Sabine Deligne
Statistical language modeling based on variable
length dependencies
Abstract:In this report, we present a stochastic language modeling tool which aims at retrieving variable-length phrases (multigrams), assuming bigram dependencies between them. The phrase retrieval can be intermixed with a phrase cluster- ing procedure, so that the language data are iteratively structured at both a paradigmatic and a syntagmatic level in a fully integrated way. Perplexity re- sults on ATR travel arrangement data with a bi-multigram model (assuming bigram correlations between the phrases) come very close to the trigram scores with a reduced number of entries in the language model. Speech recognition scores are ranked accordingly. Also the ability of the class version of the model to merge semantically related phrases into a common class is illustrated.