The goal of this paper is to present a tool for the semantic analysis of texts in Old Church Slavonic and Church Slavonic. The need for a tool capable of enabling semantic analysis of texts in this historical language emerged within the framework of a collaborative project that aims to apply Natural Language Processing (NLP) techniques to religious texts in various languages, using the Nicene-Constantinopolitan Creed as a case study. This research is part of a broader scholarly initiative dedicated to the 1700th Anniversary of the Ecumenical Council of Nicaea [1] (325-2025).

The Creed is a doctrinal text agreed by bishops convened in council that conveys the fundamental tenets of the Christian faith. As a text regularly recited in the liturgy and quoted in catechesis it exerts a significant impact on the target culture. Accordingly, its semantic implications, particularly the theological significance of linguistic and terminological variation, are of primary relevance. As shown by Gezen’s analysis of the Creed in Istoriia slavianskogo perevoda Simvola very (published in 1884; second edition: 2015), the examination of manuscript evidence reveals a wide range of textual variation, which changes according to the historical period and the regional context. More specifically, Gezen’s study demonstrates not only the multiplicity of these variations, but also makes it possible to classify the Slavic translation of the Nicene-Constantinopolitan Creed into three distinct editions. Among these, the second edition survives in two different redactions which, according to the scholar, differ both in terms of text and script: one is written in Glagolitic characters, the other in Cyrillic. Furthermore, variation is also attested within the same redaction. This is the case, for example, with the third edition, in which both adjectives истиннагѡ (istinnago - true) and животворѧщаго (zhivotvoriashchago - life-giving) appear in the eighth article ([Napolitano 2025]), as will be discussed in greater detail in the following sections of this paper. Building on Gezen’s analysis, it emerges that up to the seventeenth century three different versions of the Creed could still be found in liturgical texts. This view seems to be confirmed by the comparative analysis aimed at assessing their linguistic and historical significance. Additional differences become evident when comparing, for instance, pre-reform texts other than those used by Meyendorff, particularly those originating from the Polish-Lithuanian Commonwealth, where Catholic influence was more prominent.

The present study is study thus originates from the intent to investigate the Slavonic translations of the Creed (Сѵмволъ вѣры Символ веры - Simvol very) and to explore the impact of different geographic areas and historical phases on the language’s transmission. A particularly important point of reference is the revision of the Creed undertaken during the liturgical reform of Patriarch Nikon, [2] with a specific focus on assessing the role of the Greek liturgical tradition in the implementation of these revisions.

Following a review of the relevant literature on NLP for Old Church Slavonic (OCS) and Church Slavonic in paragraph 2, the paper discusses the creation of a reference dataset for the semantic retrieval tool (paragraph 3). This phase is particularly complex due to the extreme scarcity of resources available for this language, the fragmentation of digitized texts across multiple unsystematized digital libraries, and the persistent issues related to character standardization, especially the widespread use of non-Unicode characters. The following paragraphs will present the methodology adopted for the development of the tool for the semantic retrieval (paragraph 4), not only from a technical perspective but also through the analysis of evaluation forms completed by a selection of expert scholars (paragraph 5). This testing group includes six researchers specializing in Slavic philology, many of whom are actively involved in digital humanities projects related to OCS and Church Slavonic manuscripts. Their feedback will serve to assess the perceived reliability of the tool’s output with respect to the primary objective of analyzing the Nicene-Constantinopolitan Creed, and to identify the strengths and limitations of this novel contribution to NLP applied to (Old) Church Slavonic.

Despite the growing effectiveness of transformer-based text representations developed through large language models (LLMs), these techniques remain underdeveloped for historical languages such as Old Church Slavonic (OCS) and Church Slavonic, especially due to the lack of training data.

This scarcity of data has had a significant impact on the work conducted in the field of this project. In addition, the focus will be on diachronic information, which is of particular relevance for research on Old Church Slavonic and Church Slavonic.

The following pages will briefly present the most relevant projects and studies in the field of Old Church Slavonic and Church Slavonic research, in order to outline the context in which our tool has been developed. This overview also aims to highlight the current limitations and future challenges that must be addressed to consolidate LLMs in this field of study and to improve their results.

