Comparing Language Similarity across Genetic and Typologically-Based Groupings

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 385–393, Beijing, August 2010 Comparing Language Similarity across Genetic and Typologically-Based Groupings Ryan Georgi University of Washington rgeorgi@uw.edu Fei Xia University of Washington fxia@uw.edu William Lewis Microsoft Research wilewis@microsoft.com Abstract Recent studies have shown the poten- tial benefits of leveraging resources for resource-rich languages to build tools for similar, but resource-poor languages. We examine what constitutes “similarity” by comparing traditional phylogenetic lan- guage groups, which are motivated largely by genetic relationships, with language groupings formed by clustering methods using typological features only. Using data from the World Atlas of Language Structures (WALS), our preliminary ex- periments show that typologically-based clusters look quite different from genetic groups, but perform as good or better when used to predict feature values of member languages. 1 Introduction While there are more than six thousand languages in the world, only a small portion of these lan- guages have received substantial attention in the field of NLP. With the increase in use of data- driven methods, languages with few or no elec- tronic resources have been difficult to process with current methods. The morphological tagging of Russian using Czech resources as done by (Hana et al., 2004) shows the potential benefit for using the resources of resource-rich languages to boot- strap NLP tools for related languages. Projecting syntactic structures across languages (Yarowsky and Ngai, 2001; Xia and Lewis, 2007) is another possible way to harness existing tools, though such projection is more reliable among languages with similar syntax. Studies such as these show the possible bene- fits of working with similar languages. A crucial question is how we should define similarity be- tween languages. While genetically related lan- guages tend to have similar typological features as they could inherit the features from their com- mon ancestor, they could also differ a lot due to language change over time. On the other hand, languages with no common ancestor could share many features due to language contact and other factors. It is worth noting that the goals of historical lin- guistics differ from those of language typology in that while historical linguistics focuses primarily on diachronic language change, typology is more focused on a synchronic survey of features found in the world’s languages: what typological fea- tures exist, where they are found, and why a lan- guage has a feature. These differences between the concepts of ge- netic relatedness and language similarities lead us to the following questions: Q1. If we cluster languages based only on their typological features, how do the induced clusters compare to phylogenetic groupings? Q2. How well do induced clusters and genetic families perform in predicting values for ty- pological features? Q3. What typological features tend to stay the same within language families, and what fea- tures are likely to differ? These questions are the focus of this study, and for the experiments, we use information from World Atlas of Language Structures (Haspelmath et al., 2005), or WALS. 385 ID# Feature Name …

Content truncated to 500 words