dayaeddin/Stochastic-CNN-SDE-model-for-Robust-Character-level-Arabic-POS-Tagging

Stochastic CNN–SDE Model for Robust Character-level Arabic POS Tagging to appear in ICASA’2026 – The First International Conference on Statistics and Its Applications, Djillali Liabes University, Algeria, April 08–09, 2026. Conference website: univ-sba.dz # Stochastic CNN–SDE Model for Robust Character-level Arabic POS Tagging This repository contains the implementation and experiments for the paper: **Stochastic CNN–SDE Model for Robust Character-level Arabic POS Tagging** to appear in **ICASA’2026 – The First International Conference on Statistics and Its Applications**, Djillali Liabes University, Algeria, April 08–09, 2026. Conference website: univ-sba.dz ## Overview We propose a character-level method for Arabic part-of-speech (POS) tagging that combines a lightweight Transformer encoder with a continuous-time stochastic differential equation (CNN–SDE) layer. The model operates on characters via an embedding layer, applies a Transformer encoder to capture contextual dependencies, and then uses a CNN–SDE layer whose drift and diffusion terms are parametrized by neural networks trained end-to-end. Using a large annotated Arabic corpus, the Transformer–CNN–SDE model achieves high accuracy and strong macro-F1, outperforming or matching strong baselines such as BERT, BiLSTM, and CRF under comparable experimental conditions. We provide a detailed evaluation including robustness across multiple random seeds, per-class metrics, and visual analyses of the learned stochastic dynamics, showing the benefits of combining character-level representations with stochastic modeling to handle the morphological richness and ambiguity of Arabic. ## Test Set POS Tag Performance | Label | Precision | Recall | F1-score | |---------------------|-----------|--------|----------| | PAD | 0.00 | 0.00 | 0.00 | | adjective | 0.47 | 0.44 | 0.46 | | adposition | 0.57 | 0.58 | 0.57 | | adverb | 0.71 | 0.67 | 0.69 | | auxiliaryverb | 0.67 | 0.64 | 0.65 | | coordinatingconj | 0.70 | 0.65 | 0.68 | | determiner | 0.66 | 0.68 | 0.67 | | foreignw | 0.57 | 0.55 …