Tracing the roots of syntax with Bayesian phylogenetics

Abstract

The ordering of subject, verb, and object is one of the fundamental components of the syntax of natural languages. The distribution of basic word orders across the world's languages is highly nonuniform, with the majority of languages being either subjectobject-verb (SOV) or subject-verb-object (SVO). Explaining this fact using psychological accounts of language acquisition or processing requires understanding how the present distribution has resulted from ancestral distributions and the rates of change between orders. We show that Bayesian phylogenetics can provide quantitative answers to three important questions: how word orders are likely to change over time, which word orders were dominant historically, and whether strong inferences about the origins of syntax can be drawn from modern languages. We find that SOV to SVO change is more common than the reverse and VSO to SVO change is more common than VSO to SOV, and that if the seven language families we consider share a common ancestor then that common ancestor likely had SOV word order, but also that there are limits on how confidently we can make inferences about ancestral word order based on modern-day observations. These results shed new light on old questions from historical linguistics and provide clear targets for psychological explanations of word-order distributions. language evolution | computational historical linguistics | psycholinguistics T he sentence "dog bites man" is easily understood by a speaker of English, but switching the order of the words either renders it incomprehensible ("dog man bites") or changes its meaning ("man bites dog"). Only one of the six possible orders of these three words-the subject (S), verb (V), and object (O)-is commonly used in English sentences. However, the ordering of these three elements is a fundamental component of syntax, and varies significantly across languages. English is an SVO language, but among the world's languages, the six possible basic word orders are, from most to least common: SOV (48%), SVO (41%), VSO (8%), VOS (2%), OVS (1%), and OSV (0%) (1, 2).Several attempts have been made to explain the crosslinguistic word-order distribution in psychological terms (3-6). All of these efforts have proceeded under the assumption that SOV is somehow optimal (e.g., requiring minimum working memory for parsing), since it is the most common word order. This is a "functionalist" approach, where some word orders are claimed to better facilitate the transmission of information. However, several linguists have noted that change from SOV to SVO appears to have been more common historically than change from SVO to SOV (7-10). If we suppose that, on the whole, languages tend to change from less to more functional word orders, then we should seek theories by which SVO is psychologically superior to SOV, not the other way around. Under this view, SOV is predominant not due to greater functionality but as a vestige of even greater dominance in the past.Traditional functionalist explanations also assume that VSO is les...

Analyzing Correlated Evolution of Multiple Features Using Latent Representations

“…Surface featurebased estimation had no other way to explain the presence of these uncommon word orders than slowing down the convergence to the stationary distribution (otherwise they go extinct). Maurits and Griffiths (2014) also reported some counterintuitive results regarding OVS and OSV. By contrast, latent parameter-based estimation appears to have explained the low frequencies partly with the stochasticity of observation associated with the latent-to-surface mapping.…”

“…We collected 50 calibration points from secondary literature (Holman et al, 2011;Bouckaert et al, 2012;Gray et al, 2009;Maurits and Griffiths, 2014;Grollemund et al, 2015). For example, we set a Gaussian prior with mean 2,500 BP (before present) and standard deviation 500 on the date of (Proto-)Hmong-Mien.…”

“…As a proof-of-concept demonstration of the proposed framework, we investigate the BWO feature, or WALS's Feature 81A (Dryer, 2013b). The cross-linguistic variation of BWO attracts attention not only from typologists but from psycholinguists (see Maurits and Griffiths (2014) for a brief review). Some claim that the fact that SOV is the most frequent order indicates its optimality, presumably in terms of functionality.…”

“…To perform statistical inference, we assume that each feature evolves along the branches of the tree according to a continuous-time Markov chain (CTMC) model, which is controlled by a transition rate matrix (TRM). Once TRMs are estimated, we can gain insights from them, for example, by simulating language evolution (Maurits and Griffiths, 2014). One problem in previous studies is that they do not adequately model a characteristic of typological features that has been central to linguistic typology, that is, the fact that these features are not independent but depend on each other (Greenberg, 1963;Daumé III and Campbell, 2007).…”

Measuring and Modeling Language Change

“…We will conclude with a discussion of open research questions for the analysis of language change and diachronic textual corpora (Nerbonne, 2010;Eisenstein, 2013;Maurits and Griffiths, 2014;Perek, 2014).…”