being where the story is set. The former group includes subject pronouns ella, egli, essi, essa, esso [she, he, they, it, respectively] which are hardly ever used in contemporary Italian except for very formal texts, currently replaced with the corresponding object forms lei, lui, loro, or Ø; the elision of the final “e” in verbs in the infinitive form, which creates a poetic or funny sound pattern; the translation of Nuova York and words that have now entered Italian as borrowings or calques like “ketchup”, liaison, “shaker”; grigliata mista [mixed grill]; some dated lexical choices; the presence of footnotes with explanations that are no longer necessary or might be glossed within the text given that, by convention, the flow of non-professional reading should not be interrupted. This kind of revision could be carried out manually, but the ageing often involves syntax and turns of phrases that either elevate the register or make sentences very unusual. For example, “beauty of words or music, of faith or rebellion, did not exist for him” (Lewis, 1917: 1; bottom cell in Fig. 1 below) is translated literally, which results in a correct but very elevated sentence; by contrast, the MT version, though maintaining the same equivalents, adds articles and postpones the verb, which makes the sentence more natural by current standards.
16There are also errors at single-word level and interpretive mistakes; in such cases, the revision involves replacing items and/or reformulation, a longer process which might benefit from MT if the output proved to be largely acceptable. For example, in Chapter 2 loafers is rendered by Boffito as facchini [porters], instead of perdigiorno or fannulloni [idlers], elevateds as ascensori [lifts] instead of sopraelevate, and on its once leisurely lawn, referring to a green around a mansion being increasingly taken over by small shops, as sul suo già aristocratico prato [on its already aristocratic lawn], where the choice of the adverb già with the meaning of “ex” premodifies a noun phrase in a way that today would only be used to indicate a change in a toponym or in office. An interesting case is “she was a creature of action to whom this constricted town had denied all action except sweeping” (Lewis, 1917: 19, emphasis mine; second last cell in Fig. 2): Boffito used the verb scopare, which in itself is correct, but commonly used today as a vulgar reference to sexual intercourse, with the result that the passage may be interpreted not as “her only possibility for action was housework”, but as “given her sense of constriction, the only (rebellious) possibility for action was being sexually promiscuous”. Obviously, the previous section would cast some doubt over that interpretation, to be soon dispelled by reading on, but the word needs to be changed. As the term is probably tagged as taboo, DeepL uses its synonym spazzare, which is correct but needs completion, e.g. spazzare il pavimento [sweep the floor], while more acceptable solutions could be fare le pulizie or pulire casa [do the cleaning]. It is however remarkable that the software avoids the first translation option and in so doing bypasses a pitfall.
17For the purpose of this study, Chapters 1 and 2 were selected as they set the scene, introduce the main characters and provide the reasons for both the move to New York and Una’s subsequent decisions regarding her new life. Chapter 1 is made up of 3,670 words (3,802 in Boffito’s version; 3,740 in DeepL), and Chapter 2 of 2,094 (2,114 in Boffito; 2,121 in DeepL). The respective lengths show a slight increase in the Italian versions as is usually the case since prepositions and articles need to be added, but no relevant difference between Boffito and MT, which contributes to the feeling of reliability of the latter, since it shows no big omission or addendum. Once aligned, the texts yielded 162 and 78 vertical segments respectively, with very different lengths spanning from one to twelve lines. Drawing on personal experience in literary translation, the first step was to read Boffito’s version, mark the items that need to be changed and compare the content of each cell with the source text (ST), then check if the DeepL option was usable. Figures 1 and 2 below are samples of the work done on contiguous sections of Chapters 1 and 2:
Figure 1. A section of the aligned texts in Ch. 1 showing elements to be revised and acceptable MT sections.
Figure 2. A section of the aligned texts in Ch. 2 showing elements to be revised and acceptable MT sections.
In order to go beyond personal impressions, nine informants with translation experience (five professional literary translators, one also translating other types of texts, and three academics) were asked to compare the aligned texts—source (ST), human translation (HT), machine translation (MT) —, identify the parts of the HT that they would change and mark the corresponding MT words, segments or paragraphs only if usable in their current form or with minor changes. Three translators and one academic refused owing to time constraints. The remaining five accepted to participate to different degrees and signed an informed consent document. Their feedback is summarised here below.
