TRANSLATION ALIGNMENT

Aligning words and phrases in a translation with its source text has a long tradition. Conventional translations can include untranslated terms and English approximations side-by-side to emphasise that the source term cannot be properly translated (e.g. the opening of the Iliad: ‘Anger [mēnis], goddess, sing it, of Achilles, son of Peleus— ₂ disastrous [oulomenē] anger that made countless pains [algea] for the Achaeans’)Footnote ⁵. Interlinear translations for Greek and Latin texts appeared in the nineteenth century and mainly served readers who were forced to study these languages to fulfil various academic requirements and wanted to learn as little grammar as possible (e.g. fig. 1).

Figure 1.

Opening of Xenophon's Anabasis, in: T. Clark, The Anabasis of Xenophon: with an interlinear translation [1859], p. 9 (https://hdl.handle.net/2027/uc2.ark:/13960/t1sf2p143?urlappend=%3Bseq=15).

Contemporary translation alignment has developed along three complementary pathways. First, automatic methods generate word and phrase level alignments for source texts and translations that have been aligned at a coarse level, such as a page, paragraph or sentence. Although automatic alignment at the word and phrase level has a substantial error rate, the ability to process large bodies of textual data means that meaningful signals emerge from the noise. Automatic alignment is a fundamental tool for analysing translations at scale.

Second, manual alignments between source texts and translations can shed surprising light on the relationship between the two. The figure below provides an example from Persian poetry, a language with which many (and probably most) readers of this piece are not familiar. The underlined English words in figure 2 do not have equivalents in the Persian source.

Figure 2.

Underlined words do not correspond to anything in the Persian and add a dimension of Neoplatonic allegory.Footnote ⁶

The translator, Henry Wilberforce Clarke, has freely, and without explanation, added extensive religious language to convert a poem about drinking and sexual desire into a Neoplatonic allegory. No one without an understanding of the Persian text could have seen those additions in a traditional printed text. Anyone with a knowledge of English, whether or not they know Persian, can see words in the aligned English translation that have no counterparts in the Persian.

Third, translations can be designed from the start to align with and shed light upon the original source text. Such translations can serve as intermediaries for those who wish to understand more deeply the relationship between a more or less free literary translation and the original source text. Such translations can also serve those who wish to learn the language and can intentionally make particular features more prominent.

Figure 3 shows a born-digital translation, developed by A. Parrish and G. Crane, for the opening of Odyssey 5. The unaligned words (shown in red) now more narrowly reflect linguistic features of Greek (e.g. Greek does not need to include possessive pronouns, and the definite article is far less commonly used in Homer than in later Greek). The selected alignment also shows an effort to illustrate two grammatical features of the verb о̄rnuth’ (from ornumi, ‘to rouse, set in motion’): the verb form is middle (hence ‘to rouse herself’) and an imperfect (‘began to’ suggests that this action takes place over a period of time, though admittedly brief). Those learning Greek should be able to scan through a text and see each middle and each imperfect form glossed in such a way that the linguistic features are prominent.

Figure 3.

Born-digital translation of Odyssey 5.1–2 with unaligned words and selected alignment. Words that do not have equivalents in the other language are red. In the Greek, the only word that does not have an equivalent is δ᾽ in 5.1. The English has added ‘her’ in 5.1 and ‘she’, ‘to the’, ‘gods’ and ‘to the’ in line 5.2.

The earliest parallel text alignment work focused on automatically aligning words in source text and translation. Words that tended to occur in both source text and translation were often translation equivalents, and parallel text analysis was used to support machine translation systems of the time.Footnote ⁷ A decade later, D. Bamman, then a researcher at Perseus, applied this method to the parallel Greek/English and Latin/English corpora available in Perseus at the time. Distinguishing word senses is a notoriously difficult task, and one approach suggested that a good way to identify different word senses was to track places where translators chose very different translation terms. The Latin word oratio can be translated as ‘speech’, when it describes a secular oration, and as ‘prayer’, when it describes a religious act (see fig. 4). Bamman was able to trace the changing frequency of these word senses (and the changing importance of religious vs secular contexts) in Latin sources over more than a thousand years.

Figure 4.

English sense distribution for the Latin word oratio (96,313 instances).Footnote ⁸

Automatic translation alignment only identifies a subset of translation pairs, and it has a substantial error rate (in Bamman and Crane [n. 8] 22% of the proposed translation equivalents for oratio were false); but errors tended to be random, and the overall patterns often emerged clearly from the background noise.

