Robust model selection and the statistical classification of languages

Abstract

In this paper we address the problem of model selection for the set of finite memory stochastic processes with finite alphabet, when the data is contaminated. We consider m independent samples, with more than half of them being realizations of the same stochastic process with law Q, which is the one we want to retrieve. We devise a model selection procedure such that for a sample size large enough, the selected process is the one with law Q. Our model selection strategy is based on estimating relative entropies to select a subset of samples that are realizations of the same law. Although the procedure is valid for any family of finite order Markov models, we will focus on the family of variable length Markov chain models, which include the fixed order Markov chain model family. We define the asymptotic breakdown point (ABDP) for a model selection procedure, and we show the ABDP for our procedure. This means that if the proportion of contaminated samples is smaller than the ABDP, then, as the sample size grows our procedure selects a model for the process with law Q. We also use our procedure in a setting where we have one sample conformed by the concatenation of sub-samples of two or more stochastic processes, with most of the subsamples having law Q. We conducted a simulation study. In the application section we address the question of the statistical classification of languages according to their rhythmic features using speech samples. This is an important open problem in phonology. A persistent difficulty on this problem is that the speech samples correspond to several sentences produced by diverse speakers, corresponding to a mixture of distributions. The usual procedure to deal with this problem has been to choose a subset of the original sample which seems to best represent each language. The selection is made by listening to the samples. In our application we use the full dataset without any preselection of samples. We apply our robust methodology estimating a model which represent the main law for each language. Our findings agree with the linguistic conjecture, related to the rhythm of the languages included on our dataset.