24.05.2022 Gerhard Jäger

Mapping entries of word lists into discrete categories - such as cognate classes - results in the loss of valuable information. Recent advances in machine translation using deep learning make it possible to map sequential data into high-dimensional vector spaces with much less information loss. Applying such deep auto-encoders to multi-lingual word lists facilitates tasks such as cognate classification, phylogenetic inference, ancestral state reconstruction and missing value imputation when paired with continuous multivariate phylogenetic techniques.