The development of bilingual dictionaries to medieval translations presents diverse dificulties. These result from two types of philological circumstances: a) the asymmetry between the source language and the target language; and b) the varying available sources of both the original and translated texts. In particular, the full critical edition of Tihova of Constantine of Preslav's Uchitel'noe evangelie ('Didactic Gospel') gives a relatively good idea of the Old Church Slavonic translation but not of its Greek source text. This is due to the fact that Cramer's edition of the catenae - used as the parallel text in it - is based on several codices whose text does not fully coincide with the Slavonic. This leads to the addition of the newly-discovered parallels from Byzantine manuscripts and John Chrysostom's homilies. Our approach to these issues is a step-wise process with two main goals: a) to facilitate the philological annotation of input data and b) to consider the manifestations of the mentioned challenges, first, separately in order to simplify their resolution, and, then, in their combination. We demonstrate how we model various types of asymmetric translation correlates and the variability resulting from the pluralism of sources. We also demonstrate how all these constructions are being modelled and processed into the final indices. Our approach is designed with generalisation in mind and is intended to be applicable also for other translations from Greek into Old Church Slavonic.
The creation of the Slavonic alphabet in the second half of the 9th century marked the beginning of a completely new written tradition in Europe – first only in Glagolitic script, then in Cyrillic. The main part of this literary production consisted of translated texts from Greek, which were linked to the needs of the church practice. The studies on the relationship between source and target texts have occupied a significant place in the philological Slavic Medieval Studies (Palaeoslavistics) and an important tool for their analysis have been the bilingual dictionaries to specific scholarly edited writings. These diachronic dictionaries provide the scholars not only with a generalised view to the vocabulary of the respective work but also with an opportunity to study various specific issues concerning the Greek-Slavonic translation correlates.
The very process of developing such dictionaries is complicated not just from the philological perspective but also from the perspective of their computer support. A major technical challenge in this respect is the asymmetry between the two languages, i.e., the cases in which there is no one-to-one correlation of the words of the original and their translations. We present the approach we have developed for computer-aided creation of such dictionaries, together with the main problems we have encountered and the decisions we have made to overcome them. This paper is structured by the following sections: philological basis, technical approach, overview of specific problems and their solutions, and conclusion.
The philological background of every bilingual dictionary to a translated text is made up of the two respective sources – original and target ones. In most cases the medieval translations are not preserved as authentic documents, so scholars build their understanding of them on the basis of the preserved (or available) today copies and/or editions. Therefore, the text-critical tradition of the original text and its translation, as well as the degree of the research on them, are factors which directly afect both the content and appearance of the end lexicographic product. Here follows a brief overview of the sources in relation to the bilingual word indices to the Old Church Slavonic Uchitel’noe evangelie (UE).
UE was created by Constantine of Preslav, a prominent Old Bulgarian man of letters, probably between the years 886 and 893 [1: 3–4, 10; 2: 86]. It consists of 51 sermons commenting on the respective Sunday Gospel readings for the whole year. The main part of this commentaries is translated from Greek, but the prologue to the collection and the most of the introductory and concluding words to each sermon are authored by Constantine of Preslav.
The manuscript tradition of UE is not complicated – 4 full copies came down to us [
This means that in order to select the most probable original translation correspondences in
the word indices, the material from all preserved manuscripts needs to be critically used. The
diplomatic edition of the oldest witness by Tihova which includes the variant readings after the
other three copies [
No Byzantine collection, identical to UE, is known, but even the early researchers Gorsky and
Nevostruyev assume that Constantine of Preslav used some kind of an abridgement of John
Chrysostom’s Gospel homilies [11: 412, 423–424], and Antonij the archimandrite [12: 37, 40–50]
points out the presence of parallels in Greek Gospel catenae published by Cramer [13]. Using
the same multivolume publication, Tihova adds to her edition the Greek counterparts for the
majority of the translated parts in UE. Her omissions (mainly due to the fact that Cramer’s
Supplementum was unavailable to her) are added by Krys’ko [
The complexity of these dictionaries, imposed by the two textual traditions and the plurality of sources for each, is combined with the usual dificulties brought about by the asymmetry between the two languages. It is the reason that the lexical units from the original and its translation do not always correspond unambiguously, and also that often a given grammatical meaning is conveyed by lexical means or vice versa. UE, just like other earlier Old Church Slavonic translations, is freer than the later ones because the literary language is still in process of creation, which also means that the stable correlates are rare and the possibilities for variations (both qualitative and quantitative) are much greater. This suggests a significant number of cases beyond the standard models which need specific solving.
