Translation Alignment for Historical Language Learning: a Case Study

Abstract

This paper proposes text alignment in digital environments as a way to empower language learning. It presents the principles and goals of text alignment in Natural Language Processing, and introduces Ugarit, a web-based translation alignment editor for the collection of aligned language pairs. Then, it reports observations on the application of translation alignment in historical language courses at Tufts and Furman University between 2017 and 2019.

Introduction

The World Wide Web has introduced a massive change in the cognitive processes of reading: skimming has become the most popular way in which readers interact with written texts on digital support, with possible enormous consequences on how human beings process information and articulate complex thought [Wolf 2018]. On the other hand, digital environments have an immense, often underdeveloped, potential to improve interactive engagement and exploration of information, integrating very different media and conveying extremely complex and multifaceted messages. Readers of the future should be educated to critically approach a text through digital technologies, making full use of this potential, and going beyond the reading processes conceived for texts encoded on static supports. In this paper, we propose the application of text alignment as one method to empower the reflective engagement with digital texts, specifically in the context of language learning.

Text alignment, and more specifically translation alignment, is a type of annotation,[1] which consists in analytically matching a word or expression in a text with the corresponding ones in another text, often in a different language, in environments created to collect aligned word pairs. This is an exceptionally complex hermeneutic task, by which readers are induced to problematize the function and implications of morphosyntactic, semantic, and expressive devices of a language, even of their own native tongue.

This paper collects a set of preliminary observations on groups of students that have used translation alignment in the framework of language courses at various levels of expertise. We propose text alignment as a way to empower the perception of the complexity of a written source, but also as a method to leverage usual obstacles in the process of reading in a different language by directly engaging with original literary texts.

Principles of Translation Alignment

Translation alignment is one of the most popular applications of Natural Language Processing. It is defined as the comparison of two or more texts in different languages, also called parallel texts or parallel corpora [Kay and Röscheisen 1993] [Véronis 2000],[2] by means of automated or semi-automated methods. The main purpose is to define which parts in a source text correspond to which parts of a second text in a different language. The result often takes the form of a list of pairs of items, which can be words, sentences, or larger text chunks like paragraphs or documents. This task is often performed automatically through various methods, which often require large amounts of training data: one of the most traditional methods for automated translation alignment is statistical machine translation [Brown et al. 1990], which uses statistical methods to extract and identify correspondences of words and phrases in two parallel texts, based on already available aligned pairs.

The alignment of texts in different languages, however, is an exceptionally complex task, because of the several variables involved. It is often difficult to find perfect correspondences across languages that express ideas through different morphosyntactic constructs, with variations in word order, sentence length, and even underlying cultural significance. Machine-actionable systems are often inefficient in providing equivalences for wordplays, metaphors, and other rhetorical devices. The resulting aligned pairs may be one-to-one (one word in the source text corresponds to one word in the translation), but often align as one-to-many, many-to-many, or many-to-one. Each word correspondence may be complete or perfect (with complete overlap between two words), but also possible or incomplete (partial overlap, or both words being a translation of each other only in certain contexts [Graça et al. 2009]).

For these reasons, manually aligned word pairs are one of the main sources of training data, and are not only employed for the implementation of machine translation systems, but are increasingly useful for many other purposes, including information extraction and automated creation of bilingual lexica [Dagan et al. 1999] [Melamed 1998]. Therefore, large corpora of manually aligned word pairs are a much desired resource, which remains, however, extremely rare.

The operation of collecting training data is often configured as a Citizen Science effort: for example, Glosbe (https://glosbe.com/) provides thousands of dictionaries created from the manual alignment of word pairs performed by a community of users.[3] However, because of the hermeneutic complexity implied in the alignment across different languages, translation alignment has also a powerful pedagogical potential. We have implemented a specific web-based environment, called Ugarit, to provide a framework for language learning applications through translation alignment.

The Ugarit Translation Alignment Editor

The first public web-based environment for manual text alignment of parallel corpora was conceived within the Alpheios Project (http://alpheios.net/), and is currently hosted by the Perseids Project (https://www.perseids.org/). The editor offers an easy environment to manually match corresponding words in two parallel texts.

