Multi-Retranslation corpora: Visibility, variation, value, and virtue

....................................................................................................................................... Abstract Variation among human translations is usually invisible, little understood

German versions of a work by Shakespeare.Initial findings lead to more questions than answers.

Introduction
Our project began with a simple observation and an intuition.The observation: in any set of multiple translations in a given language, variation among them varies through the course of the text.Some text units or chunks (at any level from word, say, up to chapter or character part in a play) are more variously translated than others.The intuition: this variation can be used to project an annotation onto the translated text, indicating where and how the extent of translation variation varies.This is the essence of our online system.It uses a 'Translation Array' (a parallel multi-translation corpus, aligned to a 'base text' of the translated work) to achieve 'Version Variation Visualization'.Here, 'version' encompasses any text which can be at least partly aligned with others.But the website strapline is: 'Explore great works with their world-wide translations'. 1  If multiple translations of a work exist, then the work is enduringly popular and/or prestigious, canonical or classic, in the translating culture: typically 'great works' of scripture, literature, philosophy, etc. 2 Interest in comparing such works' multiple translations is surprisingly limited.Some large aligned retranslations corpora are publicly accessible online (works of scripture), 3 but user access is limited to two parallel texts, and no analytic tools are provided.No similar resources exist for any secular works at all, yet.This reflects the notorious 'invisibility' of translators and translations in general (Venuti, 2008).A key aim of our project is to make them visible.
Retranslations are successive translations of the 'same' source work, often somehow dependent on precursor (re)translations.The source works concerned are mostly unstable texts in their original language: what translators translate varies and changes.And so does how they do it.The gamut runs from word-for-word renderings to very free adaptations or rewritings with little obvious relation to the source.Relay translation-via a third language-introduces further variation.If translations are reprinted or otherwise re-used, they tend to be changed again.Venuti (2004) argues that retranslations (more than most translations) 'create value' in the target culture. 4A first translation of a foreign work creates awareness of it.If retranslations follow, the work becomes assimilated to the target culture.If retranslations multiply, each both reinforces the value and status of the work in the target culture, and extends the range of competing interpretations surrounding it.Retranslations therefore throw up questions going well beyond linguistic and cultural transfer, concerning 'the values and institutions of the translating culture', and how these are defended, challenged, or changed (Venuti, 2004, p. 106).
Within Translation Studies, 'retranslation studies' is underdeveloped, despite its fundamental importance for translation, linguistics, and communication, as well as comparative, transnational cultural studies.As Munday (2012) argues, retranslations are important resources, because no single utterance or text exists in isolation from alternative forms it might have taken.Any extant text is surrounded by a 'penumbra' of 'unselected forms' (Munday, 2012, p. 13, citing Grant, 2007, pp. 183-4); so any translation is surrounded by 'shadow translations' (Johannson, 2011, p. 3, citing Matthiessen, 2001, p. 83).Sets of translations by different translators (or the same translators at different moments) make visible at least some otherwise unselected forms.This offers scope for studying 'the value orientations that underlie these selections' (Munday, 2012, p. 13).Our project seeks to go even further: from the how and why of variation among translations, back to the varying capacity of the translated text to provoke variation.
The article is organized as follows: Section 2 reviews related work, including statistical studies in translation variation.Section 3 presents our software project, covering our Aligner, Corpus Overviews (including stylometric analysis), and our key innovation: an interface deploying 'Eddy Digital Scholarship in the Humanities, Vol.32, No. 4, 2017 and Viv' algorithms to explore translation variation.Section 4 presents findings of experiments using the software.Section 5 offers concluding comments.