Automatic alignment between original and translated text is a common task in corpus
linguistics [18, 19] . It is performed on languages that are rich with linguistic resources and most
often it is limited to alignment at the sentence level. Word-level alignment is a dificult task
with limited success rates [19]. When it comes to languages with less resources available,
especially old and regional ones, automation is not possible and the use of technology is limited
to supporting the work of philologists [20, 21, 22] [
Our approach is based on the scheme for the production of dictionaries to the Church Slavonic translations of the anti-Latin polemic of Gregory Palamas and Barlaam of Calabria, developed by von Waldenfels and Taseva [21]. However, due to the greater complexity of the indices to the UE, the approach has had to be revised and expanded to meet the philological goal of approximating, as close as possible, between the Greek sources and their translation. The solution presented here uses two particular open technologies: the Python programming language and the Open Ofice XML file standard. The result of the process is the generation of two word-indices: Slavonic-Greek and Greek-Slavonic. Each of these maps each lexeme from one of the texts to all the word usages corresponding to it in the other language.
This publication examines in detail the last crucial step of a wider process whose initial stages are as follows. It begins with a transcription of the source text (in this case Tihova’s edition of UE) carried out by Rabus with the program Transkribus (for the program see [23]). The text document is then corrected and enriched by a philologist who manualy introduces positional variants from the Slavonic transcriptions. Then it is automatically transformed into a spreadsheet table where each line contains a word from the text and its possible variants. The project team then enriches this table, adding Greek correspondences (including variants), lemmatising the word usages (bringing them into standard dictionary form), and adding necessary grammatical and lexicographical annotations. As a result, the completed table (see Fig. 1) contains the positional correspondences between the source and its target, annotated with the necessary philological information. This table is the input solution described here.
The structure of this table should allow for the representation of variant readings, yet be easily maintainable for philologists working in teams. The proposed solution for this trade-of is summarised in Table 1 and detailed here. Since the focal point of this work is the Slavonic translation, the used Slavonic word form (column F) in the main text, its lemma (columns H) and their sublemmas (in case where it is necessary to explain the correspondence in the other language, sequentially in columns I-K) are central. Next to these are their addresses in the text, consisting of word, page, column, and line (column E), and their contexts, containing the entire line from which the corresponding word was extracted (column G). The information in column G is not used by the software, but is useful for the annotation work of philologists. To the right of the columns for the main Slavic text is the corresponding information for the Greek main text – word usages (column L) and their adjacent lemmas (columns M-P). Where variant information is available, it is added in the free columns on the appropriate side. In the columns to the left of those for the main Slavic text is the information on Slavic variants – word usages (column A) and lemmas (columns B-D) respectively, and to the right of the main Greek text is the information on Greek variants – word usages (column Q) and lemmas (columns R-T). When more than one variant reading is present, the information is combined in the same variant columns, as illustrated in the last example of Section 4.2.
In the last step of the wider process, the two types of dictionaries are created: lists and indices (see Fig. 2). The lists contain all the positional correspondences in the two texts united under a common lemma. Entries show specific word usages and a precise address. These lists are used by philologists to verify alignment and lemmatisation, to correct possible inconsistencies and errors in the collaborative enrichment. The indices present the final word indices in a form ready for publication. They contain only the summary information about the lexical correspondences, a list of the exact address of the lemma occurrences and information about their frequency in the text. For each of the two types of dictionaries, both Slavonic-Greek and Greek-Slavonic versions are created.
