Reproduce “Neural Proto-Language Reconstruction” 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
Chenxuan Cui, Ying Chen, Qinxin Wang, David R. Mortensen
Researchers from Carnegie Mellon University advanced neural proto-language reconstruction by applying data augmentation and a VAETransformer model to the Sinitic WikiHan dataset. Their methods led to improved reconstruction performance, with data augmentation reducing Edit Distance from 0.900 to 0.870 and the VAETransformer model achieving an Edit Distance of 0.8829, outperforming the standard Transformer baseline.
Key insights
- Accurate data augmentation, especially via character-level transducers on datasets like WikiHan, effectively stabilizes neural model training and improves proto-form reconstruction.
- Explicitly integrating linguistic principles, such as the Neogrammarian hypothesis through a VAE structure, guides neural models to learn more structured and informative latent representations.
- Framing proto-language reconstruction as a variational machine translation problem allows models to learn shared semantic spaces, potentially enhancing the plausibility and accuracy of reconstructed forms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.