Related Work
There has been little digital work on larger retranslations corpora, involving works of wide intrinsic interest, and none designed to facilitate access to multiple translations, and the translated work, together with algorithmic analyses.Ja ¨nicke et al. (2015) take in some ways a similar approach, but their 'TRAViz' interface offers a very different mode of text visualization, is monolingual (shows no translated text), and works best with more limited variation and shorter texts (see Section 3.3).Lapshinova-Koltunski (2013) describes a parallel multi-translation corpus designed to support computational linguistic analyses of differences between professional translations, student translations, Machine Translation (MT) outputs, and edited MT outputs.Shei and Pain (2002) proposed a similar parallel corpus, with an interface designed for translator training.These projects only offer access to filtered segments of the text corpus, and do not envisage exploring variation among retranslations.Altintas, Can, and Patton (2007) used two timeseparated (c.1950, c.2000) collections of published translations of the same seven English, French, or Russian literary classics into Turkish, to quantify aspects of language change.This raises the question whether such translations 'represent' their language.Corpus-based Translation Studies (Baker, 1993;Kruger et al., 2011) has established that translated language differs from untranslated language.We also know from decades of work in Descriptive Translation Studies (Morini, 2014;Toury, 2012) that retranslations vary for complex genre-, market-, subculture-specific and institutional factors, and individual psychosocial factors, involving the translators and others with a hand in the work (commissioners, editors), and their uses of resources including source versions and prior (re)translations.
There is no consensus on defining such factors and their interrelations.The conclusion of a manual analysis of eight English versions of Zola's novel Nana is typically vague: (. ..) specific conditions (. ..) explain the similarities and differences (. ..).The conditions comprise broad social forces: changing ideologies and changing linguistic, literary, and translational norms; as well as more specific situational conditions: the particular context of production and the translator's preferences, idiosyncrasies, and choices.(Brownlie, 2006, p. 167) The basic lesson is that translation is a humanities subject.Translators are writers.As Baker warns: Identifying linguistic habits and stylistic patterns is not an end in itself: it is only worthwhile if it tells us something about the cultural and ideological positioning of the translator, or of translators in general, or about the cognitive processes and mechanisms that contribute to shaping our translational behaviour.We need then to think of the potential motivation for the stylistic patterns that might emerge from this type of study.(Baker, 2000, p. 258) Her comment is cited by Li et al. (2011, p. 157), in their computationally assisted study of two English translations of Xueqin Cao's Hongloumeng. 5They conclude: corpus-assisted translation research can go beyond proving the obvious or the already known as long as meta-or para-texts are available for the analysis.The extent and depth of such analysis of course depends on the amount of information available in the form of meta-or other texts.(Li et al., 2011, p. 164) Genuine understanding of cultural materials requires knowledge and critical understanding of many other materials, to assess how multi-scale human factors shape texts and the effects they have (had) in their cultural world.
Non-digital studies in retranslation underline the importance of such shaping factors.Deane-Cox (2014) andO'Driscoll (2011) both recently investigated large sets of English retranslations of 19th-century French novels.They detail at length the historical contexts of each retranslation, its production and reception, and analyse short samples linguistically or stylistically.Deane-Cox's overall argument disproves the 'Retranslation Hypothesis' put forward by Antoine Berman (1990, p. 1).Berman argued that over time, successive retranslations should tend to translate the source text more accurately.In fact-as we will see-this may hold for a first few retranslations, but when they multiply, the hypothesis no longer holds.This is partly because retranslators who come late in a series must be more inventive, to distinguish their work from that of precursors and rivals.The desire for distinction is a great motivator (Mathijssen, 2007;Hanna, 2016).Critical translation studies pays close attention to such specific contextual factors, viewing each translation as an act of intervention in a particular moment in a particular place in the geographical and social world, and a trace of a translator's (and associated agents') both conscious and unconscious choices (Munday, 2012, p. 20).As Munday argues, translation is essentially an evaluative act.Translator's decisions are based on evaluations of the source text, of the implicit values of its author and intended audience, and of the expectations and values of the intended audience of the translation.

