NoorBayan/Rhetorix: Rhetorix v1.1.0: Distance-Aware Ordinal Learning for Classical Arabic Rhetoric
Abstract
Rhetorix: Distance-Aware Ordinal Learning for Classical Arabic Rhetoric Official open-source release of the Rhetorix computational framework, accompanying the research paper: "Distance-Aware Ordinal Learning for Processing-Effort Classification in Classical Arabic Rhetorical Texts" 📌 Overview Rhetorix is an end-to-end computational framework designed for predicting graded processing effort in Classical Arabic figurative discourse. It addresses the challenge of modeling ordered linguistic constructs under severe class imbalance by introducing a task-specific distance-aware optimization objective directly into transformer fine-tuning. 🚀 Key Features & Components Distance-Aware Ordinal Loss ($\mathcal{L}_{Ordinal}$): A unified PyTorch objective combining weighted Categorical Cross-Entropy ($L_{CE}$) with an expected metric distance penalty ($L_{Dist}$): $$\mathcal{L}{Ordinal} = L{CE} + \lambda L_{Dist}$$ Built-in cost-sensitive inverse class weighting to counteract severe empirical skew without altering input representations. Pretrained Transformer Pipelines: Ready-to-use fine-tuning and evaluation modules across five architectures: ARBERT (Large-scale Modern Standard Arabic) MARBERTv2 (Web-scale Dialectal Arabic) CAMeLBERT-CA (Classical Arabic / Heritage texts) AraBERTv2 (Standard Formal Arabic baseline) XLM-RoBERTa (Multilingual cross-lingual baseline) Robust & Leak-Free Validation Suite: Nested Stratified 5-Fold Cross-Validation with source-level grouping. Internal-validation checkpoint selection optimizing Quadratic Weighted Kappa (QWK). Strict partition isolation: weights and hyperparameters are computed solely from inner training data. Comprehensive Evaluation Metrics: Automated calculation of chance-corrected agreement (QWK), error magnitude (MAE), class-balanced recovery (Macro-F1), and nominal Accuracy. Error severity profiling tools distinguishing adjacent misclassifications (Severity-1) from extreme endpoint deviations (Severity-2). Non-parametric Wilcoxon signed-rank diagnostic test for paired instance-level error analysis. 📂 Dataset: RhetoriScale ($N = 965$) An expert-annotated corpus of Classical Arabic figurative expressions categorized into three ordered inferential levels: Low (0): 45 instances (4.66%) Medium (1): 764 instances (79.17%) High (2): 156 instances (16.17%) Fully pre-partitioned into leak-free cross-validation folds. Software Stack: Python 3.10+ | PyTorch 2.x | Hugging Face Transformers
