📄 AvarLab: An Integrated Digital Ecosystem for Avar, a Morphologically Rich Low-Resource Language

Summary

Avar is a Northeast Caucasian language of Dagestan that UNESCO classifies as vulnerable. It is nearly absent from NLP. Its heavy inflection, including ergative alignment, four agreement classes, irregular oblique stems and a large spatial case system, scatters its vocabulary across many rare surface forms. Existing resources are mostly isolated digitized dictionaries. The authors describe a circular problem: training tools requires annotated corpora, and building those corpora requires tools. To break it, they built AvarLab, a platform that links a trilingual lexicon, a rule-based morphological engine and corpus search. Its generate–verify loop works in four steps:

• Paradigms are derived from descriptive grammars.
• Each generated form is tagged automatically.
• The tags are checked against corpus occurrences.
• Frequently attested forms are locked, and unattested ones are passed to native speakers for review.

The current release yields over a million tagged forms from 14,768 entries. These support lookup by inflected form, corpus example retrieval and silver-standard training data. The work offers a reusable blueprint for morphologically rich, low-resource languages where corpus-first methods are not feasible.

Key points

• The system covers 14,768 lexical entries and produces over 1 million inflected forms across all parts of speech. It runs on Django and PostgreSQL and is available at avardict.upf.edu.
• A normalization step maps keyboard substitutes for the palochka (e.g., "1", "I", "|") to Unicode U+04C0 and treats digraphs as single units. A rule-based module converts Cyrillic to IPA.
• A phonotactic module, built from 29,197 dictionary entries, records 130 onset and 139 coda cluster types.
• Annotated data can be exported as JSON, CoNLL-U or spaCy-compatible datasets, with optional train/validation/test splits. A bot interface uses the same API.
• Verification against a 300,000-sentence corpus suggests that descriptive grammars over-generate, for example deeply nested spatial cases that never appear.
• Several components remain unevaluated: no gold standard exists for class or valency inference, the dependency parser is unfinished, and community features have not been tested in public deployment.