Statistical Studies
Statistical studies of differences between translations confirm this perspective, and also rain on the MT parade.They show that variation is greatest both in the most semantically significant units of a text, and in the units which are most expressive of values and affect.Babych and Hartley (2004) measured the stability of alternative translations at word and phrase level in English versions of 100 French news stories by two professional translators.They found a strong statistical correlation between instability and the scores of linguistic items in the source text for salience (tf.idf score) or significance (S-score; see Babych et al., 2003).The more important an item is for a text's meaning, the less translators tend to agree about translating it (though each one is consistent in using their selected terms).Babych and Hartley deduce that 'highly significant units typically do not have ready translation solutions and require some ''artistic creativity'' on the part of translators', and that this necessary 'freedom' makes translation fundamentally '''non-computable'' or ''non-algorithmic''' (Babych andHartley, 2004, p. 835, citing Penrose, 1989).They conclude that there are: fundamental limits on using data-driven approaches to MT, since the proper translation for the most important units in a text may not be present in the corpus of available translations.Discovering the necessary translation equivalent might involve a degree of inventiveness and genuine intelligence.(Babych and Hartley, 2004, p. 836) Munday (2012, pp. 131-54) studied seventeen English translations of an extract from a story by Jorge Luis Borges: two published translations and fifteen commissioned from advanced trainee translators.Four in five lexical units varied.Invariance was associated with 'simple, basic, experiential or denotational processes, participants and relations' (p.143).Variation mainly occurred in 'lexical expression of attitude', i.e. affect/emotion, judgment/ ethics, appreciation, or evaluation (p.24).Variation was greatest at 'critical points', where 'attitude-rich' words and phrases 'carry the attitudinal burden of the text' and communicate 'the central axiological values of the protagonists, narrator or writer' (p.146)-again, in effect, the semantically most significant items.
Translations vary most at points of greatest semantic and evaluative/attitudinal salience.MT has a long way to go, then.Its problems include identifying attitude, affect, or evaluation in a text to be translated.In a chapter on MT and pragmatics, Farwell and Helmreich (2015) discuss lexical and syntactic differences in 125 Spanish newswire articles translated into English by two professional translators: 40% of units differed, and 41% of differences could be attributed to the translators' different 'assumptions about the world' (rather than assumption-neutral paraphrasing, or error).One example is this headline: Acumulacio ´n de vı ´veres por anuncios sı ´smicos en Chile Translation 1: Hoarding caused by earthquake predictions in Chile Translation 2: Stockpiling of provisions because of predicted earthquakes in Chile (Farwell and Helmreich, 2015, p. 171) The translations make vastly different ideological, political, evaluative assumptions.'Hoarding' suggests a panicky, irrational population, responding to rumours of an unlikely event.'Stockpiling' (by the population, or the civil authorities?) is a prudent response to credible (scientific experts'?)warnings.It is impossible-without 'meta-or para-texts'-to disentangle whether the translators impute different values to the mind of the source text creator, or to its intended readers, or to the anticipated readers of the target text, and/or whether they express their own psychological and ideological values.'Acumulacio ´n', here, has major evaluative implications which could not be predicted without area-specific political and economic expertise.Perhaps a multi-retranslation corpus could be used to discover which items provoke variation, as a proxy for such knowledge?If not, what would it discover?
3 Project Description A multi-retranslation corpus will contain versions of various kinds; complete, fragmentary, edited, adapted versions; versions derived from (a version of) the original-language translated work, or from intermediaries in the translating language, and/or other languages; versions in various media; for various audiences (popular, scholarly, restricted); in mono-, bi-, or plurilingual formats; from various periods and places; produced and received under various economic, political, institutional, and cultural-linguistic conditions.An obvious lay question is: Which one is best?But the problem is already clear: By what criteria, or whose, do we judge?Models for assessing professional translations (House, 1997) are predicated on full and precise rendering of the source, but work less well with creative genres, where such 'fidelity' is often subordinated to effect in the target culture.Retranslations of poetry, plays, novels, religious, or philosophical works can be very successful (i.e.'good', for many people) without being at all complete or accurate.A related question is: Why do most retranslations have brief lives (just one publication, or media or performance use), while others-backed by some institutional authority-become canonical, and have many editions, revisions, and re-uses, over generations?Does the answer lie in linguistic, textual qualities of the translation, measured in terms of its relation to the original work?Or in some qualities of it, measured in relation to alternative versions or other target culture corpora?Or does it lie solely in institutional factors?
Our project does not comprehensively address these questions.It grew out of a particular piece of translation criticism, and the intuition that digital tools could be developed to explore patterns in variation among multiple (re)translations, in themselves, in relation to target cultural contexts, and in relation to the translated work.Before knowing any of the above-mentioned studies, Cheesman wanted to find ways to compare a large collection of German translations and adaptations of Shakespeare's play, The Tragedy of Othello, The Moor of Venice (see corpus overviews in section 3.2 below). 6His interest was as a researcher in German and comparative literature and culture.He had worked on a recent, controversial version of Othello (Cheesman, 2010), and wondered how it related to others.He manually examined over thirty translations (1766-2010) of a very small sample: a fourteen-word rhyming couplet, a 'critical moment' which is rich in affect, evaluation, and ambiguity (Cheesman, 2011). 7His study showed how differences among the translations traced a 250-yearlong conversation about human issues in the workgender, race, class, political power, interpersonal power, and ethics.Could digital tools help to explore such questions and communicate their interest to a wider public?The couplet he had selected was clearly more variously translated than most passages in the play.So he wondered if we could devise an algorithmic analysis which would identify all the most variously translated passages, to steer further research.
A proof-of-concept toolset ('Translation Array Prototype') was built, using as test data a corpus of thirty-eight hand-curated digital texts of German translations and adaptations of part of the play: Othello, Act 1, Scene 3.This is about 3,400 continuous words of the play's 28,000, in English: 392 lines and 92 speeches (in Neill's 2006 edition).The restricted sample size was due to restricted resources for curating transcriptions, and translation copyright limitations.Versions were procured from libraries, second-hand book-sellers, and theatre publishers (who distribute texts not available through the book trade).Digital transcription stripped out original formatting and paratexts (prefaces, notes, etc).The transcriptions were minimally annotated, marking up speech prefixes, speeches, and stage directions.The brief for the programmers (Flanagan and Thiel) was to build visual web interfaces enabling the user to: align a set of versions with a base text and so create a parallel multi-version corpus; 8 obtain overviews of corpus metadata and aligned text data; navigate parallel text displays; apply an algorithmic analysis to explore the differing extent to which base text segments provoke variation among translations; customize this analysis and create various forms of data output to support cultural analyses.

