D4.1.2 Whitepaper on the results using DaMSym on Nicene-Constantinopolitan Creed

Ir00000014 - Itserr D4.1.2 - Whitepaper on the results using DaMSym on Nicene-Constantinopolitan Creed 1 Document reference: ITSERR-WP4-DAMSYM-D4.1.2 Version number: 01.00 Status: FINAL Last revision date: DD/MM/YYYY by: UNIPA Verification date: DD/MM/YYYY by: Board Approval date: DD/MM/YYYY by: MUR Subject: IR00000014 - ITSERR Whitepaper on the results using DaMSym on Nicene-Constantinopolitan Creed Filename: ITSERR_WP4_Deliverable4.1.2_FINAL.docx Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 This document is available in the ITSERR WP Management document repository at: Change history Version Numbe r Date Status Summary of main or important changes 00.01 03/02/2025 WORKING Working version 00.02 15/03/2025 DRAFT Contributions collected by WP4 members 00.03 01/05/2025 DRAFT First revision - Elia Scapini 00.04 01/06/2025 DRAFT Second revision - Marianna Napolitano, Elia Scapini 00.05 12/07/2025 DRAFT Submission to reviewer 01.00 08/10/2025 FINAL Reviewed Document finalized Distribution List Name Company Role ITSERR Members All ITSERR partners 2 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Table of Contents1 Document Overview 5 Scope 5 Objectives 5 Structure 6 Team 7 State of the Art 8 Latin 8 Ancient Greek 12 Old Church Slavonic (OCS) and Church Slavonic 12 Arabic 15 Sanskrit 15 DaMSym prototypes 18 General 18 Access and Language Selection 18 Search Filters and Corpus Management 18 Semantic Retrieval Functionality: Sentence Similarity 19 Spin-off Tools: Sanskrit, Church Slavonic 20 Spin-off Tools for Sanskrit 20 Word Vector Similarity (Sanskrit Tool 1) 20 Find Occurrences and Co-occurrences (Sanskrit Tool 2) 20 Sandhi-splitter (Sanskrit Tool 3) 20 Spin-off tools for Church Slavonic 20 Script Converter (Church Slavonic Tool 1) 20 Lemmatizer (Church Slavonic Tool 2) 21 Combined Unicode Converter and Lemmatizer (Church Slavonic Tool 3) 21 Contribute Resources (Database Implementation) 21 Owner Evaluation (Implementation Control) 22 Surveys on DaMSym prototypes test 23 General 23 Adapted part of the survey to each language 24 Latin and Ancient Greek surveys 25 Church Slavonic surveys 27 Tool Usage and Current Practices 27 Desired Features in Digital Tools 28 Evaluation of the Tool Tested 28 Converter 28 Lemmatizer 28 Accuracy Ratings (1–7 scale) 29 6. Diachronic Corpus Development 29 Foreseen limitations: 29 1 Attached material contained in the D4.1.2. Attachments file. 3 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Conclusions 29 Arabic surveys 30 Sanskrit surveys 30 Perspectives of scientific advancement from the feedback of the scholar community 32 Latin 32 Ancient Greek 33 Church Slavonic 34 Arabic 35 Sanskrit 36 Conclusion 39 Transformer-based tools: Latin, Greek, Arabic 39 Sanskrit: sandhi-splitting and lemmatization 39 Slav converter and lemmatizer 40 Bibliography 42 Latin 42 Ancient Greek 44 Arabic 48 Sanskrit 50 Old Church Slavonic and Church Slavonic 52 Appendix 56 4 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Document Overview Scope The global purpose of WP4 is the semantic analysis of key terms found in our case study, namely the text of the Nicene-Constantinopolitan Creed in its various translations. The ITSERR project envisions that advanced knowledge for the history of religious studies will emerge from these studies, but also that new computer tools will be made available to scholars for extracting information from the ancient texts under consideration. Language models, particularly those based on Large Language Models (LLM), can be used to analyze the text of the Nicene-Constantinopolitan Creed in order to identify significant semantic patterns. These models will be optimized (e.g., fine-tuning operation) over a wide range of Christian texts for the different translation languages under consideration. To address this general purpose of identifying significant semantic patterns, rule-based or probabilistic architecture tools can be used. In particular, Large Language Models (LLMs) may eventually be finetuned to perform a wide range of semantic operations on the Christian texts for the various translation languages under consideration whose amount of data currently available allows it. With these models it will be possible to obtain advanced semantic representations to facilitate the identification of relationships, similarities and deep meanings for key terms in the Creed. Data mining will then be applied to this enriched corpus to extract detailed information about semantic relationships within the Creed, helping us to better understand the theological and historical significance of each term in the various translations and interpretations of the Nicene-Constantinopolitan Creed. Objectives Deliverable 4.1.2 in its original form had a deadline set for the 1st November 2024 and was written in the following format: This whitepaper illustrates the scientific advancement that can be achieved thanks to DaMSym, drawing from the experience of testers from the scholarly community. It presents the results of its operation “on the field”, describing both the scientific results and the properties database obtained by operating DaMSym on the Creed corpus. As such, this deliverable not only validates DaMSym’s prototype, but also provides researchers with new material for their studies. In any case, from the initial draft of this deliverable to the present, the comprehension of the project around DaMSym is progressed and we convened to intervene slightly modifying this description in order to align it to the state of the art in the field of NLP for ancient languages, to make it more detailed and corresponding to the direction we are following in the scholar and developing process. This update intervention has been applied also to the other 4 deliverables of WP4 (the first has been delivered, other three are ongoing) that now reflect the nature of our work in a more informative way. Therefore, the deliverable 4.1.2. has not been submitted by the 1st November 2024 but by July 11th 2025. Deliverable 4.1.2. appears now in this shape: This whitepaper illustrates the perspectives of scientific advancement that can be achieved thanks to the future platform DaMSym, drawing from the experience of the scholar community 5 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 dealing with similar research issues. It presents the results of its operation “on the field”, describing both the scientific results and the properties database obtained by operating DaMSym tools on the collected corpora in the different languages. As such, this deliverable not only validates DaMSym’s prototypes, but also provides researchers with new material for their studies. Worded like this, the deliverable 4.1.2 represents the natural continuation and progression of the first one (Deliverable 4.1.1 “Whitepaper on DaMSym”), building upon its foundational work by introducing a more technical and detailed exploration of the innovations developed within the DaMSym platform. It integrates advanced functionalities for semantic analysis, synonym management, and text processing, presenting these innovations to the scholarly community for evaluation and feedback. By incorporating case studies, prototype testing, and insights from scholars working with the languages addressed by WP4, this document not only refines the technical aspects of the project but also bridges the gap between development and practical application, ensuring that the tools meet the needs of the academic community while paving the way for future scientific advancements. Structure To best adhere to and mirror the purposes of the deliverable 4.1.2, we have opted to structure this document according to these points: state-of-the-art; DaMSym prototypes test; experience of the scholar community through surveys on DaMSym prototypes test; perspectives of scientific advancement. After an introductory part where it is outlined the scope of the document with the objectives, structure, and team involved in the project we open a chapter dedicated to the state of the art in the field. Here, the primary focus is on digital tools already available to the scholar community that works on the languages addressed by WP4, selected and reviewed with specific attention to the development and testing interests within the DaMSym platform. Therefore, this section aims to offer a detailed discussion of the state of the art, which includes an analysis of existing methodologies and tools relevant to the project. The document then delves into the testing of DaMSym prototypes, highlighting advanced functionalities such as semantic analysis, synonym management, contextual word selection, co-occurrence analysis, and visualization of terms within corpora. Specific case studies are presented, including the management and disambiguation of lemmas in Church Slavonic and the advanced search of semantically similar sentences in Ancient Greek and Latin. All these functionalities are offered to the evaluation of the scholarly community, detailing feedback and surveys conducted on the DaMSym prototypes. This section is adapted to address the specific needs and perspectives of scholars working with the languages addressed by WP4. All earnings from these investigations are budgeted in the session of the document that explores perspectives for scientific advancement, focusing on potential future developments and improvements in the field. Finally, the document concludes with a summary of findings and a bibliography, followed by an appendix that may contain supplementary materials or data relevant to the project. 6 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Overall, the document serves as a detailed report on the development, testing, and scholarly evaluation of advanced text analysis tools within the DaMSym platform, with a particular emphasis on the languages addressed by WP4. Team Marianna NAPOLITANO Unimore Premodern Slavic Usman NAWAZ Unipa – Liliana LO PRESTI Old Church Slavonic Ivana PANZECA Unipa Arabic Giovanni PUCCETTI ISTI-CNR, Pisa Arabic/Old Church Slavonic/Greek/Latin Maria CASSESE ISTI-CNR, Pisa Old Church Slavonic Igor SPANÒ Unipa Sanskrit Irfan ALÌ Unipa – Marco LA CASCIA Sanskrit Federico IEZZI Unimore Latin and Greek Elia SCAPINI Unimore Latin and Greek Marcella CORNIA Unimore (WP6) Latin and Greek Costanza BIANCHI Unimore Coptic Fabio QUATTRINI Unimore - Silvia CASCIANELLI Coptic Lorenzo BIANCHI CNR - Fabrizio SEBASTIANI Coptic The WP Leader is Fabrizio D’Avenia, while Costanza Bianchi (post-doc researcher - Unimore) and Marianna Napolitano (RTD-A -Unimore) are the Product owners. The team directly involves 4 Phd Candidates: Federico Iezzi (Unimore), Elia Scapini (Unimore), Usman Nawaz (UniPa) and Irfan Alì (UniPa). Ivanza Panzeca (UniPa) and Igor Spanò (UniPa) are the researchers (RTD-A) included in our work package. As of Oct. 29, the team expanded for the Slavic section with the participation of the Centro Nazionale (ISTI-CNR, Pisa), represented by Maria Cassese, Giovanni Puccetti and Fabrizio Sebastiani. Activities for Latin and Greek languages are fully covered by CNR-ISTI, Pisa, but the preparatory work for this deliverable - specifically, until December 2024 for Greek and until June 2025 for Latin-was covered by WP6. Humanities research for Church Slavonic is supported by WP3 (Criterion). Regarding Coptic, due to the scarcity of data, the work is covered in a partial way and is supervised by an assignee from the University of Modena and Reggio Emilia. 