More generally, NLP for OCS and Church Slavonic continues to face severe constraints, primarily due to the absence of annotated datasets sufficient for training robust computational models. An additional obstacle stems from the limited number of historical linguists capable of providing expert human evaluation of model performances, which hinders rigorous qualitative assessment. This remains a critical bottleneck, despite recent advances that have improved performance in low-resource NLP through transformer-based architectures, including historical languages.[3]

To date, research efforts have primarily focused on two major historical Slavic varieties: Old Church Slavonic in the strict sense, and Old East Slavic (which represents the attested form used in English to refer to texts written in древнерусский - drevneruskii). The Universal Dependencies framework[4] and the Stanza toolkit[5] have played pivotal roles in providing both annotated corpora and processing tools for these languages.

Previous attempts to develop NLP resources for pre-modern Orthodox Slavic texts have employed a variety of strategies. Rule-based approaches to morphological analysis have been notably explored by [Baranov et al. 2007] and by #moldovan_etal who did this wily while at the Russian Academy of Sciences[6] Alternatively, tagging methodology from Modern Russian has been proposed for developing a new pipeline in this field.[7]

Current state-of-the-art approaches to automatic Part-of-Speech (PoS) tagging and morphological analysis for OCS have achieved results comparable to those in modern NLP applications. While many corpora offer either gold-standard dependency annotations or morphological tagging, some, like the PROIEL Treebank for Old Church Slavonic,[8] provide both.

Addressing this gap is essential for advanced syntactic analysis of historical Slavic texts. In the following, several projects that have implemented these techniques will be discussed.

The BogoSlov project presents a hybrid computational framework for identifying biblical quotations and allusions in OCS texts. The system addresses the challenge of detecting intertextual references that range from verbatim citations to paraphrased allusions. It combines 1) rule-based methods using string similarity and longest common subsequence metrics with 2) semantic methods leveraging embeddings from SentenceTransformers (Sentence-Bert or SBERT) and RoBERTa on corpora from the late antiquity and early middle ages. These methods are integrated into a CTS URN [9] - based model[10] for referencing source texts (primarily the Psalter and Gospels in OCS) and linking them to homiletic and hagiographic target texts such as Vita Constantini and Vita Methodii. The framework includes a graphical annotation interface to support expert-in-the-loop validation, visualization of alignments, and manual metadata enrichment. Designed to be extensible to other historical languages such as Latin or Greek, the framework integrates symbolic and neural approaches to support philologically sensitive annotation at scale.[11] In a related study,[12] the authors propose a mixed-methods, quantitative-qualitative framework for analyzing Church Slavonic texts, using a 16th-century East Slavic manuscript and a 1519 South Slavic printed SluzhebnikSlužebnik. Transcriptions were generated using Transkribus (HTR), followed by a quantitative analysis and traditional philological validation. This hybrid approach revealed differences across regional varieties and between manuscript and printed forms, showcasing the potential of NLP-philology synergies in historical Slavic research. The project forms part of a broader investigation of Orthodox Slavic linguistic varieties (15th - 18th centuries) at the University of Łódź, which uses corpus methods and automatic handwriting recognition to study vernacular-Church Slavonic interactions. The proposed processing workflow includes three components: 1) HTR post-editing and ground truth creation; 2) sentence-level classification with transfer learning models for temporal and regional attribution; 3) token-level annotation of linguistically salient features. The pipeline was evaluated on both HTR-derived and manually corrected corpora, with attention to systematic HTR errors, tool limitations, and accessibility for non-technical scholars. This modular system enhances the state of the art in historical Slavic NLP by integrating automatic correction, supervised classification, and fine-grained annotation in a reproducible and user-adaptable framework.