H. Diakoff of the Alpheios Project, however, recognised the potential value of manual alignment of source text and translation, both because the data would be more precise and because the exercise of aligning source text and translation seemed to be a pedagogically useful exercise for language learners, both by seeing how fuzzy the relationship could be between sources and translations and by creating their own translations for alignment with the source text (where learners could see, for example, if they had left words untranslated). The Alpheios Project began developing a manual translation alignment tool in 2009.Footnote ⁹ Over ten years later M. Foradi at the University of Leipzig studied translation alignment (Foradi, n. 6). Two of her findings were particularly suggestive. First, faculty (including one of the co-authors) had been asking students to align an English translation to a source text in an unknown language (Greek or Latin) by using basic scaffolding tools (e.g. the ability to click on any inflected word in a text and see a dictionary entry). She performed an experiment to see how well readers could align a translation to a source text in an unknown language. She provided word/phrase level alignments between Persian and pre-existing, rather free English translations to German students who did not know Persian (but who had studied Arabic and could easily work with the slightly different form of Arabic script used for Persian). These students created alignments to a German translation that were just as accurate as those made by experts in both languages (Persian speakers who were fluent in German). This case study reflects a general use case: a reader uses alignments between the source text and an openly licensed translation to scrutinise the relationship between a new literary translation and the source.

Second, and perhaps more surprisingly, Foradi wanted to see if students learned Persian vocabulary as a by-product of aligning Persian and German. She had a control group of students spend equal time studying vocabulary with flashcards. The flashcard group did better on a vocabulary test on the following day, but the alignment group did better two months later. The alignment exercise thus was less efficient in the short run, but had more lasting cognitive impact. Foradi's work laid the foundation for further research (see below).

In 2016 computer scientist T. Yousef, in collaboration with digital humanist C. Palladino, developed Ugarit, a web-based translation alignment system in which 797 users have created 1,215,828 manual word/phase level alignments in 33,595 texts.Footnote ¹⁰ Formal guidelines now exist for aligning ancient Greek with Latin, English with Portuguese, and Latin with English.Footnote ¹¹ Ugarit benefited from a particularly well-designed user interface, and it also illustrates a crucial functionality for one class of ongoing work. Ugarit can display alignments between a source text and two translations, allowing editors to juxtapose an intentionally literal translation, designed to reveal as much as possible about the Greek source, with a literary translation. This three-text mode can spur a new class of review designed to make the translator's decisions accessible to a much wider audience.

The scale of the curated alignments has had two effects. First, we can study translation pairs with greater precision than was possible with the more extensive, but noisier, alignment data that Bamman was able to generate a decade earlier.

Second, Yousef had enough trusted data to train models for a new generation of automatic alignment between Greek and Latin, Greek and English, and Latin and English. This new generation of automatic alignment also benefits from access to language models that, in turn, can recognise words with similar meanings. Thus, an aligner using such a language model can infer that, if Greek lithos aligns with English ‘rock’, then it can align with English ‘stone’ as well. YousefFootnote ¹² reported that 91.5% of proposed alignments between Greek and English were correct and that the automatic alignments identified 87.3% of all possible correct alignments.Footnote ¹³ Results will vary depending on how close the translation is to the source text and on how fine-grained the intermediate chunks are (e.g. 25-word sentences vs 75-word sections), but the larger picture is clear. We can now generate far more accurate alignments between Greek, Latin and English texts and perform more comprehensive semantic analysis than was possible before. We also can generate automatic alignments that are close enough to be helpful for readers who do not know Greek or Latin.

Using Ugarit poses at least one challenge: once a translation has been entered into Ugarit, it cannot be edited. If readers discover that there was an error in transcribing a pre-existing translation or if they wish to revise a translation of their own while aligning it, they have to start over. A. Parrish (Tufts ’21) and Crane worked on a text-only alignment format, in which Greek words are interspersed with their translations.Footnote ¹⁴ The format is not as visually as attractive as that of Ugarit, but it provides great flexibility for ongoing revision.