The Ugarit Translation Alignment Editor (http://ugarit.ialigner.com/), developed in 2016 by Tariq Yousef at the University of Leipzig, was conceived as the first version of an improved interface that could perform more extensive alignment tasks than the Alpheios editor: it allows alignment in up to three texts, incorporates transliteration of non-Latin alphabets, and collects manually aligned pairs as training datasets [Yousef 2019] [Yousef and Jänicke 2020]. Moreover, it allows almost immediate visible online publication of the material.

The first version of the tool was made public in March 2017. Since then, Ugarit has registered an ever-increasing number of visits and new users.

Originally, Ugarit was developed to collect training data for the implementation of statistical machine translation of historical languages, mainly Ancient Greek, Latin, and Persian, for which few to none aligned datasets exist. Ideally, historical languages are closed systems with a finite number of words and very limited change in the foreseeable future. Therefore, it should be possible to create adequately efficient automated methods of statistical machine translation based on a relatively small training dataset.

However, after the tool was made public, the number and variety of languages included by the users has steadily increased and has gone far beyond the original intent: at the moment this article is being written, 45 languages are included in Ugarit, including Ancient Greek, Latin, Arabic, Persian, English, French, German, Italian, Portuguese, Croatian, Armenian, Japanese, Chinese, Syriac, Sanskrit, Coptic, Egyptian, Akkadian, and Ethiopic; 465 unique users, and about 24,500 parallel texts, are currently hosted.

The home page of Ugarit offers a quick overview of the languages currently available (the graph showing the proportions of the respective corpora), the number of users and texts, the available word pairs, and the newest alignments.

The workflow is simple: users can upload their corpus in text format, or import it by calling its CTS URN from the Perseus Digital Library (http://www.perseus.tufts.edu/hopper/) [Babeu 2019]. Then, the parallel texts can be manually aligned by matching words or groups of words. Resulting pairs are automatically stored in the database, and can then be exported in XML or tabular format.

Currently, Ugarit works with up to three texts in different languages, and there is virtually no limit to the amount and length of texts that can be uploaded. A clickable option enables the user to decide whether the alignment can be publicly visible on the website or not. When the default value is kept, the text appears in the homepage of the website, and other users can inspect the alignment by hovering with the mouse on individually paired tokens, which are automatically highlighted. The proportion of aligned tokens is indicated in the colored bar below the text: the green indicates the rate of matching words, the red the rate of non-matching words.

For languages with non-Latin alphabets, Ugarit offers automatic transliteration, which is visible when the pointer hovers on the desired word. This feature is currently available for Greek, Arabic, Persian,[4] Armenian, and Georgian.

Translation pairs are stored in a local database that gathers data from the manually aligned texts, but also from bilingual dictionaries and automatically aligned texts with GIZA++.[5] This database would form the bulk of training data to improve automatic translation alignment methods focusing on historical languages.

As such, the database of translation pairs serves two main functions: first, it serves search queries to provide information on all the available translations of a word. Therefore, if the user looks for a word (for instance, "life") with the search function available on the home page, the output displays all the available aligned pairs currently hosted in Ugarit. In this way, the user can access an extended number of translations of a word in many different languages.

Second, Ugarit provides training data to the Dynamic Lexicon (http://dynamiclexicon.com/), which applies the principle of triangulation to extract bilingual pairs across Ancient Greek, Persian, and Latin, using bridge pairs in Greek/English, Persian/English, and Latin/English: the Dynamic Lexicon fetches alignment information from all available aligned corpora, and it uses aligned languages as pivot to fetch translated words in other languages that may bear the same meaning [Berti and Yousef 2015]. So, each aligned pair functions as a translation vector to access more matches in other languages that are not directly aligned. This method is based on the assumption that two words in two given languages are more likely to match if they are translations of the same word in a third language [Kumar et al. 2007]. The accuracy of the triangulation can be improved by expanding the available corpus of translation pairs, considering available part-of-speech classification, or adding more pivot languages [Yousef 2015].

