Unsupervised Word Segmentation: does tone matter ?
Résumé
In this paper, we investigate the usefulness of tonal features for unsupervised word discovery, taking Mboshi, a low-resource tonal language from the Bantu family, as our main target language. In a preliminary step, we show that tone annotation improves the performance of \emph{supervised learning} when using a simplified representation of the data. To leverage this information in an unsupervised setting, we then present a probabilistic model based on a hierarchical Pitman-Yor process that incorporates tonal representations in its backoff structure. We compare our model with a tone-agnostic baseline and analyze if and how tone helps unsupervised segmentation on our small dataset.
Origine : Fichiers produits par l'(les) auteur(s)