Integrating Linguistic Information from Multiple Sources in Lexicon Development and Spoken Language Annotation
Per-Anders Jande · 2006
In this paper, two related spoken language-oriented projects are presented. Both projects deal with integrating linguistic information from multiple sources. The first project described is the development of a multi-purpose central lexicon database including phonemic representations. Special emphasis is put on central availability and facilitating incremental development. The second project described is a spoken langue annotation project aimed at creating data for data-driven pronunciation modelling. The annotation is designed to form a general description of discourse context, including variables from the discourse level down to the articulatory feature level. A multi-layer annotation scheme for spoken language is described and the information included in the annotation is presented. Models of pronunciation variation induced from the annotation are evaluated in a tenfold cross validation experiment. On average, the models produce 8.1 % errors on the phone level. Models trained on phoneme level information only produce an average error of 14.2%. This means that including information above the phoneme level in the context description can improve model performance by 42.6%. 1.