Much of the most interesting recent work focuses on applications of translation alignment to language learning. C. Palladino documents the effects of having students align one or more existing translations to a source text, forcing them to focus on each translation much more precisely than would normally be the case. She also described the strategy of having students produce aligned translations of their own, a practice that makes it easier for them to see where they may have missed a word or a phrase and for their instructor to see the precise relationship between source text and student translation.Footnote ¹⁵ Large language models that can recognise synonymous expressions should also now be able to provide immediate feedback, identifying places where students produce semantically improbable translations. F. Shamsian has, as part of a Leipzig Ph.D. dissertation, developed a localisable infrastructure for learning ancient Greek. Her immediate learner language is Persian, but the larger goal is to create a compact, easily translated and localised framework that can support the study of ancient Greek in many languages (especially from outside of Europe). She reports on how, after 30 hours of instruction, Persian speakers were able to critique existing Persian translations of the Iliad and create more accurate representations of the Greek original. The same learners were, at this point, able to begin producing the first direct translations of Plato's Crito into Persian. As a result, we have three finalised translations of Crito under development that each represent a different interpretation or approach to the text.Footnote ¹⁶ Most recently, J. Hilleary studied readers using translation alignment to work with a source text in Estonian, a contemporary non-Indo-European language with which none of the participants were familiar. Half the participants had an English translation aligned at the sentence level, and the other half had access to word and phrase level alignments. All participants were able to generate reasonable answers to fill-in-the-blank (‘cloze’) reading comprehension questions in Estonian. Word-level alignments, as predicted, greatly enhanced performance, and participants were often able not only to identify the appropriate Estonian word, but also to generate new morphological forms in Estonian.Footnote ¹⁷ When asked what additional information they would have found useful, many participants asked for a tool to understand the structure of sentences in the story, a category of data to which we turn next in our discussion of treebanks.

TREEBANKS

Treebanks provide information that supports general linguistic research, but that also helps readers make much fuller use of translation alignments. Treebanks are textual databases that contain multiple linguistic annotations for each word. Consider the opening words of the Homeric Iliad: mēnin aeide, ‘sing the wrath’. The core features are regularised dictionary form (e.g. mēnin is a form of the noun mēnis, ‘wrath’), part of speech (e.g. mēnin designates the feminine accusative singular), syntactic role (mēnin is an object) and dependency (mēnin is governed by aeide, ‘sing!’). They derive their name from the fact that they are a form of databank and because the syntactic analyses are often visualised in graphical form as upside-down trees.

Corpus linguists developed treebanks so that they could extract precise information about linguistic usage (e.g. the frequency subject-verb-object vs subject-object-verb order or different case uses with a given verb). Treebanks allow researchers to explore ideas and quantify their results, moving away from vague terms such as ‘common in poetry’ or ‘typical of late Greek prose’.Footnote ¹⁸

Figure 5 illustrates the raw data behind a treebank created using the Perseus annotation guidelines and represented as XML.Footnote ¹⁹ The subject of the main verb gignontai (which is the root node of sentence) is paides, which is in turn modified by duo (‘two children exist’, ‘there are two children’). The head attribute indicates dependency. The main verb is the root of the tree and has no dependency (thus, for gignontai: head = ‘0’). Other than the root, each word has one and only one head in the default case. (We can extend the syntax by creating additional links if we want to show that one word notionally depends on multiple words as in ‘she saw and read many books’.) Any word can have multiple dependencies: thus, duo, ‘two’, and Dareiou, ‘of Darius’, both modify paides. The CO (for ‘co-ordinated’) relation or co-reference relation indicates that Parysatis, ‘of Parysatis’, is parallel to Dareiou. With data such as this we could answer questions such as which subjects or objects appear with any verb, how often a verb has no object, which adjectives modify which nouns, or how often subjects precede or follow the main verb.

Figure 5.

A treebank automatically produced by GLAUx in the Perseus Dependency Treebank, then edited and converted to the Universal Dependency Framework by the Daphne Treebank repository of Ancient Greek Poetry (https://perseids-publications.github.io/daphne-trees/).

Treebanks can, however, also show readers how each word fits into its sentence in the source text. Learners of ancient Greek and Latin have spent thousands of years answering questions about which word depends on which and why. Treebanks present much of that information in explicit form and allow readers to see the structure of sentences laid out in a diagram. Existing treebanks contain enough information to support a range of queries. Part of speech data allows readers to ask for particular forms (e.g. the distribution of tenses in histories vs drama, the relative frequency of future participles, a form typically used to express purpose, or the frequency of the subjunctive vs the optative over time). The dependencies allow more complex queries (e.g. the number of times that the Greek verb echо̄, ‘to have’, in a given author or genre has an object vs being intransitive and designating a state or complex queries that can retrieve particular classes of conditionals by looking at the moods and tenses of verbs in primary and subordinate clauses). Additional categories of annotation can be added to support a wider range of queries (e.g. distinguishing between nouns that designate animate and inanimate entities or linking each verb in a document with its subject).

