Diacritics restoration: can we do better with neural networks and deep learning? (ileriseviye.wordpress.com)

A compact PyTorch project called nokta-ai (2025) was announced that restores Turkish diacritics (ç, ğ, ı, İ, ö, ş, ü) from ASCII text: a small model trained on an Apple M1 Pro achieved >85% accuracy, while a larger model trained on an NVIDIA A100 for under 24 hours reportedly hit >99%. This modern “smart brute force” result sits alongside earlier solutions: Deniz Yüret’s classical Emacs-pattern deasciifier (~96% accuracy) and Ayşenur Genç Uzun’s Dynet-based RNN seq2seq system (trained on 630K sentences for 3 epochs, ~86%). There are also informal reports that ChatGPT/LLMs perform very well on deasciification, prompting calls for systematic benchmarking against validated, representative Turkish corpora. Technically, the story highlights two trajectories: lightweight on-device models good enough for many users vs. larger GPU-trained networks that can push near-perfect accuracy; and the potential of modern frameworks (PyTorch/TensorFlow) to make iterative retraining easy. Key bottlenecks remain: representative multilingual/abbreviated corpora, compute for longer training, and standardized benchmarks to compare rule-based, RNN/seq2seq, transformer, and LLM approaches. If reproducible, >98–99% automated deasciification would be practical for millions of users, but cost, latency, and domain coverage (foreign terms, abbreviations) will determine whether on-device or cloud/LLM solutions dominate.