18One academic (A1) and one professional (P) expressed prejudice against MT, so their contributions were given in the form of general comments and a few examples without using the Excel files provided. Interestingly, according to A1, Boffito’s translation is still adequate, and obsolete items contribute to recreating the atmosphere of the early 20th century, while P said that though very good, the translation would need to be entirely rewritten because of its overall agedness. A1’s main objection to the MT version is its literalness, which sometimes results in redundance—un cane setter [a setter dog] instead of un setter; vetrine di negozi, calqued on shop windows, instead of simply vetrine—and at other times limits the interpretive work of the translator, as in le colline al di là [the hills beyond], which is correct but sounds incomplete in Italian, and is rendered instead as le colline circostanti [the surrounding hills] in the HT. Although appreciative of the many acceptable segments and even paragraphs of the MT, A1 considered the process too time-consuming and pointed out the risk of adopting MT choices which are correct per se, but unsuitable in the context. As anticipated, P felt that the text should be retranslated because the quantity of changes needed would result in a very unequal text and threaten coherence. The abundance of usable solutions for collocations (e.g. fare proclami instead of emettere proclami for make proclamations) and more common or modern alternatives for obsolete terms (e.g. municipio instead of palazzo comunale for town-hall) was evaluated positively, but in P’s opinion these are minor aspects, and a translator does not need such obvious suggestions. As an overall evaluation, P thought that 1) although dated and flawed by errors, the 1955 version is coherent and shows that the translator looked for solutions suitable at global level, while the MT output is hit-and-miss; 2) having the MT version is counterproductive for a translator because it is likely to prompt them to accept a term or segment that is correct and looks acceptable without investigating interpretive possibilities, and this in turn would lead them to work at segment level, thus missing the textual dimension; and 3) using MT, even just as a tool in the revision process, is bound to generate a levelling of writing styles including the perception of what is “correct” in translation. Hence, although starting from different experiences and points of view, neither informant considered the aid of MT in revision as a desirable procedure, and both focused on the risk of missing the textual dimension, thus compromising lexical choices (A1) and interpretive possibilities (P).
19The remaining three participants (A2, T1, T2) worked on the Excel files. Since their feedback was very detailed and abundant, quantitative data and language aspects have been grouped in order to present some general features. The translators/revisers showed different approaches: overall, the academic (A2) favoured a more literal translation, while the translators (T1 and T2) opted for more fluent renderings. T1 even inserted their own changes when dissatisfied with the MT. T2 marked as problematic longer sections, sometimes whole sentences, in Boffito’s version; on three occasions, T2 was impressed with the fluency and correctness of the MT segments, but overall felt that word order should be changed to reproduce emphasis and irony. Quantitatively, A2 marked as unsatisfactory 144 single terms and short phrases, and 42 segments of various length in Boffito’s version, accepting respectively 95 and 31 of the replacements found in the MT output; T1 marked 115 single terms and short phrases, and 39 segments in the HT, accepting respectively 41 and 19 suggestions; T3 highlighted 204 words and phrases and 123 segments, accepting 83 and 45 MT proposals respectively. Many items highlighted for revision in Boffito’s translation are the same for all the revisers, but a large number depend on personal taste. With reference to subject pronouns, culture-specific items, and the translation of geographical names, as expected, they all marked them as obsolete; for the two quotes discussed in greater detail above—the one in elevated register and the term that has acquired a vulgar meaning and the potential ensuing misinterpretation—all three participants accepted the DeepL versions. In some cases, the verbs in the MT output were accepted but the tense had to be changed, usually from a form that denotes one occurrence in the past (passato remoto) to one that denotes repetition or a condition in the past (imperfetto), as both Italian tenses translate the Simple past and only context determines the choice. As expected, MT solutions at single word or phrase level are more frequently accepted than segments, which in itself is still fairly useful as it speeds up the search for synonyms or correct translations, but it may affect the global cohesion of the text. On the whole, usable solutions are below 50 %, a good score, but not so interesting if one considers the time needed to prepare the material and carry out the analysis.
20The environmental bestseller Silent Spring (1962) by the biologist and nature writer Rachel Carson (1907-1964) counts among the most controversial and influential American books ever to have been published. The book and its translations into several languages provoked an outcry against pesticides and influenced the banning of DDT in the United States in 1972 and later also in other countries. Carson’s book has made an indelible mark in ecological thought because it changed people’s attitudes towards pesticides in the Western world and played a significant part in the spread of environmental movements worldwide. Carson was a particularly skilful writer who was able to create readable and rhetorically effective texts. The “unforgettable image of a ‘silent spring’” (McCrum, 2016)—a spring without birdsong because most birds have died from pesticides)—was a particularly powerful rhetorical device in Silent Spring. (Stoll, 2020 [2012]; Taivalkoski-Shilov, 2020: 126-127.)