7 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 State of the Art This chapter will discuss the tools currently available that complement the DaMSym project, including a review of ongoing projects and research centers engaged in related fields. It will examine the existing resources and methodologies that provide a foundation for the study and development of our tools. Additionally, the chapter will identify the gaps and limitations in these tools and frameworks, highlighting the specific contributions and innovations that DaMSym brings to address these challenges and advance the research landscape. Latin Moving from the general task of performing the semantic analysis of the Nicene Constantinopolitan Creed in its early translation languages, members of WP4 dedicated to Latin (and Ancient Greek) have noted the lack in state-of-the-art NLP tools applied to Latin of a semantically searchable corpus with an easy-to-use tool to perform statement retrieval of words and sentences that are semantically related to a query. In the chapters dedicated to Latin of this Deliverable 4.1.2, we introduce the solution we have developed to face the lack of a statement retrieval tool, highlighting techniques, describing main features and detailing the outcomes of a test performed by domain experts. While no currently available software suite allows Latin scholars to retrieve statements, the methodology we introduce in section 2 of this Deliverable 4.1.2 benefits from several NLP tools, models, and datasets that we discuss in this session 1, which is dedicated to the state-of-the-art in the domain of statement retrieval performed with transformer-based Pre-trained Language Models (PLMs). We have chosen to move our steps from the transformer era2 because, since the introduction of self-attention heads (Vaswani et al., 2017) and the first comparison of a BERT model (Devlin et al. 2019), this seminal technology has set new state-of-the-art in the NLP pipelines allowing for contextual embeddings and unprecedented semantic representation capabilities. This seminal architecture was soon implemented into the first Latin BERT by Bammand and Burns (2020). The training data amounted to 642.7 million words from up to the 21st century. After Latin BERT, other solutions were attempted to develop PLMs for Latin, and, by the authors' admission, not always with good results. We can rapidly refer here to the ELECTRA model from Mercelis and Keersmaekers (2022) and the following DeBERTa (Mercelis 2024) that represented interesting attempts to produce a high-risk high-gain advancement in the state-of-the-art but, as the descriptive papers state, perform undeniably poorly. A likewise poor outcome has been admitted by Luis Antonio Vasquez3. Even if a rigorous benchmark is still to be performed and remains a great desideratum, the work of Riemenschneider and Frank (2023a) represents a significant improvement in the state-of-the-art. They have produced four models for Latin, two are monolingual and two are multilingual (En, AG, La). Of the monolingual models, one is an encoder-only LaBERTa (RoBERTa-based) and one is a T5 encoder-decoder model. Similarly, multilingual models are either encoder-only (PhilBERTa) or 3 https://huggingface.co/LuisAVasquez/simple-latin-bert-uncased. 2 For references on the previous period, one can start the survey from Sprugnoli, Moretti and Passarotti (2020). 8 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 encoder-decoder (PhilTa). As has become customary, we also refer here to their paper Exploring Large Language Models for Classical Philology but, for the sake of precision, this article only deals with Ancient Greek and multilingual models. In fact, for a brief description of the Latin models, one should refer to GitHub4. After a few months, Riemenschneider and Frank (2023b) introduced SphilBERTa, a multilingual model initialized on PhilBERTa and distilled from a teaching model tailored for semantic retrieval. To face changes in Latin over centuries and aiming to create a general parser for Latin that works effectively regardless of the Latin linguistic period, Behr 2024 introduces sentence embeddings. In this technique, the model encodes sentences of the same period into vectors with higher cosine similarity compared to sentences from another period. SBERT was produced by fine-tuning Latin BERT (Bamman and Burns, 2020) following Reimers and Gurevych (2019). Behr (2024) builds on the architecture of Attardi et al. (2021) based on the turn of Dozat and Manning (2018). A Sentence-Bert (SBERT) generates historical sentence embeddings that are concatenated with the output from a BiLSTM and are then passed into the four multilayered perceptrons (MLPs). According to Behr (2024), the inclusion of SBERT model in the pipeline did not significantly change the outcomes compared to previous state-of-the-art tools. In the EvaLatin 2024 campaign (see below), the model gained second place for the Dependency Parsing task. We have outlined the development of monolingual or multilingual models intentionally tailored for Latin and historical languages. This kind of study is pursued by universities and research centers with a vertical interest in improving NLP pipelines for under-represented historical languages. However, we can identify another strand of development represented by big tech companies. In fact, among these languages, Latin is privileged to benefit from the ongoing race to develop and improve multilingual models. Indeed, even before the model from Bamman and Burns (2020), the first multilingual m-BERT by Devlin et al. (2019) was tailored to represent more than 100 languages, including Latin. From then onwards, multilingual models have grown both in number and performance so listing them all would be difficult and probably pointless. Therefore, to track the exploitation of multilingual transformer technology for Latin NLP and, in doing so, to detect the development of models until today, a good starting point is considering models and tools presented at EvaLatin campaigns (Sprungoli et al. 2020,2022,2024). Organized by the Centro Interdisciplinare di Ricerche per la Computerizzazione dei Segni dell’Espressione (CIRCSE) at the Università Cattolica del Sacro Cuore in Milan and the University of Parma, EvaLatin is an evaluation campaign for NLP tools dedicated to Latin. The third edition occurred in 2024 and was preceded by the campaigns of 2022 and 2020. Each campaign sets shared tasks, scripts, guidelines and test data to compete against pursuing the evaluation of the performances to improve the state-of-the-art for Latin NLP. The results of these evaluation campaigns are published by Sprungoli et al. (2024, 2022, 2020)5. 5 All these editions of EvaLatin were integrated in the Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA). LT4HALA is co-located with the LREC-COLING, the Joint International Conference on Computational Linguistics, Language Resources and Evaluation, jointly organized by the ELRA Language Resources Association (ELRA) and the International Committee on Computational Linguistics (ICCL). ● 2024: Dependency Parsing and Emotion Polarity Detection ● 2022: Lemmatization, PoS tagging, Morphological Feature Identification ● 2020: Lemmatization, PoS tagging 4 https://github.com/Heidelberg-NLP/ancient-language-models. 9 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 One aspect that complicates Sanskrit processing is the presence of compound words formed by the phonetic union of terms. To understand the meaning of a compound word, it is essential to identify its constituent words and, consequently, split the compound into its parts. Sanskrit word splitting resembles other word segmentation tasks in Asian languages, such as Thai (Haruechaiyasak et al., 2008), Chinese, and Japanese (i.e., Kanji). To reduce feature engineering efforts (Zheng et al., 2013; Pei et al., 2014; Cai and Zhao, 2016; Yao and Huang, 2016), in Chinese language, compound words are segmented through a sequence labeling process to assign labels to each character of the compound word and segment the component words. Most NLP systems for Sanskrit word splitting integrate Panini’s phonetic and morphological rules (Cardona, G. 1997) with lexical resources. These systems rely on the application of formal methods (Huet, 2005; Goyal et al., 2007; Kulkarni and Shukl, 2009), or they use statistical approaches like Dirichlet processes (Natarajan and Charniak, 2011), finite-state methods (Mittal, 2010), graph queries (Krishna et al., 2016), or hybrid systems (Haruechaiyasak et al., 2008). The primary challenge in splitting words in Sanskrit is identifying the most semantically accurate segmentation among all possible splits of a compound word. (Krishna et al., 2016) addressed this issue by modeling word segmentation as a query expansion task within a path-constrained random walks (PCRW) framework. Several studies have suggested using recurrent neural networks to tackle this problem. The paper (Hellwig, 2015) introduces a neural network-based method that simultaneously performs compound splitting and Sandhi resolution in Sanskrit text. It employs Long Short-Term Memory cells for labeling tasks in Sanskrit analysis. Hellwig proposed an alternative method of Sandhi resolution in Sanskrit by developing a classifier that depends on the gold standard string splits. The source sequence represents a string split in phonemes, while the target sequence presents transformations applied to each phoneme. A classifier will then be trained to correctly split compounds and apply the appropriate Sandhi resolution in the process, generating the corresponding Sandhi rule. Five possible transformation rules, namely R1-5, are defined to guide the classification of phonemes in each string. The paper (Hellwig and Nehrdich, 2018) proposes an end-to-end trained neural network for Sanskrit tokenization, which jointly performs compound splitting and resolves phonetic mergers (Sandhi), requiring neither feature engineering nor outside linguistic resources but working on parallel versions of raw and segmented text alone. The models are designed to work at a character level and allow word splitting in a sequence labeling framework for the Sanskrit language. Their best model uses convolutional and recurrent elements with shortcut connections (rcNNshort). The work in (Dave et al., 2021) also presents neural networks for splitting Sanskrit compound words, focusing specifically on Sandhi splitting. The work formulates a sequence-to-sequence prediction problem, employing recurrent neural networks (RNNs) in a two-step process. Without additional lexical resources or a priori information, the model takes a compound word as input and generates the split of compound words as output. This data-driven approach demonstrates superior performance compared to existing methods without needing extra lexical or morphological resources. The few currently available Sandhi splitting tools use rule-based implementations (Linguistics, 2 24) and have significant drawbacks in terms of rigidity and inability to handle exceptions or variations. Three prominent Sandhi splitter tools are available in the open domain. These are (i) JNU splitter (Kumar, 2007), (ii) UoH splitter (Kumar et al., 2010), and finally, (iii) INRIA Sanskrit Reader Companion (Huet, 2003) (Goyal and Huet, 2013). All these tools addressed the challenge of splitting differently, but they work on a principle that is essentially the same. Some work has included attention to solve the problem. In (Aralikatte et al., 2018), the authors presented an attention-based deep learning method to determine the split position of compound words. Based on the predicted position of the split, the method determines the Sandhi components by graphically segmenting the 16 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 compound word. Compared to rule-based Sandhi models, knowing the split position allows to massively reduce the number of possible splits to check. In (Reddy et al., 2018), the model primarily identifies word splits and the correctness of unsandhied strings from a sandhied input string, which shall not take into account morphological and semantic details. The paper shows that the attention module improved the results significantly. 