In other words, Ugarit was designed as a Citizen Science resource to collect training datasets on historical languages from a variety of sources and projects. One of the most important initiatives in this regard was the Hafez Project (http://dynamiclexicon.com/hafez/; http://www.divan-hafez.com) led by Maryam Foradi at the University of Leipzig [Foradi et al. 2017], which tested how German and Persian speakers aligned textual portions of Hafez using English as a bridge language. The results indicated that, provided with the appropriate scaffolding by using English as a bridge language, users with no knowledge of the source language could generate word alignments with the same output accuracy generated by expert users in both languages of source and target texts. The study showed that, while the long-term pedagogical effectiveness of this method is comparable with learning vocabulary with flashcards, the vocabularies learned both intentionally with flashcards and incidentally through text alignment had a significantly better retention [Foradi 2019].

These results suggested that alignment could serve as a pedagogical tool with a certain effect of long-term retention of vocabulary. Between 2017 and 2019, we conducted a pilot study investigating how students of historical languages could use translation alignment to improve their learning experience through direct engagement with original texts. The lack of scholarship on the pedagogical application of alignment as a tool for language learning, and the complexity involved in the reproducibility of the experience, compel us to start with pre-experimental observations from specific cases, where translation alignment has been applied as a standard method during language courses. Our initial observations provide the justification for a more experimental design.

Our Pilot Study

Ugarit was used in graduate and undergraduate semester-long Classics courses at the Universities of Tufts and Furman, with different focuses: language (mainly Ancient Greek and Latin) at elementary, intermediate, and upper level, literature in translation, literature surveys, and individual research projects.[6] The use of translation alignment was incorporated in the teaching practice as a structural element:

The students made continuous use of Ugarit in class, using and creating aligned translations as a reading support.

The task of aligning more lengthy chosen texts against one or two translations constituted the main type of assignment, in the form of individual research projects at various degrees of complexity: depending on the course and level, the students produced bilingual and trilingual alignments, whose accuracy was graded accordingly.

The results were discussed in the form of research papers and presentations, in class and also in public venues.[7]

The students reported their observations on the process, evaluated how it affected their understanding of the source, and performed a qualitative analysis on the languages in question, examining specific morphosyntactic, semantic, and expressive phenomena.

We observed the application of this workflow on students with very diverse language skills, including those who mastered other modern languages, and those whose native language was not English. We propose three cases of study:

1. Students who were proficient in the source language and had complete mastery of one other language (including the native tongue).
2. Students who were proficient in two different languages, and had basic knowledge of a third one.
3. Students who were proficient in one or more languages, but had no prior knowledge of the source language.

In the following paragraphs, we report the considerations formulated by the students themselves, and our observations on the outcomes of the process.

Case 1: Students of the source language

Students of Case 1 were enrolled in intermediate and upper level courses in historical languages, mainly Ancient Greek or Latin. They performed two tasks:

1. Pairwise alignment between the source text and one chosen translation in a modern language.
2. Alignment of the source text against two different translations.

The purpose of the first exercise was to set the groundwork for a structural analysis of the most important morphosyntactic and semantic devices of the source language, comparing how those features were rendered in a different language (often English), of which the students had complete mastery. In the second exercise, they built upon this experience to draw an evaluation of two competing translations, focusing on the different ways in which complex linguistic phenomena can be rendered in another language.

Example 1: Plato, Symposium 210-212. Ancient Greek-English.[8]

Graduate, native English speaker, upper level Ancient Greek. The student created bilingual alignments of Plato’s Symposium, comparing it against the translation by Nehamas and Woodruff (1989), in the context of a research project[9] that integrated different levels of linguistic digital annotation, with the goal of enabling readers with no Greek knowledge to explore the original through the dynamic reading of the aligned parallel texts. The student created a consistent alignment strategy that only matched words corresponding in meaning and grammatical function (with the partial exception of verbs for which a perfect match would not be possible, such as subjunctives or participles with predicative use).