Aligner
An electronic Shakespeare text was manually collated with a recent edition, to give us a base text inclusive of historic variants. 9Then we needed to align it segmentally with the versions.Existing open tools for working with text variants (e.g.Juxta collation software) 10 lack necessary functionality; so do existing computer-assisted translation tools; perhaps such software could be adapted; at any rate we built a web-based tool from scratch.The developer, Flanagan, explains its two main components: Ebla: stores documents, configuration details, segment and alignment information, calculates variation statistics, and renders documents with segment/variation information.Prism: provides a web-based interface for uploading, segmenting and aligning documents, then visualizing document relationships.Areas of interest in a document are demarcated using segments, which also can be nested or overlapped.Each segment can have an arbitrary number of attributes.For a play these might be 'type' (with values such as 'Speech', 'Stage Direction'), or 'Speaker' (with values such as 'Othello', 'Desdemona'), and so on.
(Flanagan in: Cheesman et al., 2012) Hand-or machine-made attributes such as 'irony', 'variant from source x', 'crux', 'body part y', 'affect z', 'syllogism', 'trochee', and 'enjambement' are equally possible.But all would require time-consuming tagging.In fact, we have worked only with 'type: Speech'.Segment positions are stored as character offsets within documents, and texts can be edited without losing this information (transcription errors keep being discovered).Segmented documents are aligned in an interactive WYSIWYG tool, where an 'auto-align' function aligns all the next segments of specified attribute.For Othello, every speech prefix, speech and 'other' string is automatically pre-defined as a segment of that type.Any string of typographic characters in a speech can be manually defined as a segment and aligned.Thiel and colleagues at Studio Nand built visual interfaces on top of Prism, including paralleltext views tailor-made for dramatic texts (base text and any translation), and the 'Eddy and Viv' view discussed below (Section 4).Thiel (2014b) documents the design process.He also sketched a scalable, zoomable multi-parallel view of base text and all aligned versions, an overview model which remains to be developed as an interface for combined reading and analysis (Thiel, 2014a). 11

Corpus overviews
Visual overviews of a corpus support distant readings of text and/or metadata features.We devised three.An online, interactive time-map of historical geography shows when and where versions were written and published (performances are a desideratum); it identifies basic genres (published books for readers, books for students, theatre texts), and provides bio-bibliographical information (Thiel, 2012).A stylometric diagram is discussed in Section 3.2.2 (Fig. 2).'Alignment maps' depict the information created by segment alignment (Fig. 1).Digital Scholarship in the Humanities, Vol.32, No. 4, 2017

Alignment maps
Alignment maps, developed by Thiel, are 'barcode'type maps which show how a translation's constituent textual parts (here: speeches) align with a similar map of the base text.Figure 1 shows thirty-five such maps, in chronological sequence.Each left-hand block represents the English base text of Othello 1.3, the right-hand block represents a German text, and the connecting lines represent alignments in the system.Within each block, horizontal bars represent speeches (in sequence top to bottom) and thickness represents their length, measured in words; Othello's longest speech in the scene (and the play) is highlighted.Small but significant differences in overall length can be noticed: translations tend to be longer than the translated texts, so it is interesting to spot versions which are complete yet more concise, such as Gundolf (1909).We can see which versions, in which passages, make cuts, reduce, expand, transpose, or add material which could not be aligned with the base text.In the centre of the figure, the German translation (Felsenstein and Stueber, 1964) of the Italian libretto (by Boito) of Verdi's opera Otello (1887) is a good example of omission, addition, and transposition.Omissions and additions are also evident in the recent stage adaptations on the bottom line.Zimmer (2007), like Boito, assigns Othello's long speech to multiple speakers.In our online system, these maps serve as navigational tools alongside the texts in Thiel's parallel-text views.Each bar representing a speech is also tagged with the relevant speech prefix, so any character's part can be highlighted and examined.Aligned segments are rapidly, smoothly synched in these interfaces, assisting exploratory bilingual reading.