Figure 6 shows the automatically annotated opening sentence of Xenophon's Anabasis, produced by the GLAUx Trees and visualised in Beyond Translation. The arguments attached to the base URL (https://beyond-translation.perseus.org/reader/) are:

The identifiers ‘tlg0032’ and ‘tlg006’ identify Xenophon and the Anabasis, while ‘perseus-grc2’ and ‘1.1.1’ specify the edition used and the traditional citation (book 1, chapter 1, section 1).

Figure 6.

A treebank for the opening of Xenophon's Anabasis.

G. Crane first wanted to be able to capture syntactic data when, in 1984, he was composing Morpheus, a rule-based system that analyses inflected Greek and Latin forms. Syntax, however, does not lend itself to rule-based programming. H. Kučera and N. Francis had, however, built the prototype for what Crane needed two decades earlier, creating a one-million-word parsed corpus of English.Footnote ²⁰ Treebanks became more prominent in the 1990s, and in 2002 Crane suggested that a dictionary such as the Cambridge Greek Lexicon should be built on a treebanked corpus.Footnote ²¹ When D. Bamman joined Perseus four years later, his first task was to begin work on a treebank of Latin, for which he received funding from the National Science Foundation.Footnote ²² Bamman found that Czech linguists, working with their own highly inflected language with relatively free word order, had developed methods that were much better suited to languages such as ancient Greek and Latin than the methods developed for languages such as English and French, with relatively thin morphology and rigid word order.

Two other groups proved to be developing treebanks for Greek and Latin. The Proiel Project in Norway created treebanks for multiple language versions of the New Testament for comparative study.Footnote ²³ The Index ThomisticusFootnote ²⁴ had begun treebanking Thomas Aquinas. D. Bamman and M. Passarrotti of the Index Thomisticus worked hard to fashion a nearly identical annotation scheme for both projects based on the Prague Dependency Treebank.Footnote ²⁵ Proiel chose a scheme that differed a bit more, but still remained largely in agreement with the other two. Data from all three projects (and their derivatives) has been successfully combined.

Support from the Alpheios Project made it possible to create treebanks for Homeric and Hesiodic epic. Individuals annotated each sentence separately, and a more senior editor resolved places where the two independent annotations diverged. Each sentence thus has three credits. When we moved on to Greek drama, we identified an individual specialist, F. Mambrini, who created the first treebanks (and a unique edition) for the seven surviving plays of Aeschylus. Funding from the Institute for Museum and Library Services, the Mellon Foundation and the Alexander von Humboldt Foundation in Germany paid for the development of a dedicated Treebank editor, Arethusa, in 2013–2014. The Mellon-funded Perseids Project assumed management of Arethusa,Footnote ²⁶ and the amount of treebanked Greek and Latin expanded over the following years, with more than 1.2 million words of Greek available in the Perseus and Proiel formats.Footnote ²⁷

Two major changes have taken shape in the intervening fifteen years. First, treebanks were used from the earliest stage of work as training data for machine-learning algorithms, and larger bodies of Greek and Latin were automatically treebanked. However, as the amount of training data has expanded and as advances in algorithms and, especially, increases in computing power have made machine-learning systems more capable, the quality of automatically generated treebanks has increased. The GLAUx ProjectFootnote ²⁸ at Leuven has published treebanks in the Perseus Dependency Treebank format with more than 19 million words, and that figure can be increased to include any texts that they choose to process. The task for annotators has shifted from creating treebanks by hand to reviewing and augmenting automatically generated data. In this environment some annotators move away from elaborate editing environments and work directly with annotations in a text-only format (greatly simplifying the needed infrastructure).

Second, a newer annotation scheme, the Universal Dependency (UD) FrameworkFootnote ²⁹, has emerged and achieved much broader adoption. UD was designed from the start for cross-language comparison, and it has developed a substantial community: more than 200 treebanks covering more than 100 languages are available in the UD framework – including 200,000 tokens of Greek converted by Perseus and Proiel each (more than 400,000 in total) into UD. Where those who became familiar with the Perseus Dependency Treebank annotation scheme could easily work with Czech, those who become comfortable with UD can move across more than 100 languages. Put another way, working closely with UD Greek and Latin treebanks provides readers with skills that are far more immediately and widely applicable than mastery of traditional Greek and Latin grammar.