Example 2: Apuleius, Metamorphoses 8.19-22. Latin-English.[10]

Undergraduate, native English speaker, intermediate Latin. The student performed a three-text alignment of a passage from the Metamorphoses by Apuleius, comparing the original and two chosen English translations (A.S. Kline, 2016 and R. Graves, 1950), to discover differences in translation strategies in a very complex and rhetorically skilful literary work. The focus of the observation was the semantics of different words and expressions with potential ambiguity, but also the different ways the translators addressed the challenges of rendering Latin syntactical structures in a completely different linguistic system. The student then created a new English translation, discussing where it distanced itself from the original and which aspects were instead retained. The comparison between two very different authorial translations proved useful to understand different strategies of addressing relevant linguistic and semantic problems, but also to help the student to find their own "translator’s voice".

Example 3: Homer, Odyssey 9.105-115. Ancient Greek-English.[11]

Undergraduate, English-Chinese speaker, upper-elementary Greek. The student aligned the original text of Odyssey 9.105-115 against two different English translations, in prose and poetry respectively (A.T. Murray, 1919 and S. Lombardo, 2000). The student used the correspondences in meaning and words across the two English renderings to individuate possible matching words in the Greek, specifically focusing on the potentially discrepant translation strategies between prose and verse. Overall, the prose translation revealed a slightly better alignment rate, but typical Greek stopwords, such as δέ or τε, were consistently not translated in both English versions.

Case 2: Students of a third language

The typical case for trilingual alignment was a student who was proficient in two languages, and wished to improve a third one. The student would perform a trilingual alignment (possibly preceded by a pairwise alignment between the two better-known languages), systematically comparing a lesser-known language against two translations. The student would use the knowledge of the first two languages to leverage the obstacles in understanding the third one, by recognizing common syntactical patterns, morphologies, or similarities in vocabulary.

Interestingly, this case was not limited to the study of historical languages: a collateral advantage of trilingual alignment was that students with proficiency in a historical language could use this opportunity to focus on another language that they knew less, often a second modern language that they were studying at the same time.

Example 1: Tacitus, Germania 1.1. English-German-Latin.[12]

Undergraduate, native English speaker, advanced Latin, basic German. The student performed a trilingual alignment of Tacitus’ Germania, against an English and a German translation, and used his knowledge of Latin to improve his mastery in German. In his final report, he observed higher matching rate between the two inflected languages than with English: rather paradoxically, comparing German and Latin resulted easier than English and Latin. The English translation chosen was, on the other hand, approached with some criticism, because of its lengthy rendering of expressions that resulted equivalent in German and Latin.

Example 2: Loqman Fables. Arabic-French-Ancient Greek.[13]

Graduate, advanced French, intermediate Ancient Greek, basic Arabic. The student performed a trilingual alignment of the famous Loqman fables, to pursue a research project that focused on gathering systematic evidence of the relation between Loqman and the Aesopic corpus, which the sage is often said to have translated [Heller and Stillmann 2012]: the student aligned the original Arabic and the corresponding text of the fables from the Aesopic corpus, using the French translation of Loqman as a pivot between the two languages. The high matching rate across the three versions, and the often verbatim correspondences between the Arabic version and the Aesopic fables, provide the first dataset of analytically collected evidence of the dependence between the Loqman and the Aesopic corpus.

Example 3: Sophocles, Ajax 646-692. Ancient Greek-Chinese-English.[14]

Graduate, Chinese native speaker, English as second language, intermediate Ancient Greek. The student performed a trilingual alignment of Sophocles’ Ajax, aligning the Greek text, the English translation by Sir R.C. Jebb (1896), and a Chinese version based on the English. While the Chinese and the English obviously displayed a higher matching rate, the student was able to use the knowledge of both languages to individuate correspondences with the Greek, and also to establish where there was no available match for Greek words or concepts.