17 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 DaMSym prototypes This section describes the tools and prototypes that WP4 is currently developing thus showing how our tools place in the field of Digital Humanities and bring forward the state-of-the-art detailed in the previous chapter. After this overview of our tools, in the next chapter, we will discuss surveys, outcomes and desiderata for further implementation. General With the emergence of transformer-based architectures, there have been significant advancements in the field of NLP for ancient languages. Although various attempts have been made to develop pipelines tailored for different NLP tasks, to date, there remains a lack of user-friendly tools that enhance large-scale accessibility to such instruments. The ITSERR project, in general, aims to overcome this limitation. Specifically, with the DaMSym platform, the intention is to provide user-friendly tools for the semantics of languages for which the respective state-of-the-art approaches have been previously outlined. In the following sections, we will provide a description of the prototype of the product that we have submitted to specialists for each language. Access and Language Selection As part of the development of DaMSym - a platform for semantic sentence retrieval based on linguistic parameters-one of the core functionalities concerns user access and language selection. The system is integrated into the ITSERR Marketplace and can be used both by authenticated users, via their institutional or university credentials, and by unauthenticated users. While anonymous access allows the use of basic features, logging in unlocks access to advanced functionalities, described in other sections of the project. Once the user accesses the tool, they are prompted to select a language from the currently supported options: Ancient Greek, Latin, Church Slavonic, Arabic, and potentially Sanskrit, should the latter be integrated into the semantic retrieval tool in the future. The selected language serves as a central parameter, as it dynamically determines the availability of corpora and the corresponding search filters. In this way, the system ensures a consistent and targeted interaction with textual resources, adapting the interface to the linguistic specificities of the chosen material. Search Filters and Corpus Management Another key component in the architecture of DaMSym is the ability to filter corpora based on specific metadata, which vary according to the selected language. This level of granularity is crucial to enable precise and meaningful searches, especially given the diversity and complexity of the textual traditions involved. For Latin and Ancient Greek, which share similar metadata structures, the user is provided with several tools to refine their search: - The selection of one or more centuries via a dropdown menu (a desired feature); - A search bar for authors, which currently supports the selection of a single author, but may be extended in future versions to allow multiple selections; - The possibility to further filter the works associated with selected authors (a feature currently under development). 18 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Similar filtering systems are planned for Church Slavonic, Arabic, and Sanskrit, adapted to the specific characteristics of each textual tradition. In these cases, the main parameters include: - The selection of a historical period via a dropdown menu; - Optional filtering by author and genre, functionalities that are considered desirable and are currently under implementation. In the case of Sanskrit, even if it is not directly integrated into DaMSym, these features will nonetheless be implemented in related tools such as the co-occurrence detector or the sandhi splitter, contributing to a coherent ecosystem of advanced linguistic tools. Semantic Retrieval Functionality: Sentence Similarity Among the core components of the DaMSym tool is the semantic retrieval feature based on sentence similarity, designed to enable users to explore multilingual text corpora by retrieving utterances that are semantically similar to one or more input sentences. After selecting the desired language and applying filters to the reference corpus, users can activate an advanced search mode that leverages transformer-based language models to assess the degree of semantic similarity between textual segments. The functionality begins with a search bar where one or more input sentences can be entered. These inputs serve as the basis for semantic search, aiming to identify portions of the selected corpus that share content or meaning with the given sentences. A significant aspect of this feature is the ability to insert multiple input sentences, each associated with a differential weight that reflects its relative importance in the search. Although the current beta version of the tool supports only one additional sentence beyond the primary query, future development plans foresee extending this capability to allow for more nuanced input weighting. The results are ranked according to a similarity score, expressed as a percentage or other confidence indicator, with the most similar results appearing at the top of the list. This ranking mechanism ensures that users can quickly access the most relevant outputs, streamlining the consultation process. Another key feature of the platform is the association of metadata with the retrieved results. In the current implementations for Greek and Latin, for instance, each retrieved sentence is accompanied by the author’s name and the title of the work. However, to ensure full contextualization, it is strongly recommended that additional structural information-such as book, chapter, and paragraph or equivalent references-be included when available. Finally, the platform includes an active user feedback feature, accessible to logged-in users. Through simple evaluation buttons (e.g., “thumbs up” / “thumbs down”), users can indicate whether a retrieved sentence is semantically relevant to the original query. This feedback is collected to build an annotated dataset that can support further model development, for example through contrastive learning or reinforcement learning approaches, contributing to the continuous improvement of the system’s performance. 19 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Spin-off Tools: Sanskrit, Church Slavonic In addition to DaMSym, the semantic retrieval tool designed to facilitate the exploration of ancient texts, WP4 is also developing several spin-off tools specifically designed for individual languages. These tools are intended to support the scholarly community engaged in the study of ancient texts by providing specialized functionalities that enhance research capabilities. The languages involved are Sanskrit, Church Slavonic, and Coptic. Spin-off Tools for Sanskrit Word Vector Similarity (Sanskrit Tool 1) The Word Vector Similarity tool allows users to identify semantically similar words based on their context within a sentence. The user provides a target word (non-lemmatized) and a contextual sentence, and the system returns a list of sentences where words semantically similar to the target word appear. The results are ranked based on vector similarity, and each retrieved sentence is accompanied by relevant metadata, such as the source of the text. This tool is particularly useful for researchers wishing to explore semantic relationships within Sanskrit corpora and analyze how words function in various contexts. Find Occurrences and Co-occurrences (Sanskrit Tool 2) The Find Occurrences and Co-occurrences tool enables users to search for the occurrences and co-occurrences of specific words or phrases within Sanskrit texts. Users can enter a non-lemmatized word, a partial word, or a phrase fragment, and the system returns all relevant occurrences within the corpus or a filtered subset. The tool also supports co-occurrence searches, allowing users to define a proximity window (e.g., 5 or 50 words) to evaluate how frequently words appear together in the same context. The system displays usage frequency data for both individual occurrences and co-occurrences, providing insights into the contextual usage and frequency of specific terms. Sandhi-splitter (Sanskrit Tool 3) The Sandhi-splitter tool automatically detects and splits sandhi compounds in Sanskrit texts, helping researchers better understand the linguistic structure of Sanskrit sentences. After pasting or typing a Sanskrit text, the system identifies sandhi phenomena and highlights them in the text, using color coding or underlining. This tool is essential for scholars working with Sanskrit, as it simplifies the process of analyzing compound words and provides greater clarity in the interpretation of Sanskrit texts. Spin-off tools for Church Slavonic Script Converter (Church Slavonic Tool 1) The Script Converter tool is designed to assist users who work with ancient or non-standard text encodings by allowing them to convert problematic text into its correct Unicode representation for proper reading, analysis, and processing. This tool enables users to paste text with encoding or visualization issues into a text area, or upload text files in various formats (such as .txt, .doc). The system then processes the text, normalizing it into its correct Unicode form, and displays the 20 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 converted text in a separate output area. Users can easily copy the processed text for further use. Additionally, the tool provides options for downloading the processed text in txt. format. Lemmatizer (Church Slavonic Tool 2) The Lemmatizer tool helps users working with Church Slavonic texts by providing an automated process to lemmatize raw text. Users can either manually input text into a dedicated input field or upload Church Slavonic text files in multiple formats (such as .txt, .doc, or .docx). Once the text is processed, the system returns a lemmatized version, displaying it in a readable and structured format on the interface. This tool is invaluable for scholars seeking to work with standardized base forms of words for deeper linguistic analysis. Combined Unicode Converter and Lemmatizer (Church Slavonic Tool 3) The Combined Unicode Converter and Lemmatizer tool is an integrated solution for Church Slavonic text processing. This tool allows users to first convert their raw text into Unicode and then automatically lemmatize it, ensuring that the text is both properly standardized and ready for further analysis or processing. Users can paste their Church Slavonic text into the input field or upload text files (txt.), and the system will first convert the text into its correct Unicode representation. After that, the tool lemmatizes the text, returning the base forms of words. Users can preview the Unicode-converted text before proceeding with lemmatization to ensure the conversion was successful. The final processed text - both Unicode-converted and lemmatized - is displayed in a readable output area, and users have the option to download it in txt. format, facilitating continued research and analysis. This tool is still under development and has not yet been subjected to user testing or surveys, as there is still a substantial amount of work to be completed. Until the ongoing manual revision of lemma and lemmatization are finalized, the idea of this combined tool remains at the conceptual stage. Contribute Resources (Database Implementation) The WP4 platform allows external contributors to play an important role in expanding its database by submitting valuable texts that are not already included. This section is dedicated to users who wish to contribute their materials, such as