Reconnaissance thématique à partir de textes dictés et Adaptation dynamique de modèles de langages thématiques
Résumé
A robust strategy for dynamic language model selection,
based on topic recognition and switching between topic
models, is proposed. It is effective because it relies
on a small set of well trained topic-dependent language
models and on reliable topic recognition. By using
perplexity as a performance measure of the LM switching
model, a tangible reduction is observed with respect to
the use of a single, general, static LM.
Different methods are proposed for topic shift detection.
Experimental results show that different strategies for
topic shift detection have to be used depending on whether
high recall or high precision are sought.
Origine | Fichiers produits par l'(les) auteur(s) |
---|