README.md

metadata

title: Kabyle POS Tagger
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.15.2
app_file: app.py
pinned: false
license: apache-2.0

Kabyle POS Tagger Demo

Interactive demo for the boffire/kabyle-pos-v2 model — a Part-of-Speech tagger for Kabyle (kab), a Berber language spoken in Algeria.

Model Details

• Base: XLM-RoBERTa-base
• Task: Token Classification (POS tagging)
• Test F1: 93.8%
• Training Data: 10,000 sentences with a 214,000-entry lexicon
• Tagset: Universal Dependencies (17 tags)

Features

• Punctuation-aware tokenization: Attached punctuation (e.g., medden.) is automatically split and tagged as PUNCT.
• Clitic handling: Hyphenated possessive, accusative, dative, and directional clitics are split and tagged correctly.
• Post-processing lookup table: A linguistically curated override table fixes misclassifications for closed-class morphemes (e.g., -nneɣ, -is, d-, -agi).
• High-contrast visualization: Color-coded tokens with confidence scores.

Supported Clitics

The app recognizes and correctly tags the following Kabyle grammatical morphemes:

Possessive Affixes

• Singular: -w/-iw, -k/-ik, -m/-im, -s/-is
• Plural: -nneɣ, -wen/-nwen, -kent/-nkent, -sen/-nsen, -sent/-nsent

Direct Object Pronouns (Accusative)

• -iyi/-yi, -k/-ik, -kem, -t/-tt, -itt, -aɣ/-yaɣ, -ken, -kent, -ten, -tent

Indirect Object Pronouns (Dative)

• -iyi/-yi, -ak, -am, -as/-asen, -aneɣ/-anaɣ, -awen, -akent, -asen/-atsen, -asent/-atsent

Directional & Copula Particles

• d-/-d/-id — Proximal particle (toward speaker / "it is")
• n-/-in — Distal particle (away from speaker)

Demonstratives & Determiners

• -agi/-a — This / These
• -nni — That / Those (previously mentioned)
• -nniḍen/-niḍen — Other / Another

Usage

Type or paste a Kabyle sentence and click Submit to see predicted POS tags with confidence scores.

Example Sentences

• Aṭas n medden i yessen.
• Taqbaylit d tutlayt deg Lezzayer.
• Yella wuccen ameqqran deg taddart.
• Tameddakelt-nneɣ teɣra adlis-is.
• D nekkni i d-yusan d imezwura.

Limitations

• Capitalized sentence-initial words may be biased toward NOUN/PROPN due to training data distribution.
• Domain bias toward short translated sentences (Tatoeba corpus).
• No diacritic normalization.

Citation

Side part of the Masakhane initiative for African NLP. See the model card for full citation details.

Acknowledgments

• Model trained by boffire (ButterflyOfFire)