Stylometric network diagram
Figure 2 depicts a stylometric analysis of relative Most Frequent Word frequencies in 7,000-word chunks of forty German versions of Othello, carried out by Rybicki using the Stylo script and the Gephi visualization tool. 12The network diagram shows (1) relations of general similarity between versions, represented by relative proximity (clustering), and ( 2) similarities in particular sets of frequency counts, represented by connecting lines; their thickness or strength represents degree of similarity.These lines (edges) can indicate intertextual relations: dependency of some kind, including potential plagiarism.Directionality can be inferred from date labels on nodes.For example, the version by Bodenstedt (1867) (near top centre) was revised in the strongly connected version by Ru ¨diger (1983).This confirms data on his title page.Other results, as we will see, are more surprising: spurs to further research.The x/y axes are not meaningful.The analysis involves hundreds of counts using differing parameters: the diagram is a design solution to the problem of representing high-dimensional data in a twodimensional plane.Removing or adding even one version produces a different layout and can re-arrange clusters.Moreover, the analysis process is so complex that we cannot specify which text features lie behind the results.Broadly, though, the diagram can be read historically, right to left: a highly formal poetic theatre language gives way to increasingly informal, colloquial style.
Nine versions are revisions, editions or rewritings of the canonical translation by Baudissin (originally 1832, in the famed 'Schlegel-Tieck' Shakespeare edition; see: Sayer, 2015).Most are quite strongly connected and closely clustered, but the apparent stylometric variety is a surprise.The long, weak line connecting the cluster to the heavily revised stage adaptation by Engel (1939) (upper left) is to be expected, but the length and weakness of the connection with Wolff's (1926) published edition (lower right) is more of a surprise.His title page indicated a modestly revised canonical text, but stylometry suggests something more radical is going on. 13 Above all, this analysis reveals the salience of historical period.Distinct clusters are formed by all the early C19 versions (mid-right), arguably all the late C19 versions (top), most of the late C20 versions (lower left), and all the C21 versions (far left).The C21 versions are all idiosyncratic adaptations (cf.Fig. 1, bottom line).It is surprising to see how similar they appear, in stylometric terms, relative to the rest of the corpus.And what do the strong links among them indicate?Mutual influence, plagiarism, common external influence?What about the lines leading from Gundolf (1909) (low centre) across to Swaczynna (1972), to Laube (1978), to Gu ¨nther (1995)?Gu ¨nther is the most celebrated living German Shakespeare translator: do these lines trace his debts to less famous precursors?Period outliers are also interesting.Zeynek (?-1948) appears to be writing a C19 style in the 1940s.The unknown Schwarz (1941) is curiously close to the famous Fried (1972).Rothe (1956) (extreme bottom left) is writing in a late C20 style in the 1950s.This throws interesting new light on the notorious 'Rothe case' of the Weimar Republic and Nazi years: he was victimized for his 'liberal', 'modern' approach to translation (Von Ledebur, 2002).
Genre is salient, too.A very distinct cluster, bottom right, includes all versions designed for study and written in prose (rather than verse).This includes our two earliest versions (1766 and 1779) and two published 200 years later (1976,1985).Strongly interconnected, weakly connected with any other versions, this cluster demonstrates the flaw in the approach of Altintas et al. (2007).
Differences in the use of German represented by distances across the rest of graph cannot be due to any general historical changes in the language.They reflect changes in the specific ways German is used by translators of Shakespeare for the stage, and/or for publications aimed at people who want to read his work for pleasure.

The 'Eddy and Viv' interface
Overviews are invaluable, but the core of our system is a machine for examining differences at small scale.The machine implements an algorithm we called 'Eddy', 14 to measure variation in a corpus of translations of small text segments.Eddy's findings are then aggregated and projected onto the base text segments by the algorithm 'Viv' ('variation in variation').In an interface built by Thiel, on the basis of Flanagan's work, users view the scrollable base text (Fig. 3: left column) and can select any previously defined and aligned segment: this calls up the translations of it, in a scrollable list (Fig. 3: right columns).The list can be displayed in various sequences (transition between sequences is a pleasingly smooth visual effect) by selecting from a menu: order by date; by the translator's surname; by length; or (as shown in Fig. 3) by Eddy's algorithmic analysis of relative distinctiveness.Eddy metrics are displayed with the translations, and also represented by a yellow horizontal bar which is longer, the higher the relative value.
We defined 'segment', by default, as a 'natural' chunk of dramatic text: an entire speech, in semiautomated alignment.Manual definition of segments (any string within a speech) is possible, but defining and aligning such segments in forty versions is time-consuming.In future work we intend to use the more standard definition: segment ¼ sentence (not that this would simplify alignment, since translation and source sentence divisions frequently do not match).Eddy compares the wording of each segment version with a corpus word list: here the corpus is the set of aligned segment versions.No stop words are excluded; no stemming, lemmatization, or parsing is performed.Flanagan explains how the default Eddy algorithm works: Each word in the corpus word list [the set of unique words for all versions combined] is considered as representing an axis in N-dimensional space, where N is the length of the corpus word list.For each version, a point is plotted within this space whose coordinates are given by the word frequencies in the version word list for that version.(Words not used in that version have a frequency of zero.)The position of a notional 'average' translation is established by finding the centroid of that set of points.An initial 'Eddy' variation value for each version is calculated by measuring the Euclidean distance between the point for that version and the centroid.Flanagan in Cheesman et al. (2012-13) This default Eddy algorithm is based on the vector space model for information retrieval.Given a set S of versions {a, b, c . ..}where each version is a set of tokens {t 1 , t 2 , t 3 . . .t n }, we create a set U of unique tokens from all versions in S (i.e. a corpus word list).For each version in S we construct vectors of attributes A, B, C . . .where each attribute is the occurrence count within that version of the corresponding token in U, that is: We construct a further vector Z to represent the centroid of A, B, C . . .such that Then, for a version a, the default Eddy value is calculated as: This default Eddy formula is used in the experiments reported below, coupled with a formula for Viv as the average (arithmetic mean) of Eddy values.Other versions of the formulae can be selected by users, 15 e.g. an alternative Eddy value based on angular distance, calculated as:

