Ossetic-COT: Designing a morphologically annotated corpus and morphological analyzer for Ossetic

Technical Summary: Ossetic-COT

Problem Statement
Ossetic, an Iranic language spoken primarily in the North Caucasus, lacks a morphologically disambiguated corpus adhering to the Universal Dependencies (UD) schema. While two existing corpora exist—the Ossetic National Corpus (ONC) and the Corpus of Oral Texts (COT)—neither meets the requirements for high-quality morphological analysis. The ONC relies on context-free automatic annotation (Uniparser-Morph), resulting in ambiguous tags with multiple possibilities per word. The COT, consisting of oral texts, contains manual interlinear glossing but only offers incomplete, coarse-grained annotation that is difficult to convert to standard UD formats. This absence of a standardized, disambiguated resource hinders the development of NLP tools and limits the corpus's utility for educational and linguistic research.

Methodology
The authors address this gap by constructing a morphologically annotated corpus for the Iron dialect of Ossetic and training a BERT-based morphological analyzer.

1. Corpus Construction and Annotation:

   • Source: The study utilizes the COT (60,000 tokens), filtering out Russian content and retaining 5454 sentences (73,042 tokens).
   • Conversion Workflow: The process involved five stages: filtering non-target language content; manual sentence segmentation and tokenization; mapping original tags to the UD v2 inventory (enriching underspecified pronominal entries and adding missing feature values); manual refinement by experts in Ossetic morphology; and automated conversion to CoNLL-U format.
   • Annotation Guidelines: To accommodate Ossetic's specific linguistic idiosyncrasies, the authors extended standard UD guidelines:
     • POS Disambiguation: Implemented context-dependent rules to distinguish between NOUN, ADJ, ADV, and ADP for substantives that lack clear morphological boundaries (e.g., distinguishing adjectives from adverbs based on dependent structures).
     • Auxiliaries: Introduced a distinct AUX tag for light verbs in complex predicates, copulas, and analytic constructions, distinguishing them from lexical verbs.
     • Pronouns: Adopted a context-based approach where determiners functioning as verbal arguments are tagged as PRON.
     • Case System: Added Dir (directive) and Rcs (recessive) cases. Crucially, the authors introduced Num (numerative) as a value for the Case feature to distinguish it from the genitive, particularly for inanimate counted objects, applying Differential Object Marking (DOM) principles.
     • Non-finite Forms: Extended the VerbForm feature to include Inf2, Part2, Dest, and Red (Reduplicated) to cover Ossetic's extensive repertoire of non-finite forms.
     • Prefixes: Introduced Preverb=Yes and Prenoun=Yes features to handle aspectual prefixes and nominal prefixes (ӕнӕ, ӕд) respectively.

2. Model Training:

   • Pre-training: Two BERT-based models were pre-trained on the unlabeled ONC (approx. 1 million sentences): a standard multilingual BERT (mBERT) and an "Ossetic BERT" with a custom embedding layer and WordPiece tokenizer trained on Ossetic data.
   • Fine-tuning: Both models were fine-tuned on the newly created annotated COT dataset for the task of morphological labeling. The models were evaluated using a train/validation/test split (8:1:1).

Key Contributions

• Ossetic-COT Corpus: The creation of the first morphologically annotated corpus for Iron Ossetic conforming to the UD v2 schema, containing 5454 manually disambiguated sentences.
• Annotation Framework: A set of language-specific annotation guidelines that resolve ambiguities inherent to Ossetic morphology (e.g., the fluid boundary between parts of speech, the numerative case, and non-finite verbal forms).
• Baseline Morphological Analyzer: A trained BERT-based morphological analyzer that serves as a benchmark for the language.

Results

• Performance: The multilingual BERT (mBERT) based analyzer achieved a word-level tag accuracy of 95.60% and a sentence accuracy of 61.54%. The custom "Ossetic BERT" performed slightly lower (95.47% accuracy), leading the authors to select the mBERT model as the final baseline.
• Error Analysis: The model performed well on most part-of-speech categories but struggled with specific homonymies, particularly between Noun/Adp, GEN/NUM, and various person/number combinations. The authors attribute these errors to data scarcity, lexical limitations of the oral corpus, and inherent ambiguities where context is insufficient for definitive disambiguation.
• Feature Extraction: The model successfully resolved most morphological features, though errors persisted in identifying specific non-finite verbal forms and distinguishing auxiliary from lexical verb uses.

Significance and Claims
The paper claims that the Ossetic-COT corpus and the accompanying analyzer represent a foundational step for Ossetic NLP. The corpus is presented as a resource suitable for training neural morphological taggers, conducting linguistic research, and educational applications, as its annotation captures the majority of the language's morphological features. The authors modestly position the provided model as an "initial neural morphological tagger" and a benchmark for evaluating dataset complexity. They explicitly note that the model's performance may degrade on out-of-domain data or complex literary registers due to the oral nature and limited lexical diversity of the training data (COT), identifying the integration of literary data from the ONC as the primary direction for future work.