21The Finnish translation by Pertti Jotuni was published by Tammi Publishers in 1963 (2nd edition 1963, 3rd edition 1970). According to a previous study on Jotuni’s translation, it is rather complete, but important paratexts have been omitted: two epigraphs where Carson cites a verse from Keats’s poem “La Belle Dame sans Merci” and a thought by E.B. White, a dedication to Albert Schweitzer, the fifty-five page long bibliography, and the index. Jotuni’s translation also contains some shorter omissions. (Taivalkoski-Shilov, 2020: 133, 136-137.) Consequently, there would be grounds to revise it, all the more so since a Finnish translation of Keats’s poem, by Leevi Lehto, was at last published in 2020.
22Chapters One (“Fable for Tomorrow”) and Eight (“And No Birds Sing”) were selected for revision since they are particularly central for the rhetorical image of the silent spring. Chapter One contains 683 words (514 in Jotuni’s translation; 508 in DeepL), and Chapter Eight has 8,371 words (6,331 in Jotuni; 5,943 in DeepL). The chapters were processed as described above. The Excel file containing Chapter One had 40 vertical segments and the one containing Chapter Eight 378. These files, where the actual revision was done, contained five separate columns: the leftmost column contained segments from the source text, followed by Jotuni’s version that was compared to Carson’s, then a separate column for revision suggestions and another for observations, justifications or references related to the information mining required for the task. The rightmost column contained the DeepL version. The words or passages in the segments that needed revision were indicated in bold because in most cases a minor replacement of a couple of words would be enough. The DeepL suggestions were considered only if revision was needed: if they were usable without editing they were highlighted in green; if they were usable with post-editing, they were highlighted in yellow. In Chapter One, revision suggestions were made regarding fifteen segments, i.e., 37.5 % of the chapter. Six out of these fifteen segments were revisable with the help of DeepL (three without editing and three with post-editing). In Chapter Eight, revision suggestions concerned 206 segments, i.e., 54.5 % of the whole (there were overlaps, for instance the same bird name, e.g. “catbird”, omitted in more than one segment or the rendering of (DDT) spraying by the Finnish word pölytys [powdering] that was suggested to be replaced by ruiskutus [spraying]). DeepL proved helpful in 65 out of these 206 segments, 19 without editing and 46 with post-editing. All in all, these figures show that the translation would need substantial revision, and that DeepL would only be limitedly productive in the reprocessing of Jotuni’s translation.
23As the revision began, it turned out that Jotuni had systematically left out the names of several bird species from his translation, which were either rendered with a general word or omitted. Example 1 illustrates this tendency in Chapter Eight:
- 2 Words in bold face are used in quotations throughout the article to emphasise certain points.
Example 1:
ST: These include three of the thrushes whose songs are among the most exquisite of bird voices, the olive-backed, the wood, and the hermit.2 And the sparrows that flit through the shrubby understory of the woodlands and forage with rustling sounds amid the fallen leaves—the song sparrow and the white-throat—these, too, have been found among the victims of the elm sprays. (Carson, 2002 [1962]: 110.)
TT: Näiden lajien joukosta mainittakoon kolme erilaista rastasta, joiden laulu kuuluu ihmeellisimpiin. Jalavan pölytyksen uhreiksi ovat joutuneet myös ne varpuset, jotka lentelevät metsämaan vesakoissa ja keräävät ravintonsa pudonneiden lehtien joukosta niitä kahisuttaen. (Jotuni, 1970 [1963]: 100.)
[Back translation: Among these species let us mention three of the thrushes, whose song is most exquisite. The sparrows that fly here and there in the thickets of the woodlands and rustle fallen leaves when gathering their food, are among the victims of the elm powderings.]
Jotuni’s translation solutions can be explained by the difficulties of information mining in the 1960s and, according to a publisher of Tammi Publishers, who was interviewed via email for this study, by a less strict translation tradition (“if there was any”) and a looser attitude towards the source text author’s copyrights. If the translation were to be republished today, all bird names would be restored (e-mail communication 3 May 2023).
24Jotuni’s translation also contained misinterpretations and some other short omissions. As a whole, however, his version rendered well the semantic meaning of the source text and had stylistic qualities that would justify its republication, provided the errors, omissions, some outdated terms or expressions and clumsy renderings were corrected. It also soon became evident that DeepL (DeepL Pro for the pair English-Finnish) was not a very useful revision aid in this particular revision task. It had systematically made errors in the translation of the bird species names, rarely suggesting the right name, sometimes rendering two different names by the same target version and in some cases by producing pure nonsense, as in Example 2:
Example 2:
ST: [T]he nesting population in the neighborhood seems to consist of one dove pair and perhaps one catbird family (Carson, 2002 [1962]: 103).
