kariminf/chdzdt

Character-based Algerian Dialect Transformer Character-based Algerian Dialect Transformer (chDzDT) is a lightweight transformer model that operates at the character level. The model can be described using the symbolic notation $c\cdot h \cdot \frac{\partial z}{\partial t}$, which metaphorically captures its architecture: character-level processing ($c$), deep representations ($h$), and transformation over time ($\frac{\partial z}{\partial t}$). It is specifically designed to address the challenges posed by morphologically rich, noisy, multilingual, and multi-script dialects such as Algerian Arabic. To this end, the model is built to be robust against orthographic variation and morphological obfuscation, which are common in informal text. chDzDT is intended for word-level encoding based purely on character-level morphological features. This makes it well-suited to tasks involving word morphology, such as inflection or derivation. In addition, it can serve as a preprocessing module to enhance traditional language models; particularly in handling out-of-vocabulary (OOV) tokens. **This project is a part of a larger project called DzDT started on 2023. It was separated on August 2025 to handle only the morphology part of the project.** # Training a chDzDT model - To specify the configuration of the model, please see this example config/chdzdt_train/chdzdt_train.json - To train a new model ```sh >> exec/train/chdzdt.trn.py new ``` - To resume training a model ```sh >> exec/train/chdzdt.trn.py resume ``` # More information Please refer to this Article: arXiv:2509.01772 You can cite it as: ``` @misc{2025-aries-chdzdt, title={chDzDT: Word-level morphology-aware language model for Algerian social media text}, author={Abdelkrime Aries}, year={2025}, eprint={2509.01772}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={arxiv.org, } ``` # License Copyright (C) 2025 Abdelkrime Aries Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compli …