A quantitative approach to the contrast and stability of sounds

Abstract

The phonology of a language (like other components of grammar) undergoes change in the course of time. Languages differ as to which phonemic contrasts are made and also which changes their phonological system undergoes. How intimately related these two aspects of a phonological system are remains an open question, though usually some relationship is expected. In this paper we look at two aspects of phonological systems and language change: 1) the relative stability of phonemes as presumed indirectly from a cross-linguistic study of cognates; 2) expressions of phonemic contrast with respect to a cross-linguistic study of word forms which differ in only one sound. Relative stability of sounds Ideally, we should be using diachronic data in order to directly investigate the historical stability of phonemes. However, due to the lack of suitable diachronic corpora for a cross-linguistically representative set of languages, we decided to experiment with assessing the stability of sounds indirectly by basing our work on synchronic data that is available for a wide range of languages. In particular, we decided to use the data collected as part of the ASJP (Automated Similarity Judgment Program) database (version 12, Wichmann et al. 2010), since the database includes data on a wide range of languages and in a phonetically transcribed form. Generally, historical linguists have tacitly assumed consonants to be more reliable/stable than vowels in the search for cognates as the basis for reconstructing sound changes (Campbell 2004; Walchli 2010). But can it be quantitatively defended that vowels are generally less stable than consonants? And is there a general stability cline in the sounds of the languages (either for individual families or universal)? In addressing these questions, we experimented with automatically comparing items in related languages. Since the vocabulary items in the Swadesh list are expected to be culturally neutral and stable over time, areal influence is kept to a minimum and diachronic conclusions are potentially justified. We further make a simplifying assumption that the same Swadesh item in related languages is a cognate. This is not true for all items (e.g., English tree and German Baum are not cognates, yet fill the corresponding slot in the Swadesh list), but across languages in our approach cases like this can be considered to be noise in the data. Our experiments show that setting up genealogical relationships with synchronic data on Swadesh list items yields reasonably accurate results when comparing a restricted set of languages. In Figure 1, for example, an automatically created neighbor net based on the Levenshtein distance of corresponding Swadesh items groups languages in accordance with expert classifications (see also Brown et al. 2008 for similar results). So despite of the sparse data available for individual languages we assume that interesting conclusions can be drawn when comparing languages within language families.