In plain words: A system deciphers lost languages by learning how letters changed between related languages, then pairing words at lowest total cost instead of guessing matches. It beat the best earlier Ugaritic results by 5.5% and translated 67.3% of Linear B words related to ancient Greek.
Abstract
In this paper we propose a novel neural approach for automatic decipherment of lost languages. To compensate for the lack of strong supervision signal, our model design is informed by patterns in language change documented in historical linguistics. The model utilizes an expressive sequence-to-sequence model to capture character-level correspondences between cognates. To effectively train the model in an unsupervised manner, we innovate the training procedure by formalizing it as a minimum-cost flow problem. When applied to the decipherment of Ugaritic, we achieve a 5.5% absolute improvement over state-of-the-art results. We also report the first automatic results in deciphering Linear B, a syllabic language related to ancient Greek, where our model correctly translates 67.3% of cognates.
Jiaming Luo, Yuan Cao, Regina Barzilay
arXiv:1906.06718 · cs.CL · submitted Jun 16, 2019
abstract · pdf · html · Accepted by ACL 2019
Given that there is a history of using classical NLP methods for sequence alignment / noisy channel decoding, I would've expected a more extensive discussion of how NNs might be able to overcome limitations of simpler methods.
But it seems the opposite is true--here they're using classical approaches to overcome limitations of their neural approach. The paper concludes by observing the "utmost importance of injecting prior linguistic knowledge" into the model. This "linguistic knowledge" is outlined in Section 3, and basically appears in the model as a regularization term based on a classical noisy channel / word alignment model. These regularization terms basically just encourage the neural network to behave like the classical models. And "neural" approach only performs marginally better than the (Berg-Kilpatrick & Klein 2011) paper they're comparing to, which takes a more classical combinatorial approach.