At the same time published UD treebanks attracted the attention of the latest natural language processing pipelines: Stanza and SpaCy.Footnote ³⁰ Because Perseus and Proiel published UD versions for parts of their treebanks, Stanza and SpaCy can automatically generate initial treebanks for Greek and Latin texts. Automatic parsing for Greek (based on just over 400,000 tokens) lags behind the automated parsing based on data in the Perseus/Prague scheme (based on 1.2 million tokens), but it will improve as new data becomes available in UD. Although treebanks for classical Latin lag behind those available for classical Greek, more than 1 million words of Latin from various periods are available in UD format, and automatic treebanking of Latin has begun to build upon this foundation.Footnote ³¹

INTEGRATION OF TRANSLATION ALIGNMENTS AND TREEBANKS

Treebanks and translation alignments were designed to complement each other. D. Bamman used both resources to model a dynamic lexicon that could, more than fifteen years ago, automatically generate treebank and alignment data to provide basic lexicographic information for arbitrarily large corpora.

Figure 7 illustrates an early application that draws upon both treebanks and alignments. The table on the left first uses treebank data to identify objects of the verb faciо̄, ‘I do’, and then draws upon translation alignments to gloss the Latin words. The figure on the right provides a model for a dynamic lexicon entry that uses automatically generated translation alignments and treebanking. The methods for automatic alignment and treebanking were much less accurate at the time when this work was done than they are now, but the most common patterns emerged against the noisy background. We are now at a point, sixteen years after D. Bamman's pioneering work,Footnote ³³ to implement such a dynamic lexicon for ancient Greek and Latin.

Figure 7.

Left: common objects for faciо̄ by author; right, mock-up of a dynamic lexicon entry for the Latin verb liberо̄, ‘to free’ (figures from Bamman and Crane 2008)Footnote ³².

Bamman's work on the Dynamic Lexicon used treebank and alignment data to complement each other, and the goal from the start was always to provide both treebanks and alignments in a single reading environment. Integration of treebanks and alignments in a single reading environment has two main purposes. First and foremost, the goal of an integrated reading environment is to make primary sources intellectually more accessible. With an integrated environment, readers do not have to switch between reading environments for treebanks and alignments that are not aligned with each other (e.g. readers who start with treebanks cannot automatically move to translation alignments and vice versa).

Second, publications are only born-digital insofar as they can circulate across different projects and platforms. Reviews of born-digital publications necessarily not only include expository prose (such as that in this Profile of translation alignments and treebanks), but also demonstrate that third parties can in fact reuse those born-digital publications. In this, digital publication resembles not just copyright law (which requires original expression), but also patent law, which (at least in the United States) has the enablement requirement. A new invention must be described in terms that are sufficiently clear, concise and exact so that other people can use it. Likewise, a publication is not truly a digital publication unless other people can reuse it.

We look for several features when evaluating born-digital resources such as treebanks and translation alignments. (1) Data must be provided under a license that allows it to be uploaded and redistributed by a third party. In practice, those working on digital ancient Greek and Latin have converged around variations of the Creative Commons open licenses. (2) The data must be available in a sufficiently documented format so that other projects find it more effective to reuse existing data rather than to start over and build their own. Documentation includes transparency not only for the structures (e.g. Text Encoding Initiative TEI XML) but also for accuracy of the data (e.g. were the treebank annotations or alignments produced by automated systems or created/curated by human annotators?) (3) The data must have credits in a machine-actionable format that other projects can analyse and retain. This third element represents arguably the biggest current challenge to true digital publication. The Perseus Dependency Treebanks for Homer and Hesiod identify the annotators for each and every sentence. When F. Mambrini, an early contributor to the Perseus Treebanks and an expert in that form, created a version of the Perseus Homer Treebank in the Universal Dependency Framework, he did not include the sentence level credits because he had no standardised way of representing these credits in the UD format. Likewise, the latest edition of the GLAUx Treebanks includes both automatically and manually treebanked sentences. It identifies whether a sentence was manually or automatically annotated, but it does not (yet) have a mechanism to identify specific human annotators.

To establish these three features, we were able to use the Beyond Translation reading environment. We developed Beyond Translation with support from a variety of sources as a step towards a new version of the Perseus Digital Library. Beyond Translation did allow us to create the first (though surely not the last) reading environment that combined treebanks, translation alignments with other categories of annotation (metrical analysis, links to grammars and lexica, textual variants, named linking). The point is not that any one particular system has integrated this data, but rather that it was in fact practical to integrate alignments and treebanks.