A Bayesian Model for Discovering Typological Implications

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 65–72, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics A Bayesian Model for Discovering Typological Implications Hal Daum´e III School of Computing University of Utah me@hal3.name Lyle Campbell Department of Linguistics University of Utah lcampbel@hum.utah.edu Abstract A standard form of analysis for linguis- tic typology is the universal implication. These implications state facts about the range of extant languages, such as “if ob- jects come after verbs, then adjectives come after nouns.” Such implications are typi- cally discovered by painstaking hand anal- ysis over a small sample of languages. We propose a computational model for assist- ing at this process. Our model is able to discover both well-known implications as well as some novel implications that deserve further study. Moreover, through a careful application of hierarchical analysis, we are able to cope with the well-known sampling problem: languages are not independent. 1 Introduction Linguistic typology aims to distinguish between log- ically possible languages and actually observed lan- guages. A fundamental building block for such an understanding is the universal implication (Green- berg, 1963). These are short statements that restrict the space of languages in a concrete way (for in- stance “object-verb ordering implies adjective-noun ordering”); Croft (2003), Hawkins (1983) and Song (2001) provide excellent introductions to linguistic typology. We present a statistical model for auto- matically discovering such implications from a large typological database (Haspelmath et al., 2005). Analyses of universal implications are typically performed by linguists, inspecting an array of 30- 100 languages and a few pairs of features. Looking at all pairs of features (typically several hundred) is virtually impossible by hand. Moreover, it is insuf- ficient to simply look at counts. For instance, results presented in the form “verb precedes object implies prepositions in 16/19 languages” are nonconclusive. While compelling, this is not enough evidence to de- cide if this is a statistically well-founded implica- tion. For one, maybe 99% of languages have prepo- sitions: then the fact that we’ve achieved a rate of 84% actually seems really bad. Moreover, if the 16 languages are highly related historically or areally (geographically), and the other 3 are not, then we may have only learned something about geography. In this work, we propose a statistical model that deals cleanly with these difficulties. By building a computational model, it is possible to apply it to a very large typological database and search over many thousands of pairs of features. Our model hinges on two novel components: a statistical noise model a hierarchical inference over language fam- ilies. To our knowledge, there is no prior work directly in this area. The closest work is repre- sented by the books Possible and Probable Lan- guages (Newmeyer, 2005) and Language Classifica- tion by Numbers (McMahon and McMahon, 2005), but the focus of these books is on automatically dis- covering phylogenetic trees for languages based on 2Indo-European cognate sets (Dyen et al., 1992). Data The database on which we …

Content truncated to 500 words