Case 3: No prior knowledge of the source language

The third case involved students, enrolled in courses on culture or literature, who had no prior knowledge of the source language and were also not studying it. The students were given short sections of a source text and corresponding translations in various languages, and performed several tasks aimed at gradually building a critical approach to the original:

1. Individuate single words in the source text, such as "holy", "honor", or "love", and align them with the corresponding words in a translation in a chosen language. The goal of this exercise was to develop an understanding of the depth of crucial cultural concepts, by assessing the many different ways in which the same word in the source text could be translated.
2. Compare how specific expressions were rendered in two translations, and then align them against the source text.[15] The students focused on different ways of conveying morphosyntactic constructs across languages, critically approaching different translation strategies.
3. Complete bilingual or trilingual alignment of a short passage, with intensive and systematic analysis of the linguistic and expressive content. The purpose of this exercise was to let students assess the linguistic differences between the languages in question, and partly to demonstrate to them that they could, in fact, approach a text in a language that they did not know at all, by means of translation alignment.

Interestingly, we observed that students who had mastery of more than one language almost immediately learned to use trilingual alignment to leverage the difficulties in dealing with a language that they had never seen before. This strategy proved effective in various cases: many students completed the first two assignments using both English and a third modern language (often Spanish or French); non-native English speakers used their native language to better understand complex words and expressions; students who mastered another classical language tested if they could use it to better engage with phenomena like inflected forms, participial uses, and subordination. Overall, trilingual alignment proved to be an effective strategy to engage with the source text, as already assessed in Case 2. By the end of the course, students were able to align relatively long passages against one or more translations. Example 1: Herodotus, Histories 1.35. Ancient Greek-English-Latin.[16]

Undergraduate, native English speaker, advanced Ancient Greek, no Latin. The student used trilingual alignment to verify whether the knowledge of Greek, with English as a bridge language, could serve to leverage the unfamiliarity with a text in Latin. The passage chosen, from Herodotus’ Histories, was aligned in the original Greek, against the English translation by A.D. Godley (1920), and the Latin version by J. Schweighaeuser (1824). The experiment proved successful: the student was able to recognize cognate words and match them with the corresponding Greek, and to use the knowledge of inflected languages to compare parts of speech between Latin and Greek (for example, to identify the accusative fratre-m based on the similar ending with the Ionic Greek ἀδελφεό-ν).

Example 2: Euripides, Bacchae 395-399. Ancient Greek-Chinese-English.[17]

Undergraduate, Chinese native speaker, English as second language, no Ancient Greek. The student created a trilingual alignment of Euripides’ Bacchae, comparing the English translation by I. Johnson (http://jelks.nu/libri/classics/bacchae.html) and a revised Chinese version based on the English (as there were no available Chinese translations of the original Greek). As a Chinese native speaker, the student used Chinese as a bridge to establish accurate correspondences between the English and the Ancient Greek words that had never been seen before. The result was an investigation into the meaning of Ancient Greek concepts through a meaningful comparison with similar Chinese terms, often with an emphasis on imperfect grammatical and semantic correspondences between the Greek and its translation (such as τὸ σοφόν, "cleverness", translated as "being clever").

Observations

At the end of the course all the students, no matter what level of mastery they had of the languages, developed an acute sense of how limited their understanding of the source text was from the translation, as much as the critical recognition that they could convey in-depth knowledge of critical cultural concepts expressed in the source language, by contrasting them with their modern language translation and investigating into their various meanings [Crane 2019]. The recognition of the limited capabilities of translation to efficiently convey concepts in Classical languages, while at first caused disappointment and criticism, convinced the students of the need to look more in depth into the original language to understand its inherent characteristics and the way it expressed ideas [Palladino 2020].

Students of the source language developed a tangible sense for the fluidity of translation by evaluating different strategies employed by professional scholars to approach complex phenomena; they systematically approached complex morphosyntactic constructs and were compelled to discuss their function and significance, while assessing the necessary imperfect character of any translation of them. In addition, students who were experimenting with learning a new modern language could approach unknown expressions and vocabulary by using their skills in other languages as a bridge to convey similar constructs.

One of the most relevant outcomes of this study was that students with no prior knowledge of the source language could start learning it by directly reading original literary texts. This approach could be revolutionary in the field of slow reading and language learning, and it also promises non-trivial consequences for passive users: readers with no knowledge of a source language could easily use already available alignments to perform a dynamic reading of the original aligned with its translation, gaining a basic understanding of vocabulary and syntax in the original.