[Scherrer, Rabus, and Mocken n.d.] laid foundational groundwork with New Developments in Tagging Pre-modern Orthodox Slavic Texts, comparing a statistical tagger (TnT)[13], MarMoT,[14] which is used in particular to address problems with large tagsets, such as full morphological tagging, and a bi-LSTM character-level tagger (CLSTM) on annotated [Eckhoff et al. 2018]; [Eckhoff and Haug 2021]. Their findings show that modern tagging architectures can achieve accuracies up to 95–96% without orthographic normalization.[15] The study also explored cross-linguistic and cross-corpus transfer learning, demonstrating that synthetic or dialectal resources do not reveal substantial performance improvements over models trained on low resources historical languages. This underlines the importance of high-quality annotated corpora and establishes morphosyntactic tagging as a reliable foundation for future semantic modeling in premodern Slavic NLP.

Building on this foundation, more complex PoS tagging systems have been proposed. One of the most advanced systems to date is described in [Berdichevskii, Eckhoff, and Gavrilova 2016], which combines a statistical PoS tagger (TnT)[16] with a rule-based tagging module and a series of pre-processing steps. This hybrid model achieved an accuracy of 92.7% for part-of-speech tagging and 81.5% for morphological features. While these results are encouraging, recent advances in statistical and neural tagging models, as well as in normalization techniques, suggest that further improvements are possible. Subsequent work has built upon BEG16’s framework using the TOROT treebank, which has since been expanded, increasing both training data size and test set diversity. Additional sources, such as the Old Church Slavonic portion of the PROIEL corpus, have been included to broaden the linguistic basis of experimentation. Beyond the original use of the TnT tagger, more experiments have incorporated MarMoT and taggers based on deep neural networks. Both the PROIEL and TOROT datasets have been converted to the Universal Dependencies format, allowing for standardized PoS, morphological, and syntactic annotation: this is a key step for enabling cross-linguistic comparisons and annotation projection. In terms of evaluation, the field has moved beyond simple accuracy and Hamming distance. Micro-averaged F1-scores are being used to provide more granular insight into tagging performance.

Regarding character normalization, earlier approaches focused on lowercasing and removal of diacritics, ligatures, and orthographic variants. More recent experiments updated normalization routines and also explored tagging on unnormalized data, a scenario highly relevant for digital Paleoslavic studies. Efforts have been made to integrate external resources, such as the unannotated PLDR corpus (parallel Old Russian-Modern Russian) and the Modern Russian SynTagRus treebank in UD format, but these have not yielded substantial gains in accuracy.[17].

In parallel to these tagging efforts, the focus of research on OCS and Church Slavonic has increasingly shifted toward approaches using words and sentence embeddings. SuchThese approaches open for more novel downstream tasks such as text classification and embedding evaluation, semantic similarity and semantic retrieval. This paradigm shift also raises the issue of the representation’s quality automatically learned from historical language data.

Whithin In this framework, a major bottleneck for these downstream tasks lies at the tokenization stage. As [Dorkin and Sirts 2024] highlight, standard tokenization tools are poorly suited to the linguistic properties of Old Church Slavonic and Old East Slavic,. These languages, which frequently result in high out-of-vocabulary rates and unrecognized characters. Therefore, the development of custom tokenizers for specialized embedding models is essential for achieving meaningful representations in this domain.

Benchmark datasets and pre-trained models fine-tuned for language-specific tasks are notably sparse for historical Slavic languages. Embedding-based models, originally designed for applications in contemporary languages, often fail to align with the goals of diachronic linguistic research applied to historical languages. In fact, the main objective of diachronic research pursued so far, is to uncover patterns and mechanisms of orthographic and grammatical variation within the data, rather than to facilitate semantic question answering for accessing document content. In contrast, the focus of our study, as outlined in the following sections, is the development of a tool designed to retrieve sentences semantically related to a specific input query.