transcriptions of manuscripts, fragments of works, or entire texts that can enrich the platform's content. Contributors can upload these texts in common formats like .txt or .pdf, ensuring flexibility and ease of use. Additionally, a series of fields are provided to gather essential metadata about the contributed work, including the author, title, estimated period of writing, and any available printed editions. There is also space for user comments. However, any text submitted by contributors is not immediately visible on the platform. Instead, it will undergo a review process by the platform owners, who will assess its quality and relevance before it is made publicly available. Another essential feature for contributors is the ability to report errors and suggest corrections to existing texts on the platform. When a contributor opens a specific text, they will find an option to "report error." This allows them to identify text strings they believe are incorrect and propose possible corrections. The process of suggesting fixes is straightforward: contributors select the problematic text and provide their suggestions. However, these changes will not be implemented 21 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 immediately. Instead, the correction suggestions will be submitted for evaluation by the platform owners, who will review them before deciding whether to adopt or reject the proposed changes. The platform also allows contributors to help improve the quality of metadata associated with existing texts in the database. Once a contributor accesses a specific text, they will see an "add metadata" section where they can propose missing or incomplete information. This includes key details such as the author, title, period of composition, and geographic area. Contributors can also suggest changes or updates to existing metadata. As with the error corrections, any metadata suggestions will not be applied automatically. Instead, the platform owners will review the contributions and decide whether to accept or reject them. This ensures that the metadata remains accurate and consistent. Owner Evaluation (Implementation Control) For platform owners, there is a dedicated section within DamMSym to manage the evaluation and approval of external contributions. Only owners have access to this area, where they can review any updates or suggestions made by contributors. A "notifications" section informs the owners of new contributions, such as error corrections or metadata additions, that are awaiting review. The platform owners have full control over these contributions and can either accept or reject them. If the owners decide to accept a suggestion, they may also modify or adjust the proposed changes before making them live. This process ensures that the platform maintains high-quality and accurate content while involving the scholarly community in the ongoing enhancement of the database. 22 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Surveys on DaMSym prototypes test This chapter is dedicated to the description of the surveys that were designed and administered to collect the experience of the scholarly community in its first approach to the beta version of the tools elaborated by the researchers of WP4. This chapter is supposed to be descriptive, outlining the nature of the surveys both in general and for the single languages alongside the conditions of the administration (profile of candidates, time, provided material etc.). Therefore, the evaluation part is reserved for the next chapter, dedicated to the discussion of the outcomes and to the qualification of next work steps. General Deliverable 4.1.2 explicitly recalls the “experience of the scholarly community”. We have considered that a good starting point to collect the experience of the scholarly community would have been the administration of a survey. WP4 initially produced a common draft for a general survey that was meant to be adapted to each language of interest of DaMSym, depending on the tools to be tested and on the state-of-the-art regarding each subject. We hypothesized four sections as the macro-structure of the survey: ● An initial section dedicated to what the tester knows and what shortcomings he finds in the available tools ● The main body of the survey is composed by two sub-sections ○ How the tester evaluates the guided test experience ○ How the tester evaluates the free test experience ● The conclusion requires a comprehensive assessment from the tester Regarding the twofold body of the survey, composed of a guided phase and an unsupervised phase, we have adopted different strategies depending on the necessities of the specific languages and tools. Each solution will be detailed below, but we can anticipate that, in general terms, we opened to the possibility to compose a vademecum for the testers, or to directly guide them in the test side-by-side during online meetings. Clearly, languages with tools provided with user interfaces were facilitated in offering testers a hands-on trial, while languages with prototypes in the form of source code and scripts opted to carry out the tester’s requests live by technician. As natural as it is, the tester must be an expert in the field and is required to have a good knowledge of the language. Besides that, we have also fixed some core guidelines for the selection of the testers. The minimum accepted is four scholars per language: ● 1/2: structured researcher ● 2/2: Ph.D. students in thesis delivery ● 1/2: Ph.D. students at the beginning of their career Another general rule we have established requires that surveys are arranged to take between 20 and 25 minutes from each tester. Despite that, the tests ended up demanding more time than initially planned. Moreover, for reasons of convenience, surveys are administered via a digital form that is easily filled using a personal computer, allowing also for a more feasible compounding and comparison of the results. The best solution proved to be Google Forms. 23 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 All the questionnaires with the respective answers can be found in the appendix to this document in the shape of an excel table. When provided in combination with the survey, also the vademecum to lead the user will be attached below. Adapted part of the survey to each language This section provides in-depth information about the single declension of the surveys that each language has adapted depending on its prerogatives. Latin and Ancient Greek surveys This survey is equally applicable to both Ancient Greek and Latin, as the tools being evaluated offer identical functionalities and are built on the same underlying architecture. The capabilities provided 24 Fig. 1: Screenshot of the first general part of the Ancient Greek survey Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 for analyzing Ancient Greek sources are mirrored in the tools designed for Latin, ensuring a consistent user experience across both languages. By maintaining this uniformity, the survey aims to gather comprehensive feedback that can be applied to enhance the tools for researchers working with either language. This approach allows for a streamlined evaluation process, ensuring that insights and improvements can be universally implemented to benefit all users, regardless of their focus on Ancient Greek or Latin. The survey is structured into five sections, each designed to gather detailed feedback on the use and effectiveness of a tool for analyzing Greek sources, with a particular focus on its functionality, usability, and potential improvements. The survey begins by examining the tools and features typically used by testers for Greek source analysis, highlighting the necessary improvements or additions they find essential. It then transitions into a general description of the tested tool, asking testers to describe its functionality, share their general impressions, and rate its intuitiveness. Suggestions for enhancing the tool's practicality and intuitiveness are also solicited, alongside questions about the tester's familiarity with artificial intelligence technologies and models like BERT. The third section focuses on a guided test, where testers followed a specific use case created by WP4. It surveys the relevance of the model's responses to the testers' queries, particularly within the first 20 responses, and gathers impressions on the type of answers provided. The fourth section shifts to an individual test, where testers were free to explore the tool independently. It asks for descriptions of the queries prompted, the relevance of the model's responses, and any additional impressions on the answers provided. Finally, the survey concludes with a section on the final balance, seeking feedback on any errors or inconsistencies found in the model's suggestions, its ability to capture the contextual meaning of Ancient Greek sentences, and the ease of interpreting the results. Testers are also asked to propose improvements for scaling the model to meet colleagues' needs, discuss how the model can enrich their research, and evaluate whether it offers new functionalities. The necessity of specific training for the tool's public distribution is also assessed, along with an opportunity for free comments. Throughout, the survey emphasizes the importance of detailed, constructive feedback to refine and enhance the tool's effectiveness and usability. The vademecum for both Ancient Greek and Latin is designed to guide users through a structured yet flexible testing process, ensuring a comprehensive evaluation of the tool's capabilities. The test is expected to take approximately 20-25 minutes, and all responses will be collected via email, with anonymity preserved during analysis. The test is divided into two main parts, followed by an overall assessment. The first part involves a hypothetical workflow that a scholar might follow using the tool, while the second part provides a "free exploration" space, where users are encouraged to document their steps critically and clearly, as these will form the basis of the analysis. In both the Ancient Greek and Latin tests, users are presented with a user interface that allows them to query a corpus of sources. Both the query and the corpus have been transformed into multidimensional vectors, which numerically condense various explicit and implicit information within the sentences. This enables searches that are not strictly bound to verbatim rules or n-grams, but rather expand the range of results to include semantically relevant correlations. 