IAIIZI
Work remains to be done on testing the different algorithms, including the necessary normalization for variations in segment length. 16 Essentially, Eddy assigns lower metrics to wordings which are closer to the notional average, and higher metrics to more distant ones.So, Eddy ranks versions on a cline from low to high distinctiveness, or originality, or unpredictability.It sorts commonor-garden translations from interestingly different ones.
Viv shows where translators most and least disagree, by aggregating Eddy values for versions of the base text segment, and projecting the result onto the base text segment.Viv metrics for segments are displayed if the text is brushed, and relative values are shown by a colour annotation (floor and ceiling can be adjusted).As shown in Fig. 3, the base text is annotated with a colour underlay of varying tone.Lighter tone indicates relatively low Viv (average Eddy) for translations of that segment.Darker tone indicates higher Viv.Shakespeare's text can now be read by the light of translations (Cheesman, 2015).
Sometimes it is obvious why translators disagree more or less.In Fig. 3, Roderigo's one-word speech 'Iago -' has a white underlay: every version is the same.The Duke's couplet beginning 'If virtue no delighted beauty lack. ..' (the subject of Cheesman's initial studies), has the darkest underlay.As we knew, translators (and editors, performers, and critics) interpret this couplet in widely varying ways.In the screenshot, the Duke's couplet has been selected by the user: part of the list of versions can be seen on the right.MTs back into English are provided, not that they are always helpful.
Unlike the TRAViz system (Ja ¨nicke et al., 2015), ours does not represent differences between versions in terms of edit distances, and translation choices in terms of dehistoricized decision pathways.Our system preserves key cultural information (historical sequence).It can better represent very large sets of highly divergent versions.The TRAViz view of two lines from our Othello corpus (Ja ¨nicke et al., 2015, Figure 17) is a bewilderingly complex graph.With highly divergent versions of longer translation texts, TRAViz output is scarcely readable.Crucially there is no representation of the translated base text.The Eddy and Viv interface is (as yet) less adaptable to other tasks, but better suited to curiosity-driven cross-language exploration. 17

Eddy and 'Virtue? A fig!'
To illustrate Eddy's working, Table 1 shows Eddy results, in simplified rank terms ('high', 'low', or unmarked intermediate), for thirty-two chronologically listed versions of a manually aligned segment with a very high Viv value: 'Virtue?A fig!' (Othello 1.3.315).An exclamation is always, in Munday's terms, 'attitude-rich', burdened with affect; this one is a 'critical point' for several reasons.'Virtue' is a very significant term in the play, and crucially ambiguous: in Shakespeare's time it meant not only 'moral excellence' but also 'essential nature', or 'life force', and 'manliness'. 18The speaker here is Iago, responding to Roderigo, who has just declared that he cannot help loving the heroine, Desdemona: '. . . it is not in my virtue to amend it'.Roderigo means: not in my nature, my power over myself, my male strength.But Iago's response implies the moral meaning, too.Then,  The lowest and highest seven Eddy rankings are indicated.Eddy's lowest-scoring translation is 'Tugend?Quatsch' (#16, #26).'Tugend' is the modern dictionary translation of (moral) 'virtue'.'Quatsch' is a harmless expression of disagreement: a bowdlerized translation (bowdlerization is clear in most versions here). 19The Eddy score is low because most translations (until 1985) use 'Tugend' and several also use 'Quatsch'.Eddy's highest score is for 'Charakter?Am Arsch der Charakter!' (#30).This is Zaimoglu's controversial adaptation of 2003, with which Cheesman's work on Othello began (2010).No other translation uses those words, including the preposition 'am' and article 'der'.'Charakter' accurately translates the main sense of Shakespeare's 'virtue' here, and 'Arsch' fairly renders 'A fig!' This is among the philologically informed translations of 'virtue' (as 'energy', 'strength', 'power'), a series which begins with Schwarz (1941) (#13).It is also among the syntactically expansive translations, with colloquial speech rhythms, which begin with Zeynek (?-1948) (#15). 20 Both series become predominant following the prestigious Fried (1972) (#21).
Reading versions both historically and with Eddy, in our interface, makes for a powerful tool.
Here the historical distribution of Eddy rankings confirms what we already know about changes in Shakespeare translation.The lowest mostly appear up to 1926.The highest mostly appear since 1972 (recall Figure 2: lower left quadrant).Ranking by length in typographical characters is not often useful, but with such a short segment its results are interesting, and similar to Eddy's.Most of the shortest are up to 1947, and most of the longest since 1972: that shift towards more expansive, colloquial translations, again.
Similar historical Eddy results are found for many segments in our corpus.An 'Eddy History' graph, plotting versions' average Eddy on a timeline, can be generated: it shows Eddy average rising in this corpus since about 1850.This may be a peculiarity of German Shakespeare.It may be an artefact of the method.But it is conceivable that, with further work, the period of an unidentified translation might be predicted by examining its Eddy metrics.
Eddy and Viv results for any selected segments, based on the full corpus or a selected subset of versions, can be retrieved and explored in several forms of chart, table, and data export.The interactive 'Eddy Variation' chart, for example, facilitates comparisons between one translator's work and that of any set of others (e.g.her precursors and rivals).It plots Eddy results for selected versions against segment position in the text; any version's graph can be displayed or not (simplifying focus on the translation of interest); when a node is brushed, the relevant bilingual segment text is displayed.
Eddy's weaknesses are evident in Table 1, too.It fails to highlight the only one-word translation (#29), or the one giving 'fig ' for 'fig' (#19), or the one with the German equivalent of 'fuck' (#20), expressing the obscenity which remains concealed from most German readers and audiences.We still need to sort ordinary translations from extraordinary and innovative ones in more sophisticated ways.Eddy also fails to throw light directly on genetic and other intertextual relations.Some are indicated in the 'Intertexts' column in Table 1: the probable influence of some prestigious retranslations is apparent in several cases, as is the possible influence of some obscure ones.Such dependency relations require different methods of analysis and representation.Stylometric analysis (Section 3.2.2) provides pointers.More advanced methods must also encompass negative influence, or significant non-imitation.Table 1 shows-and this result is typical too-that the canonical version (#5), the most often read and performed German Shakespeare text from 1832 until today, is 'not' copied or even closely varied.That is no doubt because of risk to a retranslator's reputation.Retranslators must differentiate their work from what the public and the specialists know (Hanna, 2016).
The tool we built is a prototype.

