Skip to Main content Skip to Navigation
Theses

Notions d'événements distants et d'évenements impossibles en modélisation stochastique du langage : application aux modèles n-grammes de mots et de séquences

Abstract : A statistical language model (LM) deals only with specific linguistic events and so, is not sufficient to describe the entire language. One must then combine several LM in order to recover a great part of the language. I propose, instead of systematically and linearly combining all the LMs, to select the best one following the linguistic context. I define a measure of the prediction capacity of a LM depending on the history. I apply this measure to the comparison of several distant n-gram LMs. This improves by 5.4% the performance of the linear combination. Moreover, this method allows to define phrases as new lexical units in the vocabulary. These new units outperforms a baseline model by 21% in terms of perplexity and 12.5% in terms of word error rate for a speech recognition task. An other contribution of this work consists in automatically collecting impossible events in french language. 60 millions of impossible bigrams are collected using heuristics based on information theory.
Document type :
Theses
Complete list of metadata

https://hal.univ-lorraine.fr/tel-01746703
Contributor : Thèses UL Connect in order to contact the contributor
Submitted on : Thursday, March 29, 2018 - 10:42:59 AM
Last modification on : Saturday, October 16, 2021 - 11:26:08 AM

Links full text

Identifiers

  • HAL Id : tel-01746703, version 1

Collections

Citation

David Langlois. Notions d'événements distants et d'évenements impossibles en modélisation stochastique du langage : application aux modèles n-grammes de mots et de séquences. Autre [cs.OH]. Université Henri Poincaré - Nancy 1, 2002. Français. ⟨NNT : 2002NAN10048⟩. ⟨tel-01746703⟩

Share

Metrics

Record views

22