Internet evolution and progress in full automatic French language modelling
Dominique Vaufreydaz, Mathias Géry · 2005
The World Wide Web is the greatest information space ever seen, distributed all over the world, in many languages, on many various topics. We first describe the evolution of a French subset of this space during the last 3 years. During this time, the size of automatically extracted text for language modelling has multiplied by 6.5. Moreover, French coverage has grown from 140,000 to 200,000 lexical forms. So, we show that we can get more and more reliable data to train our trigram models. Recognition experiments, made on a French "state of the art" evaluation set, show that word accuracy increased from 51% up to 62.30% using two different models automatically computed on Web corpora. The first corpus was gathered at the beginning of 1999 and the last one at the end of 2000.