'Viv' in Venice
An initial Viv analysis of Othello 1.3, involving all the ninety-two natural 'speech' segments, was reported (Cheesman, 2015). 21It found that the 'highest' Viv-value segments tended to be (1) near the start of the scene, ( 2) spoken by the Duke of Venice, who dominates that scene, but appears in no other, and (3) rhyming couplets (rather than blank verse or prose).There are twelve rhyming couplets in the scene; two are speech segments; both were in the top ten of ninety-two Viv results.
No association was found between Viv value and perceptible attitudinal intensity, or any linguistic features.We did find some high-Viv segments associated with specific cross-cultural translation challenges.Highest Viv was a speech by Iago with the phrase 'silly gentleman', which provokes many different paraphrases.But some lower-Viv segments present similar difficulties, on the face of it.There was no clear correlation.Still, four hypotheses emerged for further research.
Hypothesis 1: Based on rhyming couplets having high Viv-value: retranslators diverge more when they have additional poetic-formal constraints. 22 Hypothesis 2: Based on finding (1) above: retranslators diverge more at the start of a text or major chunk of text (i.e. at the start of a major task).
Hypothesis 3: Based on finding (2) above: retranslators diverge more in translating a very salient, local text feature in a structural chunk (in this scene: the part of the Duke) and less in translating global text features (e.g.here: Othello, Desdemona, Iago).
Hypothesis 4 relates to 'low' Viv findings.It was somehow disappointing to find that speeches by the hero Othello and the heroine Desdemona, including passages which generate much editorial and critical discussion, had moderate, low, or very low Viv scores.Famous passages where Othello tells his life story and how he fell in love with Desdemona, or where Desdemona defies her father and insists on going to war with Othello, surely present key challenges for retranslators.Perhaps passages which have been much discussed by commentators and editors pose less of a cognitive and interpretive challenge, as the options are clearly established. 23This hypothesis could be investigated by marking up passages with a metric based on the extent of associated annotation in editions and/or frequency of citation in other corpora.For now, we have speculated that the hero's and heroine's speeches in this particular scene do exhibit common attitudinal, not so much linguistic, but dramatic features.In the low-Viv segments, the characters can be seen to be taking care to express themselves particularly clearly; even if very emotional, they are controlling that emotion to control a dramatic situation.Perhaps translators respond to this 'low affect' by writing less differently?But it is difficult to quantify such a text feature and so check Viv results against any 'ground truth'.
There is another possible explanation: in the most 'canonical' parts of the text (here: the hero's and heroine's parts), retranslators perhaps tread a careful line between differentiating their work and limiting their divergence from prestigious precursors. 24Such 'prestige cringe' would relate to the above-mentioned negative influence, or non-imitation of the most prestigious translations (Section 3.4).Precursors act, paradoxically, as both negative and positive constraints on retranslators.
Hypothesis 4: in the most canonical constituent parts of a work, Viv is low, as retranslators tend to combine willed distinctiveness with caution, limiting innovation.
In the initial analysis, the groups of speeches assigned highest and lowest Viv values had suspiciously similar lengths.Clearly the normalization of Eddy calculations for segment length leaves something to be desired.The next and latest analysis focused on segments of similar length to investigate our hypotheses.

