Скачать или смотреть Computational Historical Linguistics

How did Old Chinese diverge into the diverse Chinese (Sinitic) languages of today? The answer is that languages in the same language family began as dialects of the same language (the proto-language) and accumulated more and more sound changes over time, eventually becoming mutually unintelligible. Linguists have reconstructed words in these proto-languages in a painstakingly laborious process known as the comparative method, wherein linguists manually group related words (cognates) from daughter languages together and predict the phonemes in the proto-language. Instead, we can treat the problem as a seq2seq problem that takes in the pronunciations of words in modern related languages as input and outputs the pronunciations of the protoforms. Using RNNs, Transformers (work submitted to ACL 2023), and variational autoencoders, our lab has automatically reconstructed Middle Chinese and Latin protoforms, the former of which entailed the curation of a dataset of modern Chinese languages (Chang et al 2022). The talk will conclude with ongoing work on joint unsupervised phylogenetic inference and protoform reconstruction, the problem of learning both the language family tree and the protoforms at the same time, which more closely resembles the iterative process that historical linguists follow.

Slides: https://docs.google.com/presentation/...