Improved Reconstruction of Protolanguage Word Forms

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 65–73, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Improved Reconstruction of Protolanguage Word Forms Alexandre Bouchard-Cˆot´e∗ Thomas L. Griffiths† Dan Klein∗ ∗Computer Science Division †Department of Psychology University of California at Berkeley Berkeley, CA 94720 Abstract We present an unsupervised approach to re- constructing ancient word forms. The present work addresses three limitations of previous work. First, previous work focused on faith- fulness features, which model changes be- tween successive languages. We add marked- ness features, which model well-formedness within each language. Second, we introduce universal features, which support generaliza- tions across languages. Finally, we increase the number of languages to which these meth- ods can be applied by an order of magni- tude 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. 1 Introduction A central problem in diachronic linguistics is the re- construction of ancient languages from their modern descendants (Campbell, 1998). Here, we consider the problem of reconstructing phonological forms, given a known linguistic phylogeny and known cog- nate groups. For example, Figure 1 (a) shows a col- lection of word forms in several Oceanic languages, all meaning to cry. The ancestral form in this case has been presumed to be /taNis/ in Blust (1993). We are interested in models which take as input many such word tuples, each representing a cognate group, along with a language tree, and induce word forms for hidden ancestral languages. The traditional approach to this problem has been the comparative method, in which reconstructions are done manually using assumptions about the rel- ative probability of different kinds of sound change (Hock, 1986). There has been work attempting to automate part (Durham and Rogers, 1969; Eastlack, 1977; Lowe and Mazaudon, 1994; Covington, 1998; Kondrak, 2002) or all of the process (Oakes, 2000; Bouchard-Cˆot´e et al., 2008). However, previous au- tomated methods have been unable to leverage three important ideas a linguist would employ. We ad- dress these omissions here, resulting in a more pow- erful method for automatically reconstructing an- cient protolanguages. First, linguists triangulate reconstructions from many languages, while past work has been lim- ited to small numbers of languages. For example, Oakes (2000) used four languages to reconstruct Proto-Malayo-Javanic (PMJ) and Bouchard-Cˆot´e et al. (2008) used two languages to reconstruct Clas- sical Latin (La). We revisit these small datasets and show that our method significantly outperforms these previous systems. However, we also show that our method can be applied to a much larger data set (Greenhill et al., 2008), reconstructing Proto- Oceanic (POc) from 64 modern languages. In ad- dition, performance improves with more languages, which was not the case for previous methods. Second, linguists exploit knowledge of phonolog- ical universals. For example, small changes in vowel height or consonant place are more likely than large changes, and much more likely than change …

Content truncated to 500 words