Corpora and Processing Tools for Non-standard Contemporary and Diachronic Balkan Slavic

Teodora Vukovic, Nora Muheim, Olivier Winistörfer, Ivan Simko, Anastasia Makarova, Sanja BradjanRANLP Student Research Workshop 2019 · DOI

Introduction

The Balkan Slavic (BS) linguistic area, encompassing Bulgarian, Macedonian, and the transitional Torlak dialects, represents a complex landscape of linguistic variation. This region is a primary component of the Balkansprachbund, where centuries of contact have led to shared structural features—such as the loss of case systems and the development of post-positive definite articles—that transcend genetic linguistic boundaries. Despite this richness, digital linguistic resources have historically focused on standardized forms, leaving non-standard and diachronic varieties under-represented. Researchers at the University of Zurich are addressing this "standard language bias" by developing a unified digital infrastructure designed specifically for the study of non-standard Balkan Slavic.

The Challenge of Standard Language Bias

Most Natural Language Processing (NLP) tools, including part-of-speech (PoS) taggers and syntactic parsers, are optimized for the rigid norms of official languages like standard Serbian or Bulgarian. However, dialects and historical manuscripts exhibit "micro-variation," where grammatical structures can shift significantly across small geographical distances or short time spans. Because these varieties are often oral or found in non-standardized historical texts, they lack the large-scale "gold standard" datasets required for modern computational analysis. Until now, this data gap has made it nearly impossible to perform rigorous, computational comparisons of linguistic change across the region.

Methodological Framework

To bridge this gap, the project employs a standardized pipeline to ensure that data from disparate regions and eras are mutually comparable. The methodology involves:

  • Transcription and Digitization: Utilizing Exmaralda for spoken audio and Transkribus (an AI-powered OCR tool) for the transcription of historical manuscripts.
  • Annotation Standards: Adopting the MULTEXT-East tag-set for morphosyntactic description (MSD), which is better suited for the morphological complexity of Eastern European languages than generic frameworks. These annotations are subsequently mapped to Universal Dependencies (UD) to ensure broader compatibility with global linguistic tools.
  • Tool Adaptation: Rather than building new tools from scratch, the team adapts high-performing models like the ReLDI tagger. By fine-tuning these models with a combination of standard language data and small amounts of manually corrected dialect-specific data, they achieve high accuracy on non-standard inputs.

Core Datasets

The infrastructure is built around three primary corpora that are methodologically aligned to allow for cross-linguistic inquiry:

  1. The Torlak Corpus: This corpus focuses on the transitional dialects between Serbian and Bulgarian. It includes contemporary fieldwork from the Timok and Lužnica regions (~350k tokens) alongside historical data from the early 20th century.
  2. The Macedonian Dialect Corpus: A pioneering collection of spoken Macedonian, aggregating transcripts from the 1950s–70s with modern field data from multi-ethnic regions like Ohrid and Prespa.
  3. Pre-Standardized Bulgarian Corpus: Focusing on the 16th to 19th centuries, specifically the Damaskini tradition (vernacular religious texts), this corpus captures the Bulgarian language in a state of flux before modern norms were established.

Key Results and Insights

The project demonstrates that standard language tools can be effectively leveraged for dialects if given a relatively small amount of manually corrected "gold standard" data for fine-tuning. The adapted PoS tagger for the Torlak dialect achieved a remarkable 92.9% accuracy for PoS labels and 93.9% for lemmas.

Beyond technical metrics, the project’s most significant contribution is the uniformity of its processing. By using the same tagging standards across Torlak, Macedonian, and Bulgarian, researchers can finally perform large-scale contrastive analyses. This allows for the investigation of diachronic and diatopic questions, such as whether the definite article behaves the same way in 17th-century vernacular Bulgarian as it does in modern Torlak dialects.

Limitations and Future Outlook

While the project has successfully integrated diverse sources—from 18th-century manuscripts to modern YouTube recordings—challenges remain. Historical Bulgarian data currently requires more manual annotation to improve tagger accuracy, and some older Macedonian collections suffer from fragmentary metadata regarding speaker backgrounds.

Future work focuses on the development of UD-style treebanks for sophisticated syntactic modeling and the open-access release of all models and lexicons to the global research community. The ultimate goal is to provide a robust framework where spoken dialects, representing a more "natural" state of language evolution than prescribed standards, can be studied with computational rigor.

Summary updated

BibTeX
@inproceedings{vukovic2019corpora,
  author = {Teodora Vukovic and Nora Muheim and Olivier Winistörfer and Ivan Simko and Anastasia Makarova and Sanja Bradjan},
  title = {{Corpora and Processing Tools for Non-standard Contemporary and Diachronic Balkan Slavic}},
  booktitle = {RANLP Student Research Workshop},
  year = {2019},
  doi = {10.26615/issn.2603-2821.2019\_010},
}
Figure