Computational Methods and Long-Range Linguistic Comparison: Friends or Foes?

Abstract.  Could computers ever replace humans in the daunting task of reconstructing unattested ancestors of modern languages and building perfect phylogenies for present day linguistic families? In a recently published paper (J.-M. List et al., "The Potential of Automatic Word Comparison for Historical Linguistics", PLOS ONE, January 27, 2017), it was stated that modern automated methods of language comparison have been perfected to such a degree that they allow for identification of up to 80-90% etymological matches, previously established by trained experts "by hand" — seemingly implying that algorithms of automated comparison are robust enough to replicate and, if necessary, even replace work that was previously exclusively dependent on human potential. The truly important question in this context, however, is not whether computational methods can perform the same tasks that have already been successfully performed by human experts, but whether they can go beyond that level and help resolve issues that have hitherto proven way too challenging for said experts. Have computational methods, so far, succeeded in decisively clarifying any of these issues? In particular, have they proven useful for one of the most pressing and complicated tasks in historical linguistics — the establishment of "long distance genetic relationships" between large groupings of languages, such as, for instance, the Nostratic hypothesis of common linguistic ancestry between Indo-European, Uralic, Altaic, and several other families of Eurasia? In my talk, I will give a general overview of these issues, explain why the "big data approach" is difficult to apply to them, and point out why manual construction of various types of large linguistic databases is currently a far more important task than improving the algorithmic base for language data comparison.