Automatic Discovery of Contextual Factors Describing Phonological Variation

In this paper we describe a method for automatically discovering subsets of contextual factors which, taken together, axe useful for predicting the realizations, or pronunciations, of English words for continuous speech recognition. A decision tree is used for organizing contextual descriptions of phonological variation. This representation enables us to categorize different realizations according to the context in which they appear in the corpus. In addition, this organization permits us to consider simplifications such as pruning and branch clustering, leading to parsimonious descriptions that better predict allophones in these contexts. We created trees to examine the working assumption that preceding phoneme and following phoneme provide important contexts, as exemplified by the use of triphones in hidden Maxkov models; our results were in general accordance with the assumption. However, we found that other contexts also play a significant role in phoneme realizations. Introduction-Context Sensitivity in Realizations Phonologists claim that the context in which a phoneme occurs leads to consistent differences in how it is pronounced. For example, one phonological rule may state that the phoneme/ t / i s often flapped when it is preceded and followed by a vocalic (as in "butter"). The construction of these rules is typically an intricate process of theory formation, rule construction and then validation or disconfirmation of these rules. In this paper we describe an approach to partially automate rule construction, allowing for a larger number of examples to be examined and checked for consistencies. Our examples come from comparing transcriptions of spoken speech with a dictionary representation of the words spoken. We shall call the dictionary pronunciation symbols phonemes and define the realizations, or allophones, of a phoneme to be the set of transcription symbols corresponding to that phoneme. For example, pronunciations of the phoneme / t / inc lude the released, flapped, and unreleased realizations as characteristically occur in "tap", "butter", and "pat new", respectively. In addition, we shall refer to a context as having values. For example, the context stress has values primary, secondary, and unstressed. The approach is based on automatically forming and simplifying decision trees. Decision trees have been used for both understanding and classification of data (Henrichon and Fu, 1969). In our application, they provide a way of using context to organize the various realizations of a phoneme. The probability of a realization varies with the context in which the phoneme occurs. Contexts which have similar realization distributions are grouped together. The decision tree thus provides a method for representing the partitions of allophones with dissimilar probabilities, based on context. Decision trees can be formed automatically (Breiman et al., 1984; Quinlan, 1986) and converted to rules (Quinlan, 1987). The problem addressed here is to construct decision trees that are appropriate for use in the construction of pronunciation rules and for predicting realizations in context. In addition, an important part of the tree induction method is the discovery of appropriate descriptive categories for the formation of such trees. These categories often resemble theoretical categories, such as vocalic or plosive, and define the organization of the tree.