Dealing with distant relationships in natural language modelling for automatic speech recognition - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2000

Dealing with distant relationships in natural language modelling for automatic speech recognition

Résumé

Classical statistical language models, called n-gram models, describe natural language using the probabilistic relationship between a word to predict and the n-1 contiguous words preceding it. Obviously, the linguistic relationships present in a sentence are more complex. A first remark is that there exist distant relationships. We present here some recent work on an alternative model to n-gram models, based on the split of the history, dealing with the interpolation between distant bigram models. More precisely, our model is a cheaper alternative to high order n-grams. In conventional n-grams, when n is greater than 3, events are less frequent and statistics are not reliable. To deal with this problem, and to accurately estimate parameters, we combine a smoothed bigram with distant 3-bigram, distant 4-bigram and a cache composed of 100 words. We present new progresses obtained by using a simulated annealing algorithm in order to calculate the best parameters of this linear combination. With a 20K vocabulary and 40 million words for training, our algorithm improved the perplexity by 5.4% in comparison with the Baum-Welch algorithm. Moreover, this new model outperforms a smoothed bigram by 6.1% in terms of perplexity.
Fichier principal
Vignette du fichier
A00-R-144.pdf (94.41 Ko) Télécharger le fichier
Loading...

Dates et versions

inria-00099031 , version 1 (26-09-2006)

Identifiants

  • HAL Id : inria-00099031 , version 1

Citer

David Langlois, Kamel Smaïli, Jean-Paul Haton. Dealing with distant relationships in natural language modelling for automatic speech recognition. 4th World Multiconference on Systemics, Cybernetics & Informatics - SCI'2000, International Institute of Informatics & Systemics, 2000, Orlando, USA, pp.400-405. ⟨inria-00099031⟩
113 Consultations
79 Téléchargements

Partager

Gmail Facebook X LinkedIn More