Coevolution of Phonology and the Lexicon in Twelve Languages of West Africa

Synergetic models of language structure predict that the length of a word will depend upon various parameters such as its frequency and the number of phonemes in the language. This prediction has been used to explain word length differences within languages, but less often to explain the differences between languages. Here I show that average word length across 12 West African languages is related to the size of the phonological inventory. This is an apparent example of the adaptation of language structure to the efficient communication of information. The hypothesised mechanism by which the relationship evolves are outlined. WORD LENGTH: A SYSTEMS THEORETICAL MODEL It has long been obvious that words have very different typical lengths in different languages, from the monosyllables of Chinese and other Asian languages, to the many-syllable roots of Hawaiian and the languages of Australia. Nonetheless, the word in its uninflected stem form is considered a basic, universal linguistic unit which is comparable across languages. The question thus arises of why words should be of such different composition in different languages. Systems-theoretical linguistics treats a language as a dynamical, self-organising system whose structure is optimised to its function of communicating and representing information. The pressure exerted by function on structure is not uni-directional, however. Rather, languages evolve under several competing motivations, such as the minimisation of memory load and the minimisation of ambiguity, the minimisation of articulatory effort and the maximisation of acoustic distinctiveness. The interactions between these different pressures can be formalised into systems-theoretical models, and predictive statements about language structure produced (Kohler, 1986, 1987). According to the synergetic models, the length of an individual word will be a function of the number of segments in the phonological inventory of the language, the word's frequency, the number of words in the lexicon, and the degree of redundancy which the language requires due to the noisy nature of the human speech channel. Where L is the length of a word, then: L=a(Segments)(Frequency)(Lexicon)(Redundancy) (1) where a,b,c,d and e are constants. This statement about the distribution of word length within a language can be modified to make predictions about typical word length across languages. If we take a sample of words of all different frequencies from a language, and average their length, then their mean X will be distributed as: X = a (Segments)(Lexicon)(Redundancy) (2) We can reasonably assume that the redundancy parameter is constant across words and across *Address correspondence to: Daniel Nettle, Merton College, Oxford OX1 4JD, Great Britain. Tel.: +44 1865 286474. Fax.: +44 1865 276361. E-mail: [email protected] COEVOLUTION OF PHONOLOGY AND LEXICON 241 speech communities, given that the human speech mechanisms are the same everywhere. Furthermore, although there are differences in lexicon size between different languages, their effects will be negligible as long as all lexicons are large and d is small. We can therefore take (Lexicon) and (Redundancy) as constants when comparing across languages, giving: X = a (Segments') (3) The prediction that there will be a relationship of the form given in (3) has already been tested and found to be correct for ten unrelated languages (Nettle, 1995). In this paper, I repeat the analysis for twelve West African languages, and then investigate the mechanisms which lead to the synergetic relationship. TESTING THE PREDICTION Methods The data used to test prediction (3) were gathered as part of a wider investigation of the areal linguistics of West Africa (Nettle, 1996). The twelve languages were chosen for the quality of information available, and though they are all genetically or areally related in some way, the relationships are sufficiently uniform not to compromise statistical independence. To test the prediction, two data are needed for each language: (i) A phonological inventory This was obtained from published sources in all cases. The number of contrastive segments, henceforth S, was added up, using uniform criteria outlined in Nettle (1996). The figure desired is not simply the number of phonemes in the language by the traditional definition. Rather, it is the total number of possible contrasting segments which is of interest. This means for tone languages that each vowel/tone combination must be counted separately. Thus, a language with five vowels and three contrastive tones is deemed to have 15 possible contrasts on a vowel position. The measure S is the simple sum of vowel and consonant contrasts, and therefore takes no account of phonotactic rules operating in the language. It is thus a simplification, whose accuracy depends upon the similarity of phonotactics from language to language. (ii) An estimate of the average word length (X) This was obtained by measuring the length in segments of randomly sampled lexical stems in a dictionary of the language, and taking their mean. It is important to stress that the unit under''" investigation here is the uninflected lexical, stem. The cross-linguistic distribution of word token lengths in actual texts is heavily affected by the morphological typology of different languages, and so would require a much more complex model than that presented here. The results of Nettle (1995) suggested that a sample of 50 word stems was sufficient to produce an estimate of X. There is, however, a danger that the word length distribution will be confounded by the size of the dictionary used, since smaller dictionaries will contain more common words, which in turn tend to be shorter. An analysis in Nettle (1995), however, showed that the influence of dictionary size is negligible as long as all dictionaries contain more than 1000 entries.