TL;DR
This paper introduces the first parallel corpus for the Ruska Romani dialect and Russian, created from 1920s-30s literature. This resource, available in JSON and RNC XML formats, is crucial for linguistic research, NLP tool development, and the preservation of the Ruska Romani language and culture.
What Problem Does This Paper Address?
The field of Natural Language Processing (NLP) and computational linguistics has seen tremendous growth, largely fueled by the availability of large, parallel text corpora. These corpora are essential for training machine translation systems, developing language models, and conducting in-depth linguistic analysis. However, many minority languages, including various Romani dialects, remain severely under-resourced, lacking such foundational datasets. This scarcity hinders research and the development of NLP tools that could benefit these communities. Specifically, the Ruska Romani dialect, spoken by a significant portion of Romani people in Russia, lacked a readily available, machine-readable parallel corpus with Russian, despite existing literature translations. This paper aims to bridge this gap by creating and releasing such a corpus.
Methods
The authors employed a multi-stage approach to construct the parallel corpus:
- Data Acquisition: They sourced Russian literature translated into Ruska Romani from the 1920s and 1930s, a period of active minority language policy in the Soviet Union.
- Sentence Alignment:
- A subset of translations was manually aligned with their original Russian counterparts.
- This aligned subset was used to fine-tune a multilingual sentence embedding model (LaBSE).
- The fine-tuned model, integrated with the
lingtrain-alignerlibrary, was then used to automatically align the remaining sentence pairs. - A chain score metric was used to evaluate and refine the alignment process, automatically resolving conflicts by splitting or merging sentences.
- Cosine similarity was used as a secondary check, with sentence pairs below a threshold of 0.5 being manually reviewed and corrected by a linguist.
- Morphological Annotation:
- The
uniparser-soviet-romanilibrary, a morphological analyzer for Ruska Romani, was used.
- The
- This library provides lemmas, part-of-speech tags, and other grammatical features, though it doesn't perform disambiguation.
- For tokens not covered by the library (e.g., loanwords, proper nouns), a custom multi-step approach was developed. This involved checking for Ruska Romani suffixes, removing them, and then leveraging Russian morphological analysis (PyMorphy2) to infer properties.
- This rule-based approach increased the annotation coverage by 4%.
- Format Conversion: The annotated data was converted into two formats: JSON and the Russian National Corpus (RNC) XML format.
Datasets
- Source Texts: Russian literature translated into Ruska Romani from the 1920s and 1930s. Specific titles include "Dubrovskij," "Posle bala," "Tri medvedya," and others listed in Table 1.
- Training Data for Alignment Model: Sentence pairs from randomly selected titles: "Dubrovskij," "Malen'kie rasskazy," "Tri medvedya," "Posle bala." This subset contained 24,700 Russian tokens and 25,170 Ruska Romani tokens.
- The Parallel Corpus: The final corpus contains 88,742 Russian tokens and 84,635 Ruska Romani tokens, with 74,291 of the latter being grammatically annotated. This corpus is a newly introduced benchmark for Ruska Romani NLP research.
Key Insights and Ideas
- Leveraging Historical Texts: The paper demonstrates the value of historical translated literature as a rich source for creating parallel corpora for under-resourced languages, especially when direct contemporary data is scarce.
- Hybrid Alignment Approach: Combining a fine-tuned multilingual embedding model with heuristics and manual review proved effective for achieving high-quality sentence alignment, even for a less-resourced language pair.
- Custom Annotation Rules for Under-Resourced Languages: The development of specific rules to handle loanwords and proper nouns, by integrating analysis from a related, better-resourced language (Russian), is a practical strategy for improving morphological annotation coverage.
- Importance of Representation: Making this corpus publicly available, particularly within the RNC platform, significantly contributes to the visibility and preservation of the Ruska Romani dialect and its associated culture.
Key Results
- The developed parallel corpus contains 88,742 Russian tokens and 84,635 Ruska Romani tokens.
- 74,291 Ruska Romani tokens were grammatically annotated, achieving 88% coverage after applying custom rules (up from 84% using the library alone).
- The sentence alignment model achieved a best chain score of 0.74 during training.
- Manual review identified and corrected 881 sentence pairs (11% of those assessed) for errors.
Limitations and Future Work
- Disambiguation: The morphological annotation does not perform disambiguation, meaning multiple possible analyses are provided for each token.
- Russian Annotation: The paper relies on RNC annotation tools for Russian sentences, which are not publicly available.
- Text Type: The corpus exclusively contains translations of Russian literature and does not include spoken language data.
- Language State: The texts are from the 1920s-30s and may not fully reflect the current state of the Ruska Romani dialect.
Future work could involve expanding the corpus with more texts, including spoken language, and developing disambiguation capabilities for the morphological annotations. The authors also suggest the corpus can be used for stylometry studies, bilingual dictionary creation, and the development of NLP/MT tools for Ruska Romani.
Summary updated