Two dedicated software tools are used to create the lists and indices using shared logic in the software implementation, the main elements of which are described in this section. The lists are generated by a program called integrator. The indices are created by a program called generator. Each of the two tools performs three steps, executed sequentially for each of the two dictionaries – Slavonic-Greek and Greek-Slavonic. The first two of these steps are common, the third is diferent for each of the tools: 1. Adaptation 2. Aggregation 3. Export Adaptation of the content extraction table. This step takes into account which direction the dictionary is currently being built for and creates in-memory tables similar to the input table, but with transformations that allow all the information needed for a single word use to be contained in a single row. This is relevant in cases of quantitative asymmetry (see next section), where this condition is not satisfied for the input table. If there are variants in the information that enters the new row, the step considers separately the columns for the main text and the ones for the variants. This way the necessary adaptive transformation can be applied for each of the two cases separately. This transformation is an application of the divide-and-conquer principle. In this case it is important, firstly, for performance reasons. Secondly, it reduces the task of constructing the list or index to the simpler subtask of generating entries from a single row in the adapted table. The simplified subtask is solved in the following step. Aggregation of alphabetically ordered indices. This step is implemented using nested sorted mapping1 standard data structures, which allow the construction of specialised alphabetically sorted reference tables functionally similar to the desired end result. Because this structure is shared by the integrator and the generator, it must collect the necessary information for both tools to allow the data needed for each to be retrieved and displayed. To make this possible, a grouping is created in the sorted mapping structure according to the following hierarchy: → ( → (2 → (3 →))) where the parenthesis indicate optional sublemmas, included only if present in the input table.
We call Alignment the structure shown in Fig. 3. It consists of the combination of Usage (i.e. the word usage) in the source and target languages, an Address, and information about whether the text is a Biblical quotation. The Address allows to locate the word usage up to the line (containing also page and column) or up to a line span (represented by the self-reference in address indicating the end of the span in Fig. 3). Biblical quotations are noted in the input tables in bold and italics, and accordingly appear in the same way in the output dictionaries, an example being present in Fig. 7. Usage in turn has several important properties. First, these include 1In Python called SortedDict, available as part of https://pypi.org/project/sortedcontainers/ information about the respective word, its lemmas and, if there are repetitions at this address, a counter specifying the particular repetition of the word in the line, if any. The language of the Usage is also stored because of the direction-agnostic approach (see last paragraph of this section), which applies the same logic to the source-to-target and target-to-source dictionaries. Last but not least, the structure stores information about the source (Source) of the usage, and possible alternative spellings (Alternative), divided into main and a mapping (var ) of variants to their alternative spellings. This structure contains the necessary information for the needs of the dictionaries created in the next step.
Export of the dictionaries to a word processing document formatted according to their respective needs – for the integrator and the generator. The result is grouped according to the hierarchy of the previous step. The lemmas are at its highest level. They are sorted alphabetically. They list the usages and possible references to other corresponding variants. In the case of sublemmas at diferent levels, they divide the usages in the corresponding lemma into separate lines formatted with indentation. Since the integrator ’s goal is to allow for the verification of the manual preparation of the input table, in addition to the lemmas, the specific word usages are displayed in the generated lists. However, for the indices these word usages are superfluous information. Rather than that, the generator adds to the indices a frequency count of occurrences for each of the languages – both in the main text and in the variants (the latter are indicated with var in superscript). In Fig. 2 the result from the sample input from Fig. 1 is shown. Note that in the example, the word in the Greek main text is missing (denoted by “om.”) and, accordingly, in the Greek dictionaries (the respective list and index), only one lemma appears (παρά). And the two lemmas in the Slavonic dictionaries (въ and оу) are paralleled with only one Greek word from an unspecified variant (παρά).
Our approach applies one more simplifying technique – the Slavonic-Greek indices and the Greek-Slavonic indices are treated as symmetric. In other words, the same programming logic is used to generate the indices in both directions. The only diference between the two is that the table columns of the source and target languages are swapped when provided to the program for the Slavonic-Greek index and for the Greek-Slavonic index, respectively. For shortness, in the next section we will only show the two-way result in one of the dictionaries – the one that better illustrates the diferences.
In order to illustrate in detail how our approach works, here we look at specific problems and our corresponding solutions. We address two categories of problems: 1) quantitative asymmetry, which we illustrate with the generated lists, and 2) variability in transcripts, better illustrated by the resulting indices.
