Feature Selection with Exception Handling Using Adaptive Distance Measures -an Example from Phonetics

The goal in this paper is to show how the classiication of patterns of phonetic features (=phones) to phonemes can be acquired. This classi-cational process is modelled by a supervised feature selection method, based on a weighted Hamming distance, augmented by Boolean functions describing exceptions. An important aspect is the diierentiation of rules and exceptions during learning. 1 Phonetic features and Phonemes The goal in this paper is to show how the classiication of patterns of pho-netic features (=phones) to phonemes can be acquired. In every language a number of diierentiable phones belong to one phoneme. To learn the phonemic pattern of a language amounts to learn a classiication of all naturally occurring phones to a phoneme. The continuum of articulatory places or acoustically deened frequency formants for a single phone can be cut up into a set of descriptive features. Phones are then represented as in IPA-notation, where a single symbol is used for each naturally attested occurence of a bundle of phonetic features, or directly by a number of phonetic segmental features. Phonetic features for German vowels, the corresponding patterns and their intended classiication are shown in Table 1, adapted from Hym75]. (Non-segmental features such as length,syllabicity, nasalization, rhotaciza-tion, diphthongs have been excluded from this analysis.) As has been noted by phonologists for a long time ((Tru58]), the grouping of phones to phonemes gives rise to phonological systems, organized 1