As [Lendvai et al. 2025] emphasize, techniques such as embedding similarity and retrieval augmented generation, a methodology originally developed by [Lewis et al. 2020], perform well on modern languages but are ill-suited for capturing the phenomena of interest in historical corpora. A further shortcoming of current approaches is the limited attention to temporal and geographic variation within historical Slavic corpora. This specific study also adopts an approach to diachronic linguistics, in which the focus lies in tracing orthographic and grammatical evolution over time rather than in fulfilling semantically-driven queries aimed at producing interpretative outcomes. To address this issue the authors also propose a classifier,[18] using manuscript metadata to assign spatial-temporal attributes, by both copying time and region at the sentence level. Building on this, [Lendvai et al. 2025] assembled a diachronically and dialectally heterogeneous corpus including South Slavic manuscripts from the 10th-11th centuries and East Slavic recensions from the 15th -17th centuries. This dataset enabled the exploration of linguistic influence between traditions. The analytical framework combined string-based representations, such as character n-grams and term frequency - inverse document frequency (TF-IDF)[19] , with computational similarity techniques, including sequence alignment, local string matching, and k-nearest neighbor (kNN) search based on cosine distance. These were systematically contrasted with neural methods using text embeddings derived from SBERT and BERT. Candidate ranking in retrieval tasks was performed via cosine similarity, providing a quantitative basis for evaluating both traditional and neural approaches to modeling linguistic variation.

Semantic similarity, which assesses how similar two sentences are, based on vectorial representations that encoded high-level meanings of the word, has become a major focus of recent research, especially in low-resource language [Deshpande et al. 2023]. Dense text retrieval with LLMs has emerged as a particularly promising avenue, though still relatively new. In particular, for LLMs to create vectorial representations of text that encode semantic meaning, they require training over large scale corpora, notable examples of this approach are the works from [Artetxe and Schwenk 2019] and [Reimers and Gurevych 2019]. Adaptation of these architectures to historical languages, with their unique orthographic variants and fragmentary documentation, remains an open and rapidly evolving research challenge.

As discussed above, in the scope of historical language, orthographic normalization must be approached with caution. Non-normalized forms often carry important information about temporal or regional provenance. These features are essential for capturing linguistic change and for understanding the sociocultural dynamics that underlie orthographic and morphosyntactic variation. Also for this reason, Lendvai’s study of the Vita of Paul and Juliana, a text with Church Slavonic and Old East Slavic versions, is a notable case study.[20] Here, sentences from the earlier recension were used as queries to retrieve aligned content from the later version, preserving historical variation while testing retrieval performance. The dataset for this experiment was created through meticulous manual alignment at the word level. Crucial to this work was the identification of parallel texts across different manuscript sources. This allows for reuse of metadata typically absent or inconsistently annotated in the original documents.

The attribution of texts to specific times and places has also been explored in a work by [Lendvai et al. 2023], who purposely fine-tuned BERT models. Their corpus includes six manually transcribed texts drawn from medieval manuscripts and early printed editions from Southern and Eastern Europe, dating from the 10th to 18th centuries. Dating and regional attribution were supported by codicological, paleographic, and linguistic criteria. All texts were written in Cyrillic and exhibit non-normalized orthographies typical of the religious, non-vernacular literary genre. The corpus spans a wide variety of Old Church Slavonic recensions shaped by distinct cultural influences, reflected in the orthographic, lexical, and morphosyntactic diversity of the material. A particularly salient phenomenon traced here is the so called Second South Slavic influence, which manifests in the 14th -15th century Rus’ scribal tradition through the incorporation of South Slavic stylistic norms into East Slavic manuscript culture.[21] BERT-based models[22] have been applied to sentence-level classification, segmenting texts into sentence-like units. For fine-tuning, the study employed publicly available models on Hugging Face, such as 1) bert-base-multilingual-uncased and Cyrillic-specific architectures like 2) KoichiYasuoka/bert-base-slavic-cyrillic-upos and 3) anon-submission-mk/bert-base-macedonian-bulgarian-cased. BERT achieved high accuracy in three attribution tasks: a) manuscript identification; b) dating by century attribution; c) regional classification, outperforming Cyrillic-specialized models. The study demonstrated the competitiveness of general-purpose multilingual transformers when fine-tuned on historical Slavic data. This is coherent with our findings, as described in our own case study from section 5. (Analysis of results).

The issue of text normalization remains unresolved, along with the need to develop analytical methodologies capable of working with non-normalized texts. Such approaches are essential to preserve the orthographic and stylistic peculiarities of the language, thereby enabling in-depth philological analysis that can engage in dialogue with semantic studies and their historical, theological, and cultural significance.

