Abstract

We present an unsupervised approach to reconstructing ancient word forms. The present work addresses three limitations of previous work. First, previous work focused on faithfulness features, which model changes between successive languages. We add markedness features, which model well-formedness within each language. Second, we introduce universal features, which support generalizations across languages. Finally, we increase the number of languages to which these methods can be applied by an order of magnitude by using improved inference methods. Experiments on the reconstruction of Proto-Oceanic, Proto-Malayo-Javanic, and Classical Latin show substantial reductions in error rate, giving the best results to date.