25 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Perspectives of scientific advancement from the feedback of the scholar community In this section, we explore the perspectives of scientific advancement as reflected in the experiences and expectations of the scholarly community that tested the DaMSym tools. Drawing from survey responses, it becomes clear that researchers see digital tools not just as aids for data retrieval, but as instruments capable of enhancing the depth and efficiency of textual analysis. Participants highlighted the potential of new functionalities to uncover meaningful connections across corpora, particularly when supported by authoritative sources and context-sensitive results. The scholarly community values tools that integrate traditional rigor with innovative features-such as intuitive interfaces, precise search capabilities, and exportable outputs-indicating that such tools can significantly advance the field of digital philology and classical studies when thoughtfully designed and adequately resourced. Latin The survey on the latin tool for semantic retrieval was performed by five selected Ph.D., researchers and scholars according to the aforementioned requirements between March 19 and April 16, 2025. The testers performed their evaluation independently one from the other and without the intervention of members of WP4, being provided with a link directing them to the tool, the link toward the survey and a vademecum. From their observations we can highlight the following: The overall impression of the testers was positive, depicting the tool as a valuable starting point for an initial source survey besides other already existing tools. The majority of the testers agreed that the tool is also quite intuitive, still all of them recommended setting up guidelines for new users, eventually with the inclusion of videos to detail commands, functionalities and suggestions into a practical user tutorial that could balance the lack of familiarity with transformer-based tools the tester shared. Dealing with the performances, the tool proved to be useful and promising, still keeping some inconsistencies in the average of valuable sentences retrieved in comparison with unacceptable results. With that, testers also faced some issues with the corpus, which contained repetitions, errors, non latin text, cryptic acronyms and unclear statements and did not allow the users to deepen their reading beyond the initial sentence retrieved. In sum, the two main issues to be addressed in the near future are that of the performance of the retrieval procedure and the cleaning of the corpus. A third element to join with them is the necessity of integrating more options in the selection of one or more authors and works. All these three improvements (retrieval performances, cleaning the corpus, integration of the user interface) were somehow expected by WP4 and will be addressed in the nearest future. A fourth one consisted in the possibility of exporting the retrieved material for personal use. Overall, the tool demonstrates significant potential for assisting research on Latin sources through semantic retrieval. It aligns well with user expectations, especially in early research phases. Continued development focusing on retrieval precision, corpus quality, interface usability, and export features would greatly enhance its scholarly value. Testers agreed that the tool could significantly speed up thematic research, aid in source discovery, and facilitate comparisons of doctrinal meanings. Its ability to combine queries was highlighted as particularly valuable and we will keep this valuable introduction. 32 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Ancient Greek The survey on the Ancient Greek tool for semantic retrieval was performed by five selected Ph.D., researchers and scholars according to the aforementioned requirements between March and April, 2025. The testers performed their evaluation independently one from the other and without the intervention of members of WP4, being provided with a link directing them to the tool, the link toward the survey and a vademecum. From their observations we can highlight the following: The overall impression of the testers was positive, depicting the tool as a valuable starting point for an initial source survey besides other already existing tools: the retrieved sentences were considered interesting, still lacking a little bit of context. The tool might be improved in usability even if some testers found it quite intuitive. In fact some of them recommended setting up guidelines for new users, eventually with the inclusion of explanation to detail commands, functionalities, sorting criteria and suggestions for a more efficient research across the corpus into a practical user tutorial that could balance the lack of familiarity with transformer-based tools the tester shared. Other suggestions regarded the topic of the black box, asking for a clearer explanation on the underlying processes of selection. Dealing with the performances, the tool proved to be useful and promising in its capabilities to capture contextual features in the retrieval process, showing better performances compared to the latin retrieval tool, still keeping some inconsistencies in the average of valuable sentences retrieved in comparison with unacceptable results. An unexpected observation from a tester reported that sometimes the results stick too much to the original text, making the retrieval more verbatim than semantic. We are therefore considering implementing a rule-based function to make the tool ignore exact matches in the retrieved passages. From the survey we did not collect strong observations on the corpus, meaning that the greek material is overall well organized and easily accessible with the retrieval tool. When it comes to the user experience with the interface, it emerged the necessity of integrating more options in the selection of one or more authors and works, or, to be more precise, the necessity to exclude author and works so to give the possibility to retrieve material from unexpected sources and forcing the neglecting of useless or obvious authors and works. In sum, the Ancient Greek tool survey opened for some integration: elimination of exact matches, neglect of specific authors and works, explanation on the ground elements of the probabilistic retrieval (black box). Overall, the tool demonstrates significant potential for assisting research on Ancient Greek sources through semantic retrieval. It aligns well with user expectations, especially in early research phases. Continued development focusing on integration of new features like the active exclusion of exact matches, works and authors would greatly enhance its scholarly value. Testers agreed that the tool could significantly speed up thematic research, aid in source discovery, and facilitate comparisons of doctrinal meanings. Its ability to combine queries was highlighted as particularly valuable and we will keep this valuable introduction but an adequate preparation and tutorial is necessary both from the technical (usage) and the speculative (probabilistic black box) point of view. 33 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Church Slavonic The results of the survey, though based on a limited sample, offer significant insights into the current state and future directions of two digital tool developments for Old Church Slavonic and Church Slavonic studies. What clearly emerges is a scholarly community that is both technologically engaged and acutely aware of the limitations of the existing infrastructure. Researchers are already navigating a fragmented ecosystem of tools, ranging from general-purpose NLP libraries like spaCy to language-specific repositories like Ruscorpora or Gorazd, often requiring creative adaptations to meet the specificities of historical Slavic texts. The responses reveal a strong desire not simply for more tools, but for integrated, linguistically-informed, and philologically responsible platforms capable of handling complex orthographic variation, non-Unicode character systems, and diachronic textual corpora. In this context, the DaMSym spin-off tools are received not as final solutions but as proofs of concept that reflect a growing awareness of the digital needs of Slavic studies. The converter and lemmatizer prototypes were evaluated with cautious optimism: appreciated for their potential, yet clearly in need of improvement in accuracy, character mapping, and contextual disambiguation. What is particularly striking is that respondents framed their expectations within a broader vision of scientific advancement: they did not only comment on usability but articulated how such tools could transform research methodologies, enable more systematic and comparative work across textual traditions, and contribute to the construction of diachronic corpora that are both computationally accessible and philologically rigorous. Several constructive suggestions were offered by participants, highlighting a clear desire for precision, transparency, and modularity. For instance, some respondents emphasized the need for a more detailed user guide and increased clarity in input/output operations, especially when dealing with non-Unicode or non-standard input. Others noted that the lemmatization process must be made more transparent, ideally through the visualization of ambiguous results and the possibility to select or correct lemmata manually. There were also calls for the inclusion of morphological tagging, which would allow the tools for Old Church Slavonic and Church Slavonic not only to return dictionary forms but to contribute meaningfully to syntactic and stylistic analysis. Moreover, one respondent suggested that an OCR/HTR pipeline would be crucial in future development, especially if the aim is to create an integrated workflow from manuscript to linguistic data. Nonetheless, the responses reflect more than a desire for technical refinement: they express a broader scientific expectation, in which digital tools are not only seen as aids to existing methodologies but as instruments capable of transforming the very practices of research. In particular, scholars emphasized the potential of computational tools to enable new forms of comparative, historical, and syntactic analysis across textual traditions and linguistic layers. In this light, tool development must go hand in hand with open data practices, interoperability with existing corpora, and a clear theoretical framework that makes explicit the linguistic assumptions behind the algorithms. What these responses make clear is that the development of digital tools for Church Slavonic studies must be an iterative and collaborative process, where scholars are not mere users but active 34 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 contributors in defining priorities and standards. The testing process has revealed the contours of a shared vision: one in which scientific advancement in the field depends on the successful integration of philological depth with computational innovation. Arabic The search engine created by Giovanni Puccetti aims to obtain semantic similarity results, searching for words or text passages within the corpus of Islamic texts7. The tool is a digital platform designed for text analysis, specifically focused on searching and comparing Arabic texts. It works by identifying and outputting results based on matching keywords between the user input and stored documents in open-source databases. For this purpose, the prototype was administered to five researchers in the field, selected for their linguistic and methodological skills. After providing a brief explanation of how the instrument works and how to carry out the test, it was carried out online individually and the specialists answered a questionnaire regarding the tools they usually use, an analysis of the tested software with related suggestions, features to add or errors to highlight. The following emerges from the questionnaire administered to the researchers: 1) Regarding the first part of the test, the discipline experts indicated the tools they usually use, demonstrating a certain familiarity with the different online models and software. In addition, they listed the features they most look for when using an online service: lexical and morphological support; digitised, searchable and copyable text based on reliable Arabic text recognition; functions for annotating, tagging and categorising Arabic texts; a platform with an intuitive interface offering automatic detection of text reuse; cross-referencing functionalities with specialised dictionaries or glossaries and grammatical analysis; and tools incorporating dialectal variations of Arabic and different scripts and styles. 