15th International Congress of Phonetic Sciences (ICPhS-15)
Our work hypothesis is that phonotactics is best described by non-categorical, probabilistic biases, embodied in the lexicon as constraints on lexical forms which are emergent properties of the operation of dynamical systems that shape language behaviour. In order to investigate if these biases are robust enough to help a learner infer the phonotactic constraints of a language, we developed a connectionist model of an aspect of vowel-to-vowel phonotactics: harmony and contour, namely the tendency for vowels to share or avoid repetition of phonic properties. A Simple Recurrent Network was trained to predict the next phone in the word. No phonetic information was supplied to the network. The network was able to learn a significant amount of phonetic knowledge only from the distributional biases embodied in the lexicon. For example, the model was able to correctly differentiate between vowels and consonants and to discriminate between the dynamics of contour and harmony.
Bibliographic reference. Françozo, Edson / Coelho, Orlando Bisacchi / Albano, Eleonora Cavalcante / Roces, Laudino / Basso, Renato / Arantes, Pablo (2003): "A connectionist simulation of v-to-'v phonotactics learning", In ICPhS-15, 1357-1360.