Prosodic and Syntactic markers in Spoken French: Why, what and how should one count? A comparative statistical perspective

We highlight how the variable frequency of prosodic, syntactic and intonosyntactic features is a strong predictor for discourse genre classification in speech. We quantify this claim by studying the mutual information these features share with the situational variables considered, allowing us to identify possibly redundant information and to adequately predict genres from intonosyntactic annotations. In our quantitative study, we compare two methods for supervised classification: decision trees and support vector machines. Furthermore, we address the issue of graphical representations of features through a principal component analysis. All these points provide valuable feedback on the role played by the intonosyntactic interface in the identification of discourse genres in spoken French.