Continuous Space Representations of Linguistic Typology and their Application to Phylogenetic Inference

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 324–334, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Continuous Space Representations of Linguistic Typology and their Application to Phylogenetic Inference Yugo Murawaki Graduate School of Information Science and Electrical Engineering Kyushu University Fukuoka, Japan murawaki@ait.kyushu-u.ac.jp Abstract For phylogenetic inference, linguistic typol- ogy is a promising alternative to lexical evi- dence because it allows us to compare an ar- bitrary pair of languages. A challenging prob- lem with typology-based phylogenetic infer- ence is that the changes of typological fea- tures over time are less intuitive than those of lexical features. In this paper, we work on reconstructing typologically natural ances- tors To do this, we leverage dependencies among typological features. We first repre- sent each language by continuous latent com- ponents that capture feature dependencies. We then combine them with a typology evaluator that distinguishes typologically natural lan- guages from other possible combinations of features. We perform phylogenetic inference in the continuous space and use the evalua- tor to ensure the typological naturalness of in- ferred ancestors. We show that the proposed method reconstructs known language fami- lies more accurately than baseline methods. Lastly, assuming the monogenesis hypothesis, we attempt to reconstruct a common ancestor of the world’s languages. 1 Introduction Linguistic typology is a cross-linguistic study that classifies the world’s languages according to struc- tural properties such as complexity of syllable struc- ture and object-verb ordering. The availability of a large typology database (Haspelmath et al., 2005) makes it possible to take computational approaches to this area of study (Daum´e III and Campbell, 2007; Georgi et al., 2010; Rama and Kolachina, 2012). In this paper, we consider its application to phyloge- netic inference. We aim at reconstructing evolution- ary trees that illustrate how modern languages have descended from common ancestors. Typological features have two advantages over other linguistic traits. First, they allow us to com- pare an arbitrary pair of languages. By contrast, historical linguistics has worked on regular sound changes (see (Bouchard-Cˆot´e et al., 2013) for com- putational models). Glottochronology and computa- tional phylogenetics make use of the presence and absence of lexical items (Swadesh, 1952; Gray and Atkinson, 2003). All these approaches require that certain sets of cognates, or words with common et- ymological origins, are shared by the languages in question. For this reason, it is hardly possible to use lexical evidence to search for external relations in- volving language isolates and tiny language families such as Ainu, Basque, and Japanese. For these lan- guages, typology can be seen as the last hope. The second advantage is that typological features are potentially capable of tracing evolutionary his- tory on the order of 10,000 years because they change far more slowly than lexical traits. A glot- tochronological study indicates that even if Japanese is genetically related to Korean, they diverged from a common ancestor no earlier than 6,700 years ago (Hattori, 1999). Even the basic vocabulary …

Content truncated to 500 words