'Viv' in two liners
Table 2 shows the grammatically complete two-line verse passages in Othello 1.3, plus prose passages of equivalent length, 25 in Viv value rank order.A subcorpus of twenty translations was selected for better comparability. 26The text assigned to each major character part here is reasonably representative of their overall part in the scene, counted in lines: Brabantio (sample eighteen lines [nine couplets]/ total sixty-one lines) 0. Hypothesis 1 seems to be confirmed, though more work needs to be done to prove it conclusively: high Viv value correlates with poetic-formal constraint.In the column 'Form' in Table 2, blank verse is the default.Unsurprisingly, rhyming couplets appear mostly in the top half of the table, including five of the top ten items.Translators enjoy responding to the formal challenge of rhyming couplets in self-differentiating ways; and they must so respond, or else they very obviously plagiarize, because these items are rare in the text and highly noticeable, for audiences or readers.
Hypothesis 2 is not confirmed: scanning the column 'Running order', there is no sign that translators differentiate their work more at the start of the scene, as they embark on a new chunk of the task.That could have been interesting for psycholinguistic and cognitive studies of translation (Halverson, 2008).
Hypothesis 3 seems to be confirmed, but we need much more evidence to be sure we have discovered a general pattern.Scanning the column 'Speaker', the Duke's segments are more variously translated than those of other speakers.Even if we exclude rhyming couplets, the Duke is over-represented in the upper part of the table.Brabantio and Iago also have some very high-Viv lines, but their segments are distributed evenly up and down the table.Not so with the Duke, who is the salient, local text feature in this scene and no other.Hypothesis 4 also seems provisionally confirmed.Othello is strikingly low-Viv, mostly.Desdemona tends to be low-to mid-Viv.Translations of their parts differ 'less' than other parts, at this scale.Why?We do not know.It could be 'prestige cringe' (Section 4.2).But it could also be specific to this text.Othello in particular refuses 'affect' in this scene, as he does throughout the first half of the play: he is in command of everything, including his emotions.He echoes a much discussed line just spoken by Desdemona ('I saw Othello's visage in his mind', 1.3.250)when he says to the Duke and assembled Senators that he wants her to go to war with him, but: I therefore beg it not, To please the palate of my appetite, Nor to comply with heat-the young affects In me defunct-and proper satisfaction.But to be free and bounteous to her mind: (. ..) (Othello 1.3.258-63)This is one of the play's cruxes-passages which editors deem corrupt and variously resolve (here, 'me' is often changed to 'my', 'defunct' to 'distinct', and the punctuation revised). 27Translators also resolve this passage variously, depending in part on which edition(s) they work with; but-as measured by Viv-not very variously, compared with other passages.Can it be that textual 'affect' is relatively less, because that is the kind of character, the mind, the 'virtue' Othello is projecting?

Concluding Comments
Findings which only confirmed what was already known would be truly disappointing (though we do need some such confirmation, to have any faith in digital tools).Digital literary studies should provoke thought.A classic example is Moretti's discovery of a rhythm of 25-30 years in the emergence and disappearance of C19 novelistic genres, which he uneasily ascribed to a cycle of biological-sociocultural 'generations': I close on a note of perplexity: faute de mieux, some kind of generational mechanism seems the best way to account for the regularity of the novelistic cycle-but 'generation' is itself a very questionable concept.Clearly, we must do better.(Moretti, 2003, p. 82) So too with 'Translation Arrays' and 'Version Variation Visualization': we must do better.
We wanted to demonstrate that this sort of approach opens up interesting possibilities for future research. 28Of course one big difference between Moretti's work and ours so far is one of scale.His team works with tens or hundreds of thousands of texts and metadata items.We are working with a few dozen versions of one play, in one target language, because that is what we have got, 29 and only a fragment of the play, because we chose to make the texts publicly accessible, which entails copyright restrictions (and some expense).Our approach requires time-consuming text curation (correction of digital surrogates against page images), 30 permission acquisition, and manual segmentation and alignment processes (more sophisticated approaches including machine learning will speed these up). 31 Moretti experimentally 'operationalizes' predigital critical concepts such as 'character-space' or 'tragic collision' (Moretti, 2013), by measuring quantities in texts: digital proxies or analogues.Eddy and Viv, on the other hand, are measuring relational corpus properties which have no obvious pre-digital analogue.What could they be proxies for?Eddy makes visible certain kinds of resemblance and difference, certain sequences, patterns of influence and distinctiveness.Critically understanding these still depends on understanding 'para-and meta-texts' (Li, Zhang and Liu, 2011).Viv's contribution is even less certain: we won't know whether its results correspond to anything 'real' about translated texts' qualities, or those of translations, or of translators, until we have studied many more cases.

Fig. 3
Fig. 3 Eddy and Viv interface (Colour online) phrase 'A fig!' is gross sexual innuendo.'Fig' meant vagina.The expression derives from Spanish and refers to an obscene hand gesture: intense affect (see Neill, 2006, p. 235).(The expression 'I don't give/care a fig!' was once commonplace, and often used euphemistically for 'fuck', a word Shakespeare never uses.) Intertexts: (P) ¼ prestigious, influential.Digital Scholarship in the Humanities, Vol.32, No. 4, 2017the Eddy is admittedly imperfect.But its real virtue lies in the power it gives to Viv, enabling us to investigate to what

Table 2 '
Viv values' in two liners in Othello 1.3 generated by twenty German versions