As mentioned in the previous section, with our approach and the program implementing it, the quantitative asymmetry is solved at the adaptation step. Several examples that demonstrate diferent cases of asymmetry follow. One-to-many (1:n) is the base example that occurs when a Greek word is translated with a Slavonic phrase. To indicate such occurrences, we use background colouring of the corresponding rows in the word-use column (F), as shown in Fig. 4. In cases where two groups immediately follow each other, they are delineated by using an empty line between the two.
What the adaptation step does is to collect translation information from the whole phrase and associate it with each of the lexemes by adding it in its corresponding row in the table. In the example below, according to the currently generated dictionary, the target words (for Slavonic in column F) remain as they are, but the lemmas (H) are given as a phrase – together – and repeated for each of resulting entries.
of the words from the phrase has only grammatical meaning. The example in Fig. 5 uses an analytic verb form for a past participle.
To indicate such occurances we use “gramm.” in the second lemma (column I) with a coloured background. The particularity in this case is that the grammatical value is valid only for the lexeme on the same line, but not for the rest of the phrase. In the dictionaries, therefore, this information should be shown only for the corresponding lemma, but not for the others in the group.
In this case, the adaptation step does not add the sublemmas of the lexeme with grammatical value to the information about the other lexemes. Thus, the corresponding feature is present only in the translation from Slavonic to Greek and leads to an asymmetry between the two dictionaries in Fig. 6 (see sublemma “gramm.”). This is the reason why diferent adapted tables are needed for the aggregation step from Slavonic to Greek and from Greek to Slavonic. Many-to-many (n:m) asymmetry is an occurrence when a Greek phrase is translated with a Slavonic phrase but there is no direct corresponding between the separate words of these phrases in the two languages as shown in Fig. 6. Our approach solves this case using the same logic. In the adaptation step, the program collects in a single row all the information needed for an entry in the dictionaries. As a consequence, the aggregation step, does not need to perform any processing other than just reading the data row by row and adding it to the sorted mapping data structure.
As a rule, the indices and lists need to include all word uses from the main text and those variants from the other copies that represent either better readings than the main text or equally possible readings in the given context. Missing, wrong or less precise variants from the Slavonic transcripts W, G and H are ignored, and not lemmatised, so the program does not include them in the dictionaries.
Diferent translation correspondences (seen in Fig. 1 and Fig. 2) is a scenario where a word other than the main one is used in the additional sources. In such circumstances, the dictionaries show the usage in both lexemes used (in the main version and in the variant), adding references to the variant lexemes to the addresses of the usages. This functionality is the reason why – in the aggregation step – the structure describing the usages also contains information about the alternatives.
Diferent copies suggest diferent correspondences is a fundamentally similar but more complicated occurrence. In the example in Fig. 7, the rows indicate four diferent combinations delineating three diferent translations ( но ѧдъ, д нородъ and д но ѧдъ ) of the Greek word μονογενής, as well as the possibility of a missing translation. Similar to the previous example, this diversity should be traceable in the generated lists and indices through the corresponding references. Variant readings in copies are relatively rare, and it is even rarer to have two diferent variant readings to the same lexeme. The solution should therefore be able to handle such cases without complicating the manual enrichment and annotation performed by philologists. The solution adopted here combines the diferent variant readings in column A and their respective lemmata in column B for the Slavonic variant. Such cases are interpreted in the aggregation step, where the combinations of variants are recorded as separate Alignment data objects.
This paper presents our work on creating dictionaries for UE. In it, we show the combined solution to two particular methodological challenges: quantitative asymmetry and variation in the sources. We solve these problems applying a stepwise approach of adaptation, aggregation and export. We demonstrate a divide-and-conquer approach (implemented in the adaptation step) that allows solving even complex cases of many-to-many quantitative asymmetry with diferent grammatical values. We also propose alignment modelling for the purpose of generating dictionaries with cross-references for text variants. As a result, in the dictionaries it is possible to indicate not only the alignment between original and translation, but also between these two and a large number of variants.
The approach has been applied to the 52 constituent texts of the UE (the introduction and the 51 sermons). More complex combinations of asymmetry and variation in the sources were also successfully modelled for this purpose. Such cases represent combinations of the examples provided in this paper. The program was written with applicability to other Greek-Slavonic translations in mind. In order to allow for this applicability, and for future extensions, we commit to publishing the software code under a free license.
