Reproduce “SwitchLingua: The First Large-Scale Multilingual and Multi-Ethnic Code-Switching Dataset” 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
Peng Xie, Xingyuan Liu, Tsz Wai Chan, Yequan Bie, Yangqiu Song, Yang Wang, Hao Chen, Kani Chen
Researchers from The Hong Kong University of Science and Technology introduced SwitchLingua, a large-scale multilingual and multi-ethnic code-switching dataset, a novel data synthesis framework called LinguaMaster, and the Semantic-Aware Error Rate (SAER) evaluation metric. SwitchLingua was validated to possess high diversity and authenticity compared to existing resources, serving as a comprehensive benchmark for code-switching ASR system evaluation.
Key insights
- A multi-agent LLM-based framework incorporating explicit linguistic principles, multi-agent collaboration, and external tool integration can effectively synthesize high-quality and authentic code-switching data.
- Rigorous synthetic data generation, as demonstrated by SwitchLingua, can surpass existing human-collected or machine-translated datasets in terms of linguistic richness, diversity, realism, and naturalness.
- Incorporating semantic understanding into ASR evaluation metrics provides a more nuanced and accurate performance assessment for code-switching, addressing the limitations of purely word or character-level error rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.