Reproduce “Large Language Model Should Understand Pinyin for Chinese ASR Error Correction” 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
Yuang Li, Xiaosong Qiao, Xiaofeng Zhao, Huan Zhao, Wei Tang, Min Zhang, Hao Yang
A Pinyin-enhanced Generative Error Correction (PY-GEC) framework for Chinese ASR integrates Large Language Models with phonetic input and a multitask training paradigm. The approach demonstrates an 8.3% relative reduction in Character Error Rate and a 3.9% relative improvement in Entity Recall compared to text-only correction baselines, effectively addressing homophone-induced errors.
Key insights
- Pinyin serves as a crucial supplementary input, enabling LLMs to more effectively resolve homophone-related ambiguities and improve Chinese ASR error correction.
- Multitask training with Pinyin-to-text and text-to-Pinyin conversion tasks explicitly aligns the internal feature spaces of text and Pinyin within LLMs, enhancing their understanding and utilization of phonetic information.
- Effective LLM-based Chinese ASR error correction can be achieved by training exclusively on synthetically generated homophonic errors, demonstrating efficiency in data acquisition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.