The following section explores the challenges related to data scarcity and text processing, with a particular focus on data collection aimed at developing a semantic retrieval tool for OCS and Church Slavonic. This development is part of the broader DaMSym project, which supports semantic similarity search not only for OCS and Church Slavonic but also for Latin, Greek, Arabic, Sanskrit, and a combined Latin/Greek multilingual configuration.

As previously discussed, the scarcity of texts in Old Church Slavonic and Church Slavonic has posed, and continues to pose, a significant challenge for the development of NLP tools for these languages, and more broadly, for the training of large language models (LLMs). This challenge can be traced to three main issues concerning the availability of plain-text resources:

1. Texts are scattered across various online repositories, stored in heterogeneous formats, and are not encoded using a consistent Unicode font, which makes their aggregated use difficult;
2. The editorial criteria adopted are often unclear, and there is no shared agreement on their accuracy or reliability;
3. Progress in OCR and HTR methodologies remains limited, and the amount of manual intervention still required to improve the performance of tools such as Transkribus and eScriptorium is too substantial to enable a large-scale creation of high-quality resources.

We accounted for these issues when collecting the diachronic dataset and when developing the semantic retrieval tool presented in this work.

For this reason, we initially chose to focus exclusively on Old Church Slavonic and Church Slavonic textual resources written in the Cyrillic script, thereby excluding those preserved in Glagolitic[23]. This decision is directly related to the approach taken in the development of the dataset and the selection of source texts. We retained the original transcriptions provided by the selected editions but opted not to include transliterations. Although the texts in question have already undergone editorial handling especially in terms of normalizationprocessing, we argue that preserving the original script allows for a more faithful and nuanced analysis, both linguistically and semantically.

This principle also informed our decision to exclude the Corpus Cyrillo-Methodianum Helsingiense (CCMH)[24] which was made available through the Language Bank of Finland (Kielipankki) and the CLARINO infrastructure.[25] While the CCMH corpus offers transliterated versions of Old Church Slavonic texts that may serve particular use cases, it does not meet the requirements of our framework for semantic modeling and linguistic annotation, which relies on maintaining the original Cyrillic representations.

Given these considerations, texts written in the Glagolitic alphabet were excluded due to the substantial challenges involved in standardizing such sources for qualitative analysis, particularly those arising from the use of Private Use Area (PUA) characters.[26] As previously noted, even working exclusively with Cyrillic-script materials has required the application of multiple Unicode fonts and, in some cases, the implementation of customized mappings between PUA and Unicode code points to ensure accurate text rendering. This process, which is essential for conducting reliable qualitative analyses, would become significantly more complex with the inclusion of additional scripts.

A further example of texts that were intentionally excluded, is the Orthodox portal Azbuyka,[27] particularly the section Богослужение – Священное Писание (Bogosluzhenie – Sviashchennoe Pisanie; Divine Liturgy – Holy Scripture). These texts represent modern editions of manuscripts reflecting liturgical language from approximately the 16th-17th centuries, a later phase in the historical development of Church Slavonic. However, they were excluded from the corpus due to the lack of reliable metadata[28], particularly concerning the precise dating and provenance of the editions.

Therefore, our current focus remains limited to Cyrillic sources for which precise metadata information could be retrieved. Once the primary issues related to PUA encoding are resolved, we intend to expand the dataset to include Glagolitic texts, enriching both its historical scope and linguistic depth in support of the broader objectives of this diachronic resource.

As previously mentioned, the inclusion of Cyrillic sources underwent further refinement based on the issue of Private Use Area (PUA) characters. For this reason, only texts with available and compatible fonts enabling accurate rendering were included. In the case of the Cyrillomethodiana,[29] in addition to the font itself, PUA-to-Unicode mappings were introduced to facilitate proper front-end visualization. These mappings were designed not only to support display within the semantic retrieval tool, but also to serve as the foundation for a Unicode character converter currently under development.