Limitations and Future Work

We hope to have shown the pedagogical potential of translation alignment. However, we are also aware that significant implementation needs to be pursued to improve some underdeveloped aspects and to allow a more systematic use of a translation alignment interface in teaching practices.

We recognize that at the moment the tool is open to every category of user, and the aligned pairs are partly the result of unsupervised work, which may affect the quality and consistency of the information collected in the database. The lack of a suitable evaluation/correction workflow not only makes the dataset not accurate enough for machine translation, but also prevents organic teacher supervision. Therefore, the integration of a voting system for the evaluation of the accuracy by teachers or expert users is one of the most needed implementations. A related problem is the current impossibility of working on collaborative projects, where multiple users can work together on the same corpus.

Secondly, not all translation pairs are equally useful as training data. We should expect alignment to produce conflicting, mutually exclusive, results: after all, translation alignment is the result of interpretation, and not all cases are easily classifiable as "right" or "wrong". In some cases, typically literary texts, it is extremely difficult to establish perfect or even partial correspondences between expressions and concepts in potentially very different languages: we have found that, in lack of specific guidelines or gold standards for each single language pair, users tend to create their own set of rules to keep a consistent alignment strategy, but these rules tend to be very different according to the particular purpose of the alignment. Therefore, we need a system to filter translation pairs: an option would be to allow users to classify different kinds of translation pairs, for example distinguishing categories such as "perfect/complete" or "partial/incomplete", and "literal" or "free/poetical".

Finally, we recognize that immediate access to part-of-speech tagging and available word pairs while performing the alignment would enormously impact on the work of non-specialized users and students, and obviously on the improvement of analytical passive reading. This very desirable additional feature is also included in our future implementation.

Notes

[1] In digital environments, annotation is the operation of attaching additional information to a resource or a specific part of it: an annotation can be unstructured, such as a comment, or structured, such as a machine-actionable citation identifying a text [Romanello and Pasin 2013].

[2] A text and its translation or translations placed alongside are defined as parallel texts. Large collections of parallel texts are called parallel corpora, such as the European Parliament Proceedings (http://www.statmt.org/europarl/) or the Bible Corpus [Christodouloupoulos and Steedman 2015]. Large research infrastructures like CLARIN (https://www.clarin.eu/resource-families/parallel-corpora) host big collections of parallel corpora in multiple languages as a resource for automatic translation and linguistic research. The use of parallel corpora in translation management systems is increasingly popular: commercial tools like SDL Trados (https://www.sdltrados.com/), MemoQ (https://www.memoq.com/), or Across (https://www.across.net/) allow professional translators to manually align sentences while performing their work, at the same time keeping consistency in the translation.

[3] For current issues in using Glosbe for statistical machine translation, see Berti (2015).

[4] Since the short vowels in Arabic and Persian are not written, the model used here needs to be improved to increase the readability of the transliterated text.

[5] GIZA++ (http://statmt.org/moses/giza/GIZA++.html) is a popular program, created as an implementation of the famous IBM models 1-5, which provides automatic alignments of words or phrases to implement statistical machine translation.

[6] See, for instance, Example 1 of Case 1, where alignment was part of a complete digital project.

[7] In 2018, Furman students presented their projects at two undergraduate conferences, at UNC Greensboro and University of South Carolina respectively.

[8] http://ugarit.ialigner.com/text.php?id=2154

[9] https://github.com/ChiaraPalladino/TuftsDCC/wiki/Jiyoung-Song,-Diotima-on-Love-(Plato,-%22Symposium%22)

[10] http://ugarit.ialigner.com/text.php?id=25919

[11] http://ugarit.ialigner.com/text.php?id=26615

[12] http://ugarit.ialigner.com/text.php?id=26669

[13] http://ugarit.ialigner.com/text.php?id=24421

[14] http://ugarit.ialigner.com/text.php?id=25279

[15] A strategy fully developed in Case 1, above.

[16] http://ugarit.ialigner.com/text.php?id=25197

[17] http://ugarit.ialigner.com/text.php?id=2134