2) The software shows the section of the text of interest at the top, then the similarity and a link to the full text. From the link it is possible to trace the author and period of the individual results. The tool is designed to allow semantic matching between an input text string and other output strings within the Open ITI project's corpus of digitised works. Each semantic match is accompanied by a degree of confidence in the tool and a semantic relationship expressed in cents. The tool is schematic and not very detailed, but simple and intuitive to use. The results are not very uniform and not very relevant to the queries entered. The actual correspondence must however be searched within the text itself contained in the corpus. To improve practicality: enhance the chronological range filter for searches; show the author and title of the work prominently in the search results, and include the extract of the reference text directly, with three or four lines; measuring similarity by the number of matching words could provide a more quantitative and direct method for comparing texts; improve semantic reconnaissance and learn to recognise the sources it uses by arranging them by subject, so as to focus on a more specific topic (Qur'anic exegesis, historical material, philosophical or literary texts). 7 https://kitab-project.org/explore/#/2022.2.7/?version=pri 35 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 3) The model returns semantically irrelevant sentences based on the query entered. 4) Experts inserted several texts in the interface (Qur'anic verses, tafsir, etc.). The outputs limited to the single pre-set century do not guarantee direct access to the portion of text in which the output is contained and it is still necessary to search the Open ITI corpus as well. 5) The model captures the genre of the sources, but not the true meaning and context of the sources, because the results are not relevant in terms of textual/contextual reuse. A model like this is more useful for finding links between sources, but it should definitely be optimised in the meaning of the sources it is asked to process by suggesting to incorporate a wider variety of sources, including texts of different genres, time periods, and both with and without diacritical marks. It should improve semantic reconnaissance and learn to recognise the sources it uses by placing them by subject, so as to focus on a more specific topic. The tool, with the appropriate implementations, could offer very useful functionality to the researcher in the future, e.g. combining several time periods, reaching the text area containing the output string, applying it to a specific literature, linking semantic meanings in different texts, etc… Sanskrit Regarding the test carried out on the tool created for the analysis of the dataset of Sanskrit texts, several experts on the Italian scene were contacted, mostly researchers or associate professors. Five of them agreed to carry out the test, four of them returned the completed questionnaire. The testers were chosen based on their proven knowledge of the Sanskrit language and based on the variety of their scientific interests and their methodological perspectives (linguistic, philological, literary, historical-religious). The test was carried out online, organising four meetings using Google's Meet platform. In fact, given the urgency of obtaining the first useful results to guide the work of the coming months, it was decided not to create an interface for users immediately. In any case, this objective will soon be realised in the coming weeks, thus making available the link with which to access the created tool. As they are experts in the field, during the meetings only a brief explanation was given on how the tool works and how the test is carried out. They were asked to formulate some queries on the search within the identified corpus of single words, pairs or even triplets of words, proposing the terms both in inflected forms (in the case of nouns) and in conjugated forms (in the case of verbs) or on the contrary, respectively in the pure thematic state or in radical forms. Each test took about 30 minutes and the experts were given the results of the various searches as they were completed so they could comment immediately. At the end of the test, each expert was sent the results of the searches carried out and the questionnaire to fill in and return. From the analysis of the answers received from the scholars who were given the test on the Sanskrit language, the following emerges: 1) Regarding the first question, concerning the type of research tools used to analyse Sanskrit texts, the experts demonstrate that they use the tools available online (dictionaries, corpora) and therefore favour the use of electronic and digital tools over analogue or paper ones. 2) Regarding the second, concerning the characteristics that guide the choice of using software or online services related to Sanskrit texts, scholars emphasise the importance of detailed linguistic analysis of the texts, focusing on three main aspects: a) Distribution of words – Identifying where specific terms appear, in which text and context. This helps to trace their use over time and to 36 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 understand their semantic relevance; b) Formation of compound words – Examining compounds to understand how words are combined and what linguistic structures emerge in the analysed texts; c) Text segmentation without sandhi – Studying texts in a form where sandhi (the phonetic changes at word junctions) is eliminated, to obtain greater precision in the analysis of lexical occurrences. Therefore, the main objective is to ensure accuracy and completeness, analysing the associations between words to identify usage patterns and potential contextual clusters, thus identifying linguistic and semantic trends in the texts. B) Regarding the third question, concerning improvements or additions to be made to existing digital tools for the analysis of Sanskrit sources, the answers highlight several aspects that are fundamental for the analysis and digitisation of Sanskrit texts: a) Access to complete and updated texts – It is essential to have the most recent versions of published texts to ensure accurate and up-to-date analysis; b) Possibility of separating sandhi – Separating words in the texts would allow for a better lexical and syntactic understanding, facilitating linguistic analysis; c) Metrical indications for poetic texts – The presence of metrical signals would help to identify the rhythmic structure and correctly interpret Sanskrit poetry; d) Uniformity in abbreviations – A coherent system of abbreviations for the texts would facilitate research and consultation in digitalised sources; e) Expansion of the corpus of digitised Sanskrit texts – Increasing the number of texts available in digital format would improve the possibilities for study and comparison between sources; f) Semantic and lexical analysis – Identifying synonyms, analysing the co-occurrence of words and studying semantic clusters would allow for a better understanding of the evolution of the meaning of terms and their relationships in Sanskrit texts. Therefore, the general objective would seem to be to improve the accessibility, linguistic analysis, and interpretation of Sanskrit texts through digital tools and advanced methodologies. 4) Regarding the fourth question, concerning the description of the tool, the answers highlight the following: The tool tested allows you to search for single words or groups of words in Sanskrit texts, also indicating their position in the verse or line of text. This tool allows you to map the distribution of a term in a wide range of textual sources, highlighting its role in the sentence. Its main value lies in its ability to identify not only the occurrences of a single term but also of multiple words together, specifying the texts in which they co-occur and their relative position. This feature helps to distinguish whether the connection between the terms is random or if there is a significant link to be explored. In summary, according to the experts, the tool offers a detailed analysis of the presence and position of words in Sanskrit texts, making it particularly useful for lexical, semantic and philological studies. 5) Regarding the fifth question, which asked testers to describe their general impressions, the evaluation highlights the effectiveness of this tool for analysing the vast corpus of Sanskrit texts. Its main advantage is the speed with which it allows users to identify the contexts in which a term appears. Although the search for single words is not particularly innovative, as other similar tools already exist, this software stands out for its ability to carry out simultaneous searches on multiple texts. Another important aspect is the function for searching groups of words, which allows their co-occurrences to be analysed in different textual contexts, an option considered unique and particularly useful for semantic investigation. 37 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 6) Regarding the evaluation of the performance of the tested tool, the experts expressed their satisfaction with the tool's performance, highlighting its speed and precision. The results of the searches carried out were relevant and correct. Even in the case of very specific queries, the tool provided exact and satisfactory answers, proving to be effective and reliable. The questions asked concerned words that often appear together in a fixed phrase or that could express integrated meanings, based on current research in which the experts are engaged. The experts emphasise the usefulness of the tool in searching for collocations and parallel passages in Sanskrit texts. They also believe that the tool can significantly speed up research, allowing queries to be made on a large corpus of sources in a single operation, instead of consulting the texts one by one. 7) Although they found the tool practical and intuitive, the following possible improvements are suggested: a) organise the results of the texts in an estimated chronological order, starting from the Vedas up to the epic poems and beyond; b) split the sandhi of the compounds. 8) Based on the queries proposed by the testers, the results obtained are reliable, as the tool's ability to determine the position in which the words appear is a useful indicator for assessing whether their relationship is causal or significant. On a scale of 0 to 10, two testers rated the tool at 8, one rated it at 9, and one rated it at 10. 