News: ACL 2024 Presentation
#post
Thrilled to share that I will be presenting our paper, “Semisupervised Neural Proto-Language Reconstruction” at ACL 2024 in Bangkok 🇹🇠next month!
We introduce the new task of semisupervised protoform reconstruction, in which the model learns to reconstruct ancestral words (protoforms) from their descendent words (reflexes, or related words in daughter languages) with only a small amount of labelled data (labels being gold reconstructions) and a large amount of unlabeled data.
We propose the new DPD (Daughter-to-Proto-to-Daughter) architecture informed by historical linguists’ comparative method, namely its underlying principle that historical sound changes are largely regular and reflexes should be derivable from protoforms.
The DPD architecture is designed to predict reflexes from its own predicted reconstructions in an end-to-end manner, so gradients flow from the reflex prediction sub-model into the reconstruction sub-model!
We find that the DPD architecture can leverage unlabeled training data to outperform baseline methods in almost all situations.
Joint work with Peirong Xie and David Mortensen
Looking forward to presenting at ACL 2024
To find out more:
#ACL2024