Towards a Better Collaboration Between a n-class and a n-gram Language Model - Inria - Institut national de recherche en sciences et technologies du numérique
Communication Dans Un Congrès Année : 1999

Towards a Better Collaboration Between a n-class and a n-gram Language Model

Résumé

This paper deals with the combination of a trigram and a triclass. This combination is done in an original manner which is different from a simple linear interpolation. The method, we propose is based on an oriented search for the potential words in a speech process which is simulated through the Shannon game. The combined model presented in this paper outperforms in terms of perplexity, rank of words (for the 100 first propositions) the trigram model and also the one we proposed in the AUPELF (linear interpolation between trigram and triclass). The oriented model reveals that the cooperation between a trigram and a triclass is better when it is used out of the classical linear interpolation.
Specomcomplet.pdf (2.87 Mo) Télécharger le fichier

Dates et versions

inria-00108063 , version 1 (19-10-2006)
inria-00108063 , version 2 (21-11-2017)

Identifiants

  • HAL Id : inria-00108063 , version 2

Citer

Kamel Smaïli, Imed Zitouni, Jean-Paul Haton. Towards a Better Collaboration Between a n-class and a n-gram Language Model . SPECOM, Oct 1999, Moscow, Russia. ⟨inria-00108063v2⟩
160 Consultations
19 Téléchargements

Partager

More