aldo-git-bit/predicting-tashlhiyt-plural

Predicting Tashlhiyt Berber plural formation: phonological pipeline, ablation study, Bi-LSTM baselines, and lexical idiosyncrasy analysis # Predicting Tashlhiyt Plural Formation A computational study of Tashlhiyt Berber (Tachelhit) nominal plural formation, combining rule-based phonological analysis with machine learning to quantify the predictability of plural patterns and identify lexically idiosyncratic forms. --- ## Overview Tashlhiyt Berber nouns form plurals through three strategies: **external** (suffixation only), **internal** (stem mutation only), or **mixed** (both). This project investigates how much of plural formation is predictable from surface phonological form versus stored lexically, using a dataset of 1,914 nouns with full inflectional paradigms. The central finding is that hand-crafted morphophonological features (syllable structure, foot type, morphological class) consistently outperform n-gram baselines and character-level neural models across 10 classification tasks, with Macro-F1 ranging from 0.59 (Final A insertion) to 0.91 (templatic mutation). Error overlap analysis reveals that only 19–25% of errors are made by both model types simultaneously, suggesting most failures reflect genuine lexical idiosyncrasy rather than underfitting. --- ## Repository Structure ``` predicting-tashlhiyt-plural/ │ ├── data/ │ ├── tash_nouns.csv # Main dataset (1,914 nouns, 53 columns) │ ├── tash_nouns_readme.txt # Full column descriptions and documentation │ ├── import_golden_syllables.csv # 72-form gold standard for syllabification │ ├── golden_syllables_expanded.csv # Expanded gold standard (140 forms) │ ├── forms_from_plural_theme.csv # Plural forms derived from themes │ ├── record_extractor_template.txt # Template for extracting records by pattern │ ├── ngram_features_macro.csv # N-gram feature matrix, macro level (n=1,185) │ ├── ngram_features_micro.csv # N-gram feature matrix, micro level (n=562) │ ├── ngram_metadata_macro.json # N-gram feature selection metadata (macro) │ ├── ngram_metadata_micro.json …