Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Integration of Several Information Sources for Robust Class-Based Statistical Language Modelling

Géraldine Damnati

France Télécom, CNET DIH/DIPS, Lannion, France

The lack of training material, which is particularly problematic in the context of human-machine spoken dialogue, is addressed in this paper. Automatic classification is known to be useful to model spontaneous speech, but contextual methods are, by nature, sensitive to the data. In test conditions similar to training, the heterogeneity of classes does not matter, as long as they faithfully reflect training. When considering some robustness issues, such as domain or vocabulary adaptation, consistency of classes becomes relevant, and classical contextual algorithms can not provide this when faced to a low amount of training data. Hence, this paper proposes a new theoretical framework for hierarchical clustering, where several information sources can be combined at the similarity criterion evaluation level. This allows a priori information to be taken into account as a complement of the contextual information. Consistent classes can be built, while remaining suited for spontaneous speech modelling.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Damnati, Géraldine (1999): "Integration of several information sources for robust class-based statistical language modelling", In EUROSPEECH'99, 1579-1582.