Reproduce “A Chat About Boring Problems: Studying GPT-based text normalization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Zhang, Travis M. Bartley, Mariana Graterol-Fuenmayor, Vitaly Lavrukhin, Evelina Bakhturina, Boris Ginsburg
Researchers at Nvidia Corporation empirically demonstrate that Large Language Models can reliably perform text normalization, achieving superior accuracy and drastically reducing unrecoverable errors compared to traditional rule-based systems. GPT-4.0 notably outperformed Google's Kestrel WFST system, making 40% fewer total errors and exhibiting only nine unrecoverable errors on a 7551-sentence test set when evaluated with a novel linguistic error taxonomy.
Key insights
- Few-shot learning effectively enables LLMs for text normalization, though prompt length can influence performance due to context management challenges.
- Integrating self-consistency reasoning consistently enhances the robustness and accuracy of LLM-based text normalization.
- A refined, linguistically-informed error taxonomy is essential for accurately evaluating text normalization systems, demonstrating that most LLM "errors" are felicitous variations rather than unrecoverable failures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.