Reliable Detection of Important Word Boundaries Using Prosodic Features

Kaufhold C, Stemmer G, Nöth E (2011)


Publication Status: Published

Publication Type: Conference contribution

Publication year: 2011

Journal

Publisher: Springer-verlag

Book Volume: 6836

Pages Range: 259-267

Abstract

Natural input dialog systems for entering address data in modern GPS units demand a significantly more robust extraction of slot information. In previous work, prosody was used to detect phrase boundaries (PB) which separate particular address parts. Only input without filler words was used. In this work, carrier sentences are allowed. A boosting approach provides twenty strong prosodic features which model the characteristic of PBs. We introduce the concept of a prosodically marked word boundary (PMB), which enables a better location of the provided information in natural input. Our results on a dataset of 5883 input samples reveal that about 67% of the found PBs indicate a PMB, while most of the remaining boundaries occur within compound words.

Authors with CRIS profile

How to cite

APA:

Kaufhold, C., Stemmer, G., & Nöth, E. (2011). Reliable Detection of Important Word Boundaries Using Prosodic Features. (pp. 259-267). Springer-verlag.

MLA:

Kaufhold, Caroline, Georg Stemmer, and Elmar Nöth. "Reliable Detection of Important Word Boundaries Using Prosodic Features." Springer-verlag, 2011. 259-267.

BibTeX: Download