The mapping table was compiled through meticulous cross-referencing with the charts found in Unicode Technical Note #41 [30] specifically the sections detailing the Glagolitic Supplement and Cyrillic Unicode blocks, as defined in Unicode Standard Version 16.0.[31] Furthermore, for diacritical marks, comparative analysis was carried out with the Unicode implementation for Ancient Greek, using the IFAOGrec Unicode[32] resource as a reference. The map includes 43 characters, organized according to the following information: 1) appearance on website; 2) PUA – PUA Unicode; 3) Cyrillic Char.; 4) Cyrillic Unicode. Among these, 16 are diacritical signs and 23 are characters. Of the latter, 2 belong to the Glagolitic alphabet block: 1) U+2C51 - Glagolitic Small Letter Yat; 2) U+2C49 - Glagolitic Small Letter Out. Both have been added in Unicode version 4.1 (2005). and belong to the Glagolitic block of the Basic Multilingual Plane. [33]

In conclusion, the dataset used for the development of the semantic analysis tool is compiled from the following repositories and sources:

1. Cyrillomethodiana (uni-sofia.bg);
2. Syntacticus (syntacticus.org);
3. Old Russian Hagiographic Literature (spbu.ru);
4. a subset of the National Corpus of the Russian Language (ruscorpora.ru), specifically the texts available through Universal Dependencies;
5. a sample from the Ruthenian Corpus (UD_Old_East_Slavic-Ruthenian);
6. the Cyrillic manuscript transcriptions from the 11th century and the transcriptions of manuscripts from the Kazanskaya Collection, available through the manuscript.ru project.

In total the documents are: 683. The semantic retrieval carried out by the tool is performed on these texts. Following the principle that the text to be retrieved should not be contained in the documents on which the tool operates, among these there are four documents that are known to contain witnesses of the Creed.

This chapter describes the technical aspects underlying the functionalities of the retrieval search system. Here, we describe the methodology used, the details of the text parsing and preparation, and the models used and the impact of these choices on the final tool.

[Feliksov 2018] has noted that the language of Orthodox doctrine, although well consolidated in liturgical practice and written tradition, lacks a structured grammar or a comprehensive descriptive model capable of accounting for the syntactic, semantic, and morphological peculiarities of its core terms. From his perspective, this absence leads to:

1. difficulties in the linguistic analysis of vocabulary: as Orthodox vocabulary contains semantically dense units deeply rooted in theological meaning, many of which do not easily conform to standard grammatical categories. The main challenges involve:
   • the polysemy of key terms (e.g., благодать –blagodat- grace),
   • their strong interdependence on liturgical context,
   • the lack of direct correspondences in modern and everyday language.
2. difficulties in grammatical and semantic encoding: the current linguistic description of Orthodox lexicon is fragmented and often limited to encyclopedic dictionaries or glossaries, which typically do not offer coherent morphological or syntactic analysis. This creates obstacles for structured teaching and for computational formalization (e.g., annotation, parsing, semantic processing) of this type of vocabulary.

To address these issues, a semantic approach is proposed, that aims to describe the complexity and richness of the vocabulary of Orthodox faith in Church Slavonic, by incorporating:

• analysis of syntactic roles within the sentence,
• discourse function in liturgical usage,
• semantic and theological relations among terms.

The linguistic methodology underpinning this study is based on the ontological theory of language, grounded in the ideas of Orthodox doctrine on being and the word, as developed by [Florenskii 1990]; [Losev 1993]; [Kamchatnov 1998]; [Bulgakov 2007]; [Aksakov 2011] and [Postovalova 2022]. Central to this theory is the concept of the “eidosphere”, understood as an instrument to designate the set of lexical and phraseological units, as well as non-predicative (nominative) word combinations "generated" by a given eidos through the act of naming. The study adopts the term “lexico-eidetic group” (LEG), as introduced by Kamchatnov. This notion is rooted in the Christian intuition that God “creates simultaneously both the thing and its idea, in their indivisibility and inseparability” [Kamchatnov and Nikolina 2008, 37].

The ontological approach to language fundamentally relies on the deductive method of semantic, accordingly, the first step toward constructing a semantic classification of Orthodox faith vocabulary should consist in a systematic description of the “world” of eidē that constitute the content of the eidosphere of “Orthodox Faith.”

