Diachrony-aware Induction of Binary Latent Representations from Typological Features

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 451–461, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Diachrony-aware Induction of Binary Latent Representations from Typological Features Yugo Murawaki Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan murawaki@i.kyoto-u.ac.jp Abstract Although features of linguistic typology are a promising alternative to lexical ev- idence for tracing evolutionary history of languages, a large number of missing val- ues in the dataset pose serious difficul- ties for statistical modeling. In this pa- per, we combine two existing approaches to the problem: (1) the synchronic ap- proach that focuses on interdependencies between features and (2) the diachronic approach that exploits phylogenetically- and/or spatially-related languages. Specif- ically, we propose a Bayesian model that (1) represents each language as a se- quence of binary latent parameters encod- ing inter-feature dependencies and (2) re- lates a language’s parameters to those of its phylogenetic and spatial neighbors. Ex- periments show that the proposed model recovers missing values more accurately than others and that induced representa- tions retain phylogenetic and spatial sig- nals observed for surface features. 1 Introduction Features of linguistic typology such as basic word order (examples are SVO and SOV) and the pres- ence or absence of tone constitute a promising resource that can potentially be used to uncover the evolutionary history of languages. It has been argued that in exceptional cases, typological fea- tures can reflect a time span of 10,000 years or more (Nichols, 1994). Since typological features, by definition, allow us to compare an arbitrary pair of languages, they can be seen as the last hope for language isolates and tiny language fam- ilies such as Ainu, Basque, and Japanese, for which lexicon-based historical-comparative lin- guistics1 has failed to identify genetic relatives. Fortunately, the publication of a large typology database (Haspelmath et al., 2005) made it pos- sible to take computational approaches to this area of study (Daum´e III and Campbell, 2007). Murawaki (2015) pursued a pipeline approach to utilizing typological features for phylogenetic inference. Exploiting interdependencies found among features, Murawaki (2015) first mapped each language, represented as a sequence of sur- face features, into a sequence of continuous latent components. It was in this continuous space that phylogenetic relations among languages were sub- sequently inferred. Murawaki (2015) argued that since the conversion and the resulting latent repre- sentations were designed to reflect typological nat- uralness, reconstructed ancestral languages were also likely to be typologically natural. In this paper, however, we show that Murawaki (2015) rests on fragile underpinnings so that they need to be rebuilt. One of the most important problems underestimated by Murawaki (2015) is an alarmingly large number of missing values. The dataset is a matrix where languages are rep- resented as rows and features as columns, but only less than 30% of the items are present after a mod- est preprocessing. What is worse, the situation is unlikely to change in the foreseeable future be- cause of the thousands of languages …

Content truncated to 500 words