TT (DeepL): naapuruston pesivä populaatio näyttää koostuvan yhdestä kyyhkyläisparista ja ehkä yhdestä kissanristiäisperheestä.
[Back translation: [T]he neighborhood’s nesting population seems to consist of one pair of lovers and perhaps one family of cat’s christening.]
- 3 https://www.birdlife.fi/lintutieto/maailman-lintulajien-suomenkieliset%20nimet/
Consequently, the revision was done mainly with the help of information searches (for instance, the website of BirdLife Finland3 that has a repertory of the Finnish names of the birds of the world), online dictionaries (Oxford English Dictionary, English-Finnish MOT dictionary), by asking friends or checking the French translation of Carson’s book. The process was quite painstaking. Having said that, comparing segment by segment in the Excel file made it easy to evaluate Jotuni’s translation and notice good and bad solutions, omissions or additions. This method could be useful in other kinds of translation analysis as well. However, as it was difficult to estimate whether it would make revision any easier in a professional setting, feedback was sought from Finnish professionals who work in book publishing. Two professionals who have experience in revision were interviewed for this purpose.
25A semi-structured format was designed for both interviews with questions addressing the policy and practice of revising old translations in the publishing houses in question and the participants’ opinion concerning the method that had been developed for this study. After obtaining informed consent from both participants, the interviews were conducted on 10 May 2023 on Microsoft Teams using the recording and transcription option offered by the software. The Excel files with the revisions were shown to the participants with Microsoft Teams’s “share” function.
26The participant in the first interview, who has more than 25 years of experience, is currently head of publishing (H) at a renowned publishing house that counts among the oldest in Finland. According to H, revision of already published translations used to be a rather marginal activity in that publishing house. It was done in case a clear error had been discovered in a translation or if an old translation that was otherwise publishable had several small errors that could be amended with proofreading. H remembers only one case where a translation had been given to a professional translator for thorough revision and the translator had complained afterwards that the task was as time-consuming as retranslating the whole book. Consequently, as professionals know how arduous a comparative revision can be, it has rarely been resorted to.
27However, as Koskinen and Paloposki (2003: 26) anticipated twenty years ago, with digital publishing things have changed. Now the publishing house’s backlist is often under scrutiny, as it is possible to rediscover works that can be turned into e-books and audiobooks, if the existing translation is of sufficient quality. The team of digital publishing does the first scouting by searching old translations from the archives and by reading the beginning. In case they consider that the translation is written in beautiful Finnish, an initial decision to publish can be made. The translation is then scanned into a PDF file and checked more thoroughly (for instance, politically incorrect expressions are always on the checklist). Usually, however, no systematic comparisons to the source text are made. This publishing house has a policy that, out of solidarity towards the translator whose work is being revised, the revision task is practically never given to another translator, but translators might be asked to revise their own translation. The revision is mostly done by the staff of the publishing house. The translator’s copyright has to be respected in the revision: all changes need to be accepted by the translator, or by their heirs. According to H, revision is not done if substantial changes are required. H considers revision particularly suitable for entertainment literature, for instance well written detective series; for classics, instead, a retranslation might be commissioned. This contrasts with Koskinen and Paloposki’s (ibid.: 34) findings from the years 2000 and 2002 in Finland; in their data both reprinting and retranslation concerned mainly classics. It would be interesting to study whether increased digital publishing has had an influence on the genres of books selected for republication.
28H found the Excel file, and especially the DeepL versions very interesting and made many observations on Jotuni’s translation, confirming, for instance, that Jotuni’s pölytys [powdering] would need to be replaced. As for DeepL, H noted that it can produce both clumsy raw translations or quite decent quality, but “a human eye” is necessary to evaluate the end result. H further observed that the Excel file permitted precise comparisons between Carson’s, Jotuni’s and the DeepL versions, but that usually revision does not entail such careful comparisons. H concluded that while the application of this method would result in a “big renovation” of Jotuni’s translation, the revision actually needed would only concern minor details, since his style was eloquent. So, replacing the segments indicated with bold letters might be enough for an excellent end result. If the revision were a real commission, the suggestions would need to be incorporated in Jotuni’s translation (possibly with some comments) and sent to Jotuni’s (1931-2007) heirs for authorisation.
