A Learned Picker for Convention-Aware Hybrid Symbolic-Neural Lemmatization in Agglutinative Languages: A Case Study on Turkish (Artefact Bundle)

Artefact bundle accompanying the paper "A Learned Picker for Convention-Aware Hybrid Symbolic-Neural Lemmatization in Agglutinative Languages: A Case Study on Turkish" (Yuce & Yuce, submitted to Computer Speech & Language, 2026). Contents: - Turkish (surface form, root) dataset (9,154 unique pairs, SHA-256 c4b3f64935ad0ddb3846ec4b80470ec9a8e10541eea67a5b320a446b67488d57) - paper_splits.json (canonical 7,322/916/916 train/val/test split, random_state=42) - Trained model weights: best_rootfinder_beta98_m0.weights.h5, best_rootfinder_swa20_m0.weights.h5 (Keras .weights.h5 format) - Per-example test predictions for every system reported in the paper (ByT5, TURNA, Zemberek, Stanza, character-Transformer, N-way RF picker) - Bootstrap and McNemar artefacts (JSON) - Source code tarball (tr_root_code_cf70f0e.tar.gz) snapshot at commit cf70f0e - Cross-lingual probing data and ablation results