The article is written within the project The Vocabulary of Constantine of Preslav’s Uchitel’noe evangelie (’Didactic Gospel’): Old Bulgarian-Greek and Greek-Old Bulgarian Word Indices, financed by the Bulgarian National Science Fund (contract КП-06-Н50/2 of 30.11.2020). didactic gospel of constantine of bulgaria: editio princeps.], Bulletin of the RAS: Studies in Literature and Language 76 (2017) 52–62. [11] A. V. Gorsky, K. I. Nevostruyev, Описание славянских рукописей Московской синодальной библиотеки. Отдел 2. Писания святых отцев. 2. Писания догматическия и духовно- нравственныя [Description of the Slavic manuscripts of the Moscow Synodal Library. Section 2. Scriptures of the Holy Fathers. 2. Dogmatic and Spiritual-Moral Scriptures], Synodal Printing House, Moscow, 1859. [12] archim. Antonij (Vadkovskij), Из истории древнеболгарской церковной проповеди. Константин, епископ болгарский и его Учительное евангелие [From the history of Old Bulgarian church preaching. Konstantin, Bishop of Bulgaria and his Didactic Gospel], Typogr. Imp. Univ., Kazan, 1885. [13] J. A. Cramer, Catenae graecorum patrum in novum testamentum, number v. 4 in Catenae graecorum patrum in novum testamentum, E Typographeo Academico, 1844. [14] W. Lamb, Conservation and conversation: New testament catenae in byzantium, in: D. Krueger, R. S. Nelson (Eds.), The New Testament in Byzantium, Dumbarton Oaks Research Library and Collection, Washington, D.C., 2016, pp. 277–299. [15] J.-P. Migne (Ed.), Patrologiae cursus completus: Series Graeca, T. 58. Joannes Chrysostomus,
Paris, 1862. [16] D. Kotova, “Слово 19 от Учителното евангелие и неговите гръцки източници” [“Sermon 19 in Constantine of Preslav’s Didactic Gospel and Its Greek Sources”], Palaeobulgarica 46 (2022) 3–28. [17] I. P. Petrov, “The Greek Sources of Učitel’noe Evangelie Revisited: Sermon 20”, Palaeobulgarica 46(2) (2022) 3–28. [18] Y.-C. Chiao, O. Kraif, D. Laurent, T. M. H. Nguyen, N. Semmar, F. Stuck, J. Véronis, W. Zaghouani, Evaluation of multilingual text alignment systems: the ARCADE II project, in: Proc. of the 5th Language Resources and Evaluation Conference, ELRA, Genoa, 2006. [19] F. J. Och, H. Ney, “A Systematic Comparison of Various Statistical Alignment Models”,
Computational Linguistics 29 (2003) 19–51. [20] T. Pataridze, B. Kindt, “Text Alignment in Ancient Greek and Georgian: A Case-Study on the First Homily of Gregory of Nazianzus”, Journal of Data Mining & Digital Humanities Special Issue on Computer-Aided Processing of Intertextuality in Ancient Languages (2018). [21] R. von Waldenfels, L. Taseva, An der Schnittstelle von Korpuslinguistik und Paläoslavistik: Wörterverzeichnisse zu einer mittelalterlichen Handschrift als Keimzelle eines anotierten digitalen Korpus, in: B. Hansen (Ed.), Diachrone Aspekte slavischer Sprachen: für Ernst Hansack zum 65. Geburtstag, Otto Sagner, 2012, pp. 243–258. [22] A. Yli-Jyrä, J. Purhonen, M. Liljeqvist, A. Antturi, P. Nieminen, K. M. Räntilä, V. Luoto, HELFI: a Hebrew-Greek-Finnish parallel Bible corpus with cross-lingual morpheme alignment, in: Proc. of the 12th Language Resources and Evaluation Conference, ELRA, Marseille, 2020, pp. 4229–4236. [23] A. Rabus, “Recognizing Handwritten Text in Slavic Manuscripts: a Neural-Network Approach Using Transkribus”, Scripta & e-Scripta 19 (2019) 9–32.