The effect of dictionary omissions on phylogenies computationally inferred from lexical data

Lexical datasets used for computational phylogenetic inference suffer a unique type of data error. Some words actually present in a language may be absent from the dataset at no fault of its curators: especially for lesser-studied languages, a word may be missing from all available sources such as dictionaries. It is thus important to be able to (i) check how robust one’s inferences are to dictionary omission errors, and (ii) incorporate the knowledge that such errors may be present into one’s inference. I introduce two simple techniques that work towards those goals, and study the possible effects of dictionary omission errors in two real-life case studies on the Lezgian and Uralic datasets from Kassian (2015) and Syrjänen et al. (2013), respectively. The effects of dictionary omission turn out to be moderate (Lezgian) to negligible (Uralic), and certainly far less significant than the possible effects of modeling choices, including priors, on the inferred phylogeny, as demonstrated in the Uralic case study. Assessing the possible effects of dictionary omissions is advisable, but severe problems are unlikely. Collecting significantly larger lexical datasets, in order to overcome sensitivity to priors, is likely more important than expending resources on verifying data against dictionary omissions.