Our work on the study of the semantic value of sentence and word is based on a different methodological premise, although it aims to capture the polysemy of specific terms and to investigate the diachronic evolution of Orthodox vocabulary.

First and foremost, our research aims to analyze individual sentences starting from the context in which they appear, thus proceeding from semantic retrieval at the phrasal level. Additionally, the research is conducted with reference to a specific text, namely the Nicene-Constantinopolitan Creed, a brief doctrinal and liturgical formulation that conveys the core principles of the Christian faith.

Gezen, as previously mentioned, in his book Istoriia slavianskogo perevoda Simvola very Istoria slavianskova perevoda Simvol veri (published in 1884; second edition: 2015), notes that the Slavic translation of the Creed predates the missionary work of Saints Cyril and Methodius, although its final redaction could only occur after the invention of the Slavic alphabet. As a text regularly recited during the liturgy, the Creed has exerted a profound influence on the target culture. Therefore, due to its catechetical role, the semantic implications of its vocabulary, particularly the theological significance of linguistic and terminological variants, are of special relevance.

This research, consequently, is motivated by the intention to trace the influence of the Greek liturgical tradition on the Slavonic versions of the Creed, and to study their reception across different regions and historical stages in the development and use of the language.

The research sets out to examine two key questions:

1. the ability of the tool to retrieve phrases that are semantically related to selected articles from the Nicene-Constantinopolitan Creed;
2. the tool’s ability to provide useful references for analyzing the textual revisions introduced in the Creed during the liturgical reform of Patriarch Nikon (1653-1656).

These revisions, which aimed to restore consistency between the Russian Orthodox liturgical tradition and the Byzantine tradition, involved not only orthographic corrections, but also terminological changes whose historical and theological significance has been widely discussed in the scholarly literature.

For this reason, the presentation of results will proceed in two parts:

• The first section will be based on an in-depth analysis of two selected articles from the Creed and will address both objectives.
• The second section will examine the results of user tests conducted independently by six scholars and professors with expertise in the field of Slavic studies and digital humanities.

These tests will focus on:

1. the functionality and usefulness of the tool
2. the accuracy of its outputs in relation to the self-selected article and version of the Nicene-Constantinopolitan Creed.

The development of a semantic tool for Old Church Slavonic (OCS) and Church Slavonic texts, presented in this study, took place within a research context marked by significant methodological and technical challenges. The main difficulties concern the a) scarcity of digitalized linguistic resources available in plain text; b) the fragmentation of existing resources across multiple online repositories; c) the fact that texts in these repositories are not always encoded using Unicode-compliant characters; d) the absence of shared normalization standards, and the intrinsic complexity of the Slavonic manuscript tradition, often characterized by fragmentary documentation and insufficient metadata. Moreover, the particularities of the scripts, orthographic features, and diacritical signs, though difficult to preserve in standard NLP tools, are essential for conducting studies on historical, geographical, and linguistic variation. Such variation is foundational not only for Slavic philological research but also for broader investigations into the history of these regions and the cultural influences and contaminations that shaped them.

One of the most important takeaways from this overview is that the issue of text normalization remains unsolved. There is an urgent need to develop analytical methodologies capable of processing non-normalized texts. Such approaches are crucial in preserving the orthographic and stylistic idiosyncrasies of the language, which are essential not only for diachronic linguistic research, but also for rigorous philological analysis that can engage meaningfully with the theological, cultural, and semantic dimensions of the texts.

In addition, the literature review clearly shows that, although transformer-based language models have achieved strong performance in modern contexts and, to some extent, in high-resource historical languages, they remain insufficiently adapted to low-resources corpora marked by high diachronic, orthographic, and dialectal variation, such as Old Church Slavonic and Church Slavonic. The solutions proposed thus far have predominantly relied on hybrid (rule-based and statistical) approaches, yielding promising results in part-of-speech tagging and morphological analysis, yet still falling short in deep semantic modeling and diachronic variation. As noted, current research efforts in the field have primarily focused on two key objectives: a) dating manuscripts and determining their geographic provenance, and b) identifying biblical citations and allusions within OCS texts.

