16. Dez 2025
Frederik Hartmann
University of North Texas, Department of Linguistics
Challenges of implementing the comparative method with neural networks
Abstract: Reconstructing proto-languages or ancestral states of living languages is a fundamental task in diachronic linguistics from which much of our knowledge of language change is sourced. Traditionally, proto-languages have been reconstructed using the comparative method, which involves the systematic comparison of phonological patterns to back-trace sound change trajectories. Despite its considerable success in historical linguistics, its current implementation has several limitations, many of which stem from the lack of cross-linguistic knowledge informing the reconstructions. Here, artificial neural networks can complement the existing methods by providing the capability for learning comparative reconstruction across language families and having cross-linguistic knowledge feed into linguistic reconstructions. Using initial experiments on simulated and real-world data as a foundation, I pinpoint the challenges ahead and outline a roadmap for neural reconstruction modeling by sketching a possible pathway for future research.