29The participant of the second interview is a publishing editor (E) at a newer, equally reputable Finnish publishing house. E has more than fifteen years of experience and usually edits new translations as well as original works before publication. E has also been involved in the publication of an anthology for which a great number of shorter already published translations were revised. On that occasion, E collaborated with the editors and the translators of the single texts. This meant reading the manuscript of the anthology prepared by the editors (in Word format), making suggestions and contacting the translators in order to discuss them and obtain permission. E says that in their publishing house, the working methods concerning the revision of older translations vary from case to case. Sometimes, the process goes as described above. In others, the translator first revises or edits their own translation, and then the editor proofreads and possibly polishes the text. In any case the translator—or the heirs or a person authorised by them—checks the revision suggestions and perhaps makes some changes of their own to the final version.
30E highlighted that in the revision process the editor’s task is to polish the language and style of the target text and that it is the translator who is responsible for the accuracy of the translation and who is expected to provide a text that is as ready for publication as possible. However, E characterises the revision process as an interactive one; the editor works in dialogue with the translator. E checks the source text only in case there is something odd in the translation and they know the source language. Consequently, for E the method with the Excel file could be useful mainly “at the translator’s end”; at the editor’s end it might slow down the process in case they had to go through the file systematically. However, the Excel file, sent by the translator as an attachment of the revised translation, could be helpful in case the editor wants to check something and especially when they do not know the source language.
31The idea of introducing TM alignment and MT in the revision of literary texts is an attempt to reduce the time and cognitive load involved in the task, and it was experimented with by the authors themselves, then submitted to Italian and Finnish professionals in the publishing sector for their own evaluation. The reactions ranged from a relative appreciation of the process for cases where the existing translations contain terminological inconsistency, some misinterpretations and omissions as in Silent Spring, to more negative positions expressed by the translators who evaluated the revisions of the Italian version of The Job, who think that alignment helps the comparison between texts and the correction of limited, factual errors (names, toponyms, punctuation, omissions), but at the same time the format does not allow revisers to build a macrotextual perspective. Also, the usability of the DeepL version varies hugely, and only rarely do whole sentences fit. The majority of our participants felt that the method we created would be rather time-consuming. Hence it would not be usable in the ordinary case of revising old translations for republication, which at least in the two Finnish publishing houses in question seems to involve mainly monolingual editing. However, sometimes comparative revision is resorted to, mainly by translators who revise their own work for republication, and in such cases this method could be useful. One of the supplementary findings of our study is that copyright issues significantly limit the opportunities for recycling old translations by revising them: if substantial changes need to be made, permission is required even if the translator is deceased. This raises the costs of revision (in terms of time and money) and makes retranslation a viable alternative especially for classics.
32Summarising the results of our study, unless the MT output is deemed exceptionally good, the process we envisaged is not likely to replace human revision at present for a number of reasons:
• Despite some suitable suggestions, the process as a whole does not prove faster than traditional revision because where longer strings are concerned a reformulation of the whole sentence or even paragraph is often necessary.
• More mature translators/editors/revisers may not be familiar with the preparatory stage—scanning older translations, aligning ST and TT, and obtaining the MT output—which would have to be carried out by the publisher. This may lead publishers to prefer more IT-inclined, less experienced revisers, which may affect both employability and the quality of the target text.
• Some of our participants see the reasons for segmenting the text, but feel hindered in their global reading experience, as it draws attention to single lexical items rather than syntax or rhythm.
• Judging from the reactions of the Italian professionals, literary translators do not seem likely to warm to the process because of their mistrust of MT, and lack of experience with translation technologies—an aspect that resonates with the fact that the two translators who accepted to carry out the task using the Excel files were the only ones already familiar with translation memories.
33We feel, however, that this hybrid approach goes in the direction of a more responsible use of MT and translation technology by providing alternatives through a process that starts from a human translation and is controlled by a human reviser. More research is of course needed in order to test its validity. First of all, we need participants who have been involved in comparative revision, for instance, translators who have revised their own work. Participants with previous experience with CAT tools might also find it easier to assess the method’s usability. A larger variety of genres and language pairs could also be included. As pointed out by one interviewee, the growing market for e-books and audiobooks has raised publishers’ interest in revision, a phenomenon that also needs further research. Therefore, in addition to testing the usability of technological solutions, there is a demand for an up-to-date analysis of the different perspectives provided by those who control publishing policies and those who actually reprocess translations.