38 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Conclusion Transformer-based tools: Latin, Greek, Arabic The evaluations of the Latin, Ancient Greek, and Arabic semantic retrieval tools reveal both shared challenges and opportunities for cross-disciplinary synergy. Despite their distinct linguistic and cultural heritage, all three tools demonstrate promising potential as innovative aids for scholarly research, particularly in streamlining early-stage source discovery and thematic analysis. Testers recognised their intuitive frameworks and alignment with foundational research needs, while also identifying critical areas for improvement that may or may not transcend linguistic boundaries. A primary commonality lies in retrieval performance and this may be affected both by the internal capabilities of the model or by the corpus condition. The Latin tool faced critiques over textual repetitions, errors, and irrelevant results, mirroring the Arabic tool’s struggles with semantic relevance and contextual accuracy. These issues underscore the necessity of rigorous corpus curation-removing inconsistencies, standardizing formats, and enabling dynamic filtering (e.g., by chronology, genre, or author). For the Arabic tool, integrating dialectal variations and diacritical marks emerged as additional priorities, reflecting the unique complexities of its corpus. Performance refinement is critical. While the Latin and Greek tools aim to balance retrieval of valuable versus irrelevant sentences, the Arabic tool must refine its semantic matching to prioritize contextual meaning over keyword overlap. User interface enhancements were universally recommended alongside proper training. All tools require improved tutorials-such as vademecum or video guides-to bridge gaps in user familiarity with transformer-based systems. Additionally, testers emphasized the need for more intuitive result displays, including direct access to extended text excerpts (Latin/Greek) or prominent author/title visibility (Arabic) but also the function of actively excluding exact matches, authors and works from the retrieval process. Export functionalities and advanced filtering (e.g., chronological or thematic parameters) were highlighted as essential for personal research workflows. Finally, clarification, refinement and expanding of the corpus is vital. Incorporating varied genres, authors, and historical periods will broaden these tools’ applicability. For Arabic, this includes diversifying beyond pre-set centuries; for Latin and Greek, enhancing author/work selection option acting improvements on the metadata already available. By addressing these shared stepsretrieval precision, interface usability, corpus optimization, and feature expansion-the tools can evolve into more performing and usable platforms. Sanskrit: sandhi-splitting and lemmatization We focus on predicting where to split compound Sanskrit words formed according to the Sandhi rules, without utilizing external resources such as phonetic or morphological analyzers. Using the Sandhi window, where transformation changes occur, we subsequently split compound words into their constituent parts based on the detected Sandhi window. To address this task, we propose a novel two-branch model based on a multi-head attention deep learning architecture. One branch is dedicated to learning features for split location prediction, which are then passed to the second branch responsible for performing the actual splitting. These two branches reinforce each other during training, thereby improving the accuracy of both tasks within a unified framework. Our model employs two Bi-LSTMs to encode character-level contextual information from both forward 39 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 and backward sequences. In the first branch, the bi-directional encoders work in conjunction with a multi-head attention module to focus on different regions of the compound word and identify the correct split location. The use of multi-head attention further enhances the contextual understanding of predicting splitting locations. In the second branch, we apply an encoder-decoder approach to generate the appropriate splits. Additionally, we used an augmented dataset that incorporates the UoH dataset, featuring challenging compound words drawn from the Digital Corpus of Sanskrit repository. This work not only advances the state of the art in Sanskrit compound splitting but also establishes a foundation for robust processing of complex morphological phenomena in digital humanities and natural language processing applications. Slav converter and lemmatizer The development and testing of digital tools for Old Church Slavonic and Church Slavonic within the DaMSym framework reveal a field marked by both promise and complexity. Despite the relatively small number of survey participants, the feedback demonstrates a clear scholarly demand for computational solutions that respond to the specific challenges posed by Slavic philological research-namely, the scattered sources, the heterogeneity of scripts, the inconsistent use of Unicode encoding, and the lack of standardization across corpora. The tested prototypes - the converter and the lemmatizer - have been evaluated as necessary preliminary steps toward the larger goal of enabling semantic retrieval and diachronic analysis. The converter, prompted by the limitations of PUA characters in existing corpora, is a tool which can allow textual normalization and visualization, especially in facilitating cross-platform readability. While some respondents encountered limitations in its immediate usability, suggestions such as expanding font support and refining character mapping have been very useful. The lemmatizer, in turn, was positively received for its utility and alignment with researchers' expectations, but also recognized as an evolving tool that would benefit from further training and integration with broader linguistic resources. Key suggestions - such as context-based disambiguation, compatibility with OCR/HTR workflows, and corpus expansion - underscore a forward-looking attitude in the community. A central issue that emerged from the responses is the fragmentation of currently available resources. Scholars rely on a patchwork of corpora, annotation standards, and toolkits - often combining several Slavic NLP libraries. This lack of interoperability significantly hinders cross-textual and diachronic comparison, and underscores the need for shared standards and centralization efforts. The idea of developing a diachronic dataset for Church Slavonic, explicitly mentioned by respondents, was met with strong enthusiasm but also with cautious realism. Participants recognized the high value of such a resource for both linguistic and cultural analysis, yet pointed to major obstacles: limited access to digitized and annotated texts, script and orthographic variation, and the absence of human and financial infrastructure to sustain a long-term editorial and technical effort. What emerges from the experience is a strong sense of shared priorities: preservation of philological integrity, linguistic transparency, and usability across diverse textual traditions. The vision expressed by scholars goes beyond tool functionality to touch on foundational methodological concerns. Calls for collaborative standardization, improved user interfaces, and 40 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 diachronic corpora indicate that the community is not merely reacting to current digital limitations but actively contributing to the definition of next-generation scholarly infrastructures. In sum, the development of tools for (Old) Church Slavonic - though still in its early stages - has the potential to open new research horizons. This potential, however, hinges on sustained dialogue between developers and scholars, careful attention to the specificity of historical Slavic material, and a collective investment in creating tools that are at once technically robust and philologically sound. 41 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 «Machine Learning for Ancient Languages: A Survey». Computational Linguistics, 2023, 703–47. https://doi.org/10.1162/coli_a_00481. Stopponi, Silvia, Nilo Pedrazzini, Saskia Peels-Matthey, Barbara McGillivray, e Malvina Nissim. «Natural Language Processing for Ancient Greek: Design, Advantages and Challenges of Language Models», 2 luglio 2024. https://doi.org/10.1075/dia.23013.sto. Yamshchikov, Ivan P., Alexey Tikhonov, Yorgos Pantis, Charlotte Schubert, e Jürgen Jost. «BERT in Plutarch’s Shadows». In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, a cura di Yoav Goldberg, Zornitsa Kozareva, e Yue Zhang, 6071–80. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics, 2022. https://doi.org/10.18653/v1/2022.emnlp-main.407. Yousef, Tariq, Chiara Palladino, e Farnoosh Shamsian. «Classical Philology in the Time of AI: Exploring the Potential of Parallel Corpora in Ancient Language». In Proceedings of the Ancient Language Processing Workshop, a cura di Adam Anderson, Shai Gordin, Bin Li, Yudong Liu, e Marco C. Passarotti, 179–92. Varna, Bulgaria: INCOMA Ltd., Shoumen, Bulgaria, 2023. https://aclanthology.org/2023.alp-1.21. Yousef, Tariq, Chiara Palladino, Farnoosh Shamsian, Anise d’Orange Ferreira, e Michel Ferreira dos Reis. «An automatic model and Gold Standard for translation alignment of Ancient Greek». In Proceedings of the Thirteenth Language Resources and Evaluation Conference, a cura di Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, et al., 5894–5905. Marseille, France: European Language Resources Association, 2022. https://aclanthology.org/2022.lrec-1.634. Yousef, Tariq, Chiara Palladino, David J. Wright, e Monica Berti. «Automatic Translation Alignment for Ancient Greek and Latin». In Proceedings of the Second Workshop on Language Technologies for Historical and Ancient Languages, a cura di Rachele Sprugnoli e Marco Passarotti, 101–7. Marseille, France: European Language Resources Association, 2022. https://aclanthology.org/2022.lt4hala-1.14. Arabic AL-Khalifa, H. Darwish, K., Mubarak, H., Ali, M., Elsayed, T. (2024). Proceedings of the 6th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT) with Shared Tasks on Arabic LLMs Hallucination and Dialect to MSA Machine Translation @ LREC-COLING 2024. Al-Thubaity, A. O. (2015). A 700M+ Arabic Corpus: KACST Arabic Corpus Design and Construction. Lang. Resour. Eval., 49(3):721–751. 48 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Belinkov, Y., Magidow, A., Romanov, M., Shmidman, A., Koppel, M. (2016). Shamela: A Large-Scale Historical Arabic Corpus. In Proceedings of the Workshop on Language Technology Resources and Tools for Digital Humanities (LT4DH), 45-53. Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazaré, P. E., ... & Jégou, H. (2024). The faiss library. arXiv preprint arXiv:2401.08281. Faghihi, Y., Teaching the Text Encoding Initiative; Context, Community and Collaboration. The Journal of Open Humanities Data, 2022. https://openhumanitiesdata.metajnl.com/articles/10.5334/johd.72. Faghihi, Y., Investigating the origins of Islamicate manuscripts using computational methods, The Digital Hum, 2021. https://blog.cdh.cam.ac.uk/2021/09/13/investigating-the-origins-of-islamicate-manuscripts-using-c omputational-methods/ Faghihi, Y., Jones, H., “Manuscript catalogues as data for research: from provenance to data decolonisation”. Digital Humanities Quarterly, 18 (3). https://www.digitalhumanities.org/dhq/vol/18/3/000751/000751.html Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The Interplay of Variant, Size, and Task Type in Arabic Pre-trained Language Models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 92–104, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based Model for Arabic Language Understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 9–15, Marseille, France. European Language Resource Association. Wissam Antoun, Fady Baly, and Hazem Hajj. 2021. AraELECTRA: Pre-Training Text Discriminators for Arabic Language Understanding. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 191–195, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Wissam Antoun, Fady Baly, and Hazem Hajj. 2021. AraGPT2: Pre-Trained Transformer for Arabic Language Generation. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 196–207, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. Habash, N., Rambow, O., Roth. R. (2009). MADA+TOKAN: A Toolkit for Arabic Tokenization, Diacritization, Morphological Disambiguation, POS Tagging, Stemming and Lemmatization. In Proceedings of the Second International Conference on Arabic Language Resources and Tools. Muhanna, E. (2016). The Digital Humanities and Islamic & Middle East Studies. De Gruyter. 