In this context, the tool introduced in our study represents an original and methodologically grounded contribution to computational research on (Old) Church Slavonic texts. The platform enables semantic queries that go beyond literal matching, identifying conceptually related expressions even when they exhibit morphosyntactic variation or reformulation. The research has focused on semantic similarity analysis using the Nicene-Constantinopolitan Creed as a case study. Due to its brevity and doctrinal centrality, the Creed functions as a core liturgical formula with high doctrinal content, providing an ideal reference point for analyzing both its source culture and patterns of reception across regions and centuries.

The results, particularly in comparing different versions of the Nicene-Constantinopolitan Creed, demonstrate the tool’s capacity to support lexical change analysis over time, offering results that are both linguistically and semantically significant.

The adoption of a semantic-oriented embedding model (Jina Embeddings V3) has proven to be a sound choice for the liturgical-Slavonic domain, where vocabulary is highly connotative and theologically dense. The system has shown its ability to return relevant outputs even in the absence of exact textual matches, such as in the replacement of “нѣсть несть кѡнца конца” with “не бꙋдетъ кѡнца не будет конца,” or the disappearance of the adjective “истиннагѡистиннаго” (“true”) referring to the Holy Spirit in pre-Nikonian reform texts.

Structured testing conducted with experts in Slavic studies and digital humanities confirmed the tool’s usefulness, while also highlighting areas for improvement, particularly regarding the conceptual coherence of retrieved results and the system’s interpretability when confronted with non-standard query formulations. These observations underscore the necessity of supplementing the platform with a methodological guide to assist users in query construction and in critically interpreting the system’s outputs.

Overall, the tool emerges as an innovative resource for semantic inquiry into (Old) Church Slavonic texts, capable of offering new perspectives in theological, historical, and linguistic research. It not only facilitates information retrieval but also serves as a conceptual exploration instrument, enabling the tracing of lexical and semantic variants across time. This approach lays the groundwork for deeper integration between traditional philology and semantic technologies in the study of premodern religious texts.

The development of our tool has confirmed the need to address the limitations outlined above, particularly the ongoing data scarcity, which remains a major bottleneck. It has also highlighted the importance of improving OCR and HTR methodologies to ensure the acquisition of reliable digital resources. Although, as noted in Pedrazzini’s 2020 study, computational methods for early Slavic texts are making steady progress, particularly in relation to the digitization of primary sources[35] (Rabus and Petrov, 2023; Lendvai et al., 2024), the use of HTR and OCR techniques for the digitization of Old Church Slavonic and Church Slavonic texts remains an unresolved challenge. The contributions of these scholars represent a critical foundation for future work, but additional resources are still required to optimize model performance in Transkribus and to advance alternative frameworks such as eScriptorium. Improvements in this area will be essential for increasing the availability of high-quality textual data, thereby enabling the eventual training of LLMs for historical Slavic languages.

A further significant challenge was the widespread use of PUA (Private Use Area) characters. These characters frequently manifested in the dataset as what is colloquially referred to as “tofu”, placeholder glyphs (e.g., boxes or question marks) that indicate missing font support.

Accordingly, future development phases of our tool will also require the integration of extended PUA-Unicode mappings to ensure consistent and accurate frontend rendering. Although this issue does not affect the tool’s semantic retrieval capabilities, it can nonetheless lead to confusion for users and compromise the overall user experience.

This paper is the outcome of a collaborative work developed within WP4 of the ITSERR project, entitled DaMSym - Data Mining: the Nicene-Constantinopolitan Symbolum, and focused on the specific case study of Old Church Slavonic and Church Slavonic. In particular, sections 1 and 6 were jointly authored by both contributors; sections 2, 3, and 5 were written by Marianna Napolitano; and section 4 was authored by Giovanni Puccetti.

This work was supported by the PNRR (National Recovery and Resilience Plan) project Italian Strengthening of ESFRI RI Resilience (ITSERR) founded by the European Union—NextGenerationEU (CUP:B53C22001770006).