Modeling language evolution with codes that utilize context and phonetic features

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 136–145, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Modeling language evolution with codes that utilize context and phonetic features Javad Nouri and Roman Yangarber Department of Computer Science University of Helsinki, Finland first.last@cs.helsinki.fi Abstract We present methods for investigating pro- cesses of evolution in a language fam- ily by modeling relationships among the observed languages. The models aim to find regularities—regular correspon- dences in lexical data. We present an al- gorithm which codes the data using pho- netic features of sounds, and learns long- range contextual rules that condition re- current sound correspondences between languages. This gives us a measure of model quality: better models find more regularity in the data. We also present a procedure for imputing unseen data, which provides another method of model com- parison. Our experiments demonstrate im- provements in performance compared to prior work. 1 Introduction We present work on modeling evolution within language families, by discovering regularity in data from observed languages. The study of evolution of language families covers several problems, including: a. discovering cognates—“genetically related” words, i.e., words that derive from a common ancestor word in an ancestral proto-language; b. determining genetic relationships among languages in the given lan- guage family based on observed data; c. discover- ing patterns of sound correspondence across lan- guages; and d. reconstruction of forms in proto- languages. In this paper, we treat a. (sets of cog- nates) as given, and focus on problems b. and c.1 Given a corpus of cognate sets,2 we first aim to 1Extending the methods to problem d. is future work. 2The members of a cognate set are posited (by linguists) to derive from a common, shared origin: a word-form in the (typically unobserved) ancestral proto-language. find as much regularity as possible in the data at the sound (or symbol) level.3 An important goal is that our methods be data-driven—we aim to use all data available, and to learn the patterns of regular correspondence directly from the data. We allow only the data to determine which rules underlie it—correspondences that are inherently encoded in the corpus itself—rather than relying on exter- nally supplied (and possibly biased) rules or “pri- ors.” We try to refrain from a priori assumptions or “universal” principles—e.g., no preference to align consonants with consonants, to align a sym- bol with itself, etc. We claim that alignment may not be the best way to address the problem of regularity. Finding alignments is indeed finding a kind of regularity, but not all regularity is expressed as alignment. The paper is organized as follows. In section 2 we review the data used in our experiments and re- cent approaches to modeling language evolution. We formalize the problem and present our mod- els in section 3. The models treat sounds as vec- tors of phonetic features, and utilize the context of the sounds to discover patterns of regular corre- spondence. Once we have obtained the regularity, …

Content truncated to 500 words