49 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Prozorov, S. M., Romanov, M. G. (2003). Principles and procedures of extracting and processing the data from Arabic sources (based on materials of historic-and-biographical literature) / Original title: Metodika izvlecheniya i obrabotki informatsii iz arabskih istochnikov (na materiale istoriko-biograficheskoi literaturi). Oriens Vostok 4: 117-127. Romanov, M. Toward the Digital History of the Pre-Modern Muslim World: Developing Text-Mining Techniques for the Study of Arabic Biographical Collections (2014). In Analysis of Ancient and Medieval Texts and Manuscripts: Digital Approaches, 229-244. Brepols. Romanov, M. (2017). Algorithmic Analysis of Medieval Arabic Biographical Collections. Speculum 92/S1. DOI: 10.1086/693970, 0038-7134/2017/92S1-0009. Savant, S. (2016). The history of Arabic books in the digital age. BAR 28: 40-43. Savant, S. (2018). Digitizing the Textul Heritage of the Premodern Islamicate World: Principles and Planes. International Journal of Middle East Studies, 50: 103-109. Sengupta, N., Sahu, S. K., Jia, B., Katipomu, S., Li, H., Koto, F., ... & Xing, E. (2023). Jais and jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models. arXiv preprint arXiv:2308.16149. Sanskrit Hellwig, O. (2015). Using recurrent neural networks for joint compound splitting and sandhi resolution in Sanskrit. In 4th Biennial workshop on less-resourced languages. Goyal, P., Huet, G., Kulkarni, A., Scharf, P., and Bunker, R. (2012). A distributed platform for Sanskrit processing. In Proceedings of COLING 2012, pages 1011–1028. Haruechaiyasak, C., Kongyoung, S., and Dailey, M. (2008). A comparative study on thai word segmentation approaches. In 2008 5th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology, volume 1, pages 125–128. IEEE. Cardona, G.: Pāṇini: A Survey of Research. Motilal Banarsidass (1997). Huet, G. (2005). A functional toolkit for morphological and phonological processing, application to a sanskrit tagger. Journal of Functional Programming, 15(4):573 614. Goyal, P., Arora, V., and Behera, L. (2007). Analysis of sanskrit text: Parsing and semantic relations. In International Sanskrit Computational Linguistics Symposium, pages 200–218. Springer. Zheng, X., Chen, H., and Xu, T. (2013). Deep learning for chinese word segmentation and pos tagging. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 647–657. 50 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Cai, D. and Zhao, H., 2016. Neural word segmentation learning for Chinese. arXiv preprint arXiv:1606.04300. Yao, Y. and Huang, Z., 2016. Bi-directional LSTM recurrent neural network for Chinese word segmentation. In Neural Information Processing: 23rd International Conference, ICONIP 2016, Kyoto, Japan, October 16–21, 2016, Proceedings, Part IV 23 (pp. 345-353). Springer International Publishing. Pei, W., Ge, T., and Chang, B. (2014). Max-margin tensor neural network for chinese word segmentation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 293–303. Linguistics, S. C. (2002-24). Sanskrit computational linguistics. Kulkarni, A. and Shukl, D. (2009). Sanskrit morphological analyser: Some issues. Indian Linguistics, 70(1 4):169–177. Natarajan, A. and Charniak, E. (2011). s3-statistical sandhi splitting. In Proceedings of 5th international joint conference on natural language processing, pages 301–308. Mittal, V. (2010). Automatic sanskrit segmentizer using finite state transducers. In Proceedings of the ACL 2010 Student Research Workshop, pages 85–90. Krishna, A., Santra, B., Satuluri, P., Bandaru, S. P., Faldu, B., Singh, Y., and Goyal, P. (2016). Word segmentation in sanskrit using path constrained random walks. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 494–504. Hellwig, O. and Nehrdich, S. (2018). Sanskrit word segmentation using character-level recurrent and convolutional neural networks. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2754–2763. Dave, S., Singh, A. K., AP, D. P., and Lall, P. B. (2021). Neural compound-word (sandhi) generation and splitting in sanskrit language. In Proceedings of the 3rd ACM India Joint International Conference on Data Science & Management of Data (8th ACM IKDD CODS&26thCOMAD), pages 171–177. Kumar, S. (2007). Sandhi splitter and analyzer for sanskrit (with reference to ac sandhi). Submitted, 2007. Huet, G. (2003). Towards computational processing of sanskrit. In International Conference on Natural Language Processing (ICON). CiteSeer. Goyal, P. and Huet, G. (2013). Completeness analysis of a sanskrit reader. In International Symposium on Sanskrit Computational Linguistics, pages 130–171. Aralikatte, R., Gantayat, N., Panwar, N., Sankaran, A., and Mani, S. (2018). Sanskrit sandhi splitting using seq2 (seq)ˆ 2. arXiv preprint arXiv:1801.00428. 51 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Reddy, V., Krishna, A., Sharma, V. D., Gupta, P., Goyal, P., et al. (2018). Building a word segmenter for sanskrit overnight. arXiv preprint arXiv:1802.06185. Old Church Slavonic and Church Slavonic 3. Afanasev, I., “A Hybrid Lemmatiser for Old Church Slavonic”, HSE Working Papers: Linguistics (2021) 1–19. Andreev, A./Miklas, H./Shardt, Y., “Proposal to Encode Combining Glagolitic Letters in Unicode”, Ponomar Project (2022). Andreev, A./Shardt, Y./Simmons, N., “Church Slavonic Typography in Unicode”, Unicode Technical Note 41 (2015) 1–97. Andreev, A./Shardt, Y./Simmons, N., “Proposal to Encode some Additional Symbols used in Church Slavonic Text (Revision 2)”, Ponomar Project, Slavonic Computing Initiative (2015). Baranov, A.N., “Metaphor and Metonymy in the Structure of the Lexical Meaning of a Word”, Kazan Linguistic Journal 1 (2016) 50–61. Berkes, L., “Perspectives and Challenges in Editing Documentary Papyri Online: A Report on Born Digital Editions through Papyri.info”, in N. Reggiani (ed.), Digital Papyrology II (Berlin: De Gruyter, 2018) 75–86. Birnbaum, D.J. et al., Character Set Standardization for Early Cyrillic Writing after Unicode 5.1. A White Paper prepared on behalf of the Commission for Computer Processing of Slavic Manuscripts and Early Printed Books to the International Committee of Slavists (2008). Devlin, J. et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”, arXiv.org (2018). Ferro, M.C./Romoli, F., “Gli attributi di Dio: Per una traduzione slavo ecclesiastico-russo-italiano del lessico religioso e teologico-filosofico”, Studi Slavistici 10/1 (2013) 121–38. Garzaniti, M., Gli slavi (Milan: Jaca Book, 1998). Garzaniti, M., La letteratura slava medievale (Milan: Jaca Book, 2003) Graf, K., “Online Tools for Handwritten Text Recognition: A Comparative Review of Transkribus and eScriptorium for Byzantine Paleography”, Archivalia, 22 January 2025. Haug, D.T.T./Eckhoff, H.M./Lindstad, A./Bech, K. et al., “Universal Dependencies for Old Church Slavonic: The PROIEL Treebank”, Universal Dependencies 2.10 (2022). Jouravel, A./Renje, E./Lendvai, P./Rabus, A., “Assessing Automatic Sentence Segmentation in Medieval Slavic Texts”, in Proceedings of the Digital Humanities 2024 Conference, Washington, DC (2024) 6–9 August. Jung, H., “On the Origin of the Glagolitic Alphabet”, SCRIPTA 5 (2013) 105–30. 52 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Kempgen, S., “Unicode 5.1, Old Church Slavonic, Remaining Problems – and Solutions, including OpenType Features”, in Slovo: Towards a Digital Library of South Slavic Manuscripts. Proceedings of the International Conference, 21–26 February 2008, Sofia, Bulgaria (Sofia: Slovo, 2008) 200–19. Lendvai, P. et al., “Domain-Adapting BERT for Attributing Manuscript, Century and Region in Pre-Modern Slavic Texts”, in N. Tahmasebi et al. (ed.), Proceedings of the 4th Workshop on Computational Approaches to Historical Language Change (Singapore: ACL, 2023) 15–21 Lendvai, P. et al., “A Workflow for HTR-Postprocessing, Labeling and Classifying Diachronic and Regional Variation in Pre-Modern Slavic Texts”, in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (n.p.: ELRA, 2024) 2039–48. Lendvai, P./Reichel, U./Jouravel, A./Rabus, A./Renje, E., “Retrieval of Parallelizable Texts Across Church Slavic Variants”, in Scherrer, Y./Jauhiainen, T./Ljubešić, N./Nakov, P./Tiedemann, J./Zampieri, M. (eds.), Proceedings of the 12th Workshop on NLP for Similar Languages, Varieties and Dialects, Abu Dhabi (2025) 105–114. Pantelia, M.C., “The TLG and the Canon”, The Classical Review 74/1 (2024) 37–9. Pedrazzini, N., “Exploiting Cross‑Dialectal Gold Syntax for Low‑Resource Historical Languages: Towards a Generic Parser for Pre‑Modern Slavic”, arXiv (2020) . Peng, J./Simon, M./Qi, P. et al., “Stanza: A Python NLP Toolkit for Many Human Languages”, Transactions of the Association for Computational Linguistics 8/1 (2020) 339–351. Rabus, A./Petrov, I.N., “Linguistic Analysis of Church Slavonic Documents: A Mixed-Methods Approach”, Scando-Slavica 69/1 (2023) 25–38. Scherrer, Y./Rabus, A./Mocken, S., “New Developments in Tagging Pre-Modern Orthodox Slavic Texts”, Scripta & e-Scripta 18 (2018) 9–33. Tomelleri, V.S., “When Church Slavonic Meets Latin: Tradition vs. Innovation”, in De Gruyter Mouton, Berlin/Boston (2022) 201–232. Totomanova, A.-M., “Electronic Research Infrastructure for Bulgarian Medieval Written Heritage: History and Perspectives”, Diacronia 14 (2021) 1–9. Totomanova, A./Atanasova, D. (ed.), Cyrillo‑Methodian Readings 2019: Proceedings of the Conference Organized by the Department of Cyril and Methodius Studies, Sofia University “Sv. Kliment Ohridski” (8 May 2019) (Sofia: Sofia University Press, 2020). Winslow, J.J., “Old Church Slavonic Phonemes: The Problem of /j/ and /ě, a/ after Palatals”, Die Welt der Slaven 67 (2022) 296–322. Ziffer, G., “Storia dello slavo ecclesiastico e del russo: A proposito di alcuni lavori di H. Keipert”, Europa Orientalis 8 (1989) 275–300. Ziffer, G., “Aspetti storico-linguistici della cristianizzazione del mondo slavo”, in R. De Giorgi/S. Garzonio/G. Ziffer (ed.), Gli studi slavistici in Italia oggi: Atti del IV Congresso Italiano di Slavistica, Udine, 20–23 settembre 2006 (Udine: Forum, 2007) 41–50. 53 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Online sources Church Slavonic fonts in Unicode, <https://sci.ponomar.net/fonts.html>. Corpus Corporum, <https://mlat.uzh.ch/>. Cyrillic blocks in Unicode standard version 16.0, <https://www.unicode.org/charts/PDF/U0400.pdf>. Cyrillomethodiana, <https://histdict.uni-sofia.bg/>. eScriptorium, <https://msia.escriptorium.fr/>. Glagolitic Supplement for Unicode, <https://www.unicode.org/charts/PDF/U1E000.pdf>. IFAOGrec Unicode, <https://www.ifao.egnet.net/publications/outils/polices/>. Manuscript, <http://manuscripts.ru/>. Old Church Slavonic Text Processing, <https://github.com/usmannawaz01/OCS-Text-Processing>. Papyri.info, <https://papyri.info/>. Perseus Digital Library, <https://www.perseus.tufts.edu/hopper/>. PHI Latin Texts, <https://latin.packhum.org/>. Shafarik Unicode Font Documentation, Beta Version 1.0, <https://www.ponomar.net/files/documentation_3.5.pdf>. SlavConvert, <https://zenodo.org/records/7821709>. Stanza NLP toolkit, “Stanza: A Python Natural Language Processing Toolkit”, Stanford NLP Group, online resource, https://stanfordnlp.github.io/stanza/. TEI P5 Guidelines, <https://tei-c.org/guidelines/p5/>. 54 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 TLG, <https://stephanus.tlg.uci.edu/>. Transkribus, <https://www.transkribus.org/it>. Unicode 16.0 Character Code Charts, <https://www.unicode.org/charts/>. Unicode chart, <https://www.ssec.wisc.edu/~tomw/java/unicode.html>. Universal Dependencies project’s Old East Slavic, <https://github.com/UniversalDependencies/UD_Old_East_Slavic-RNC>. UTF-8, a Transformation Format of ISO 10646, RFC 3629, <https://datatracker.ietf.org/doc/html/rfc3629>. 55 Ir00000014 - Itserr Status: FINAL ITSERR-WP4-DAMSYM-D4.1.2 Version: 01.00 Attachments 1. Maria Cassese, Giovanni Puccetti, Marianna Napolitano, and Andrea Esuli. 2025. DIACU: A dataset for the DIAchronic analysis of Church Slavonic. In Proceedings of the 10th Workshop on Slavic Natural Language Processing (Slavic NLP 2025), pages 101–107, Vienna, Austria. Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.bsnlp-1.12 https://aclanthology.org/2025.bsnlp-1.12/ 2. Surveys: vademecum for Latin and Ancient Greek and outcomes of the surveys for each language are included in the attachments file. Surveys are derived from Google Forms submitted to experts in the linguistic field and finally turned into excel sheets with one participant for each column. 56