Language Modeling with Sentence-Level Mixtures

Thispaperintroduces a simple mixtare languagemodelthat attempts to capture long distance conslraints in a sentence orparagraph. The model is an m-component mixture of Irigram models. The models were constructed using a 5K vocabulary and trained using a 76 million wordWailStreetJournal textcorpus. Using theBU recognition system, experiments show a 7% improvement in recognition accuracy with themixture trigrammodels ascomparedto using a Irigram model.

• Type: Conference
• Year: 1994
• Where: NAACL
• Authors: Rukmini Iyer, Mari Ostendorf, Jan Robin Rohlicek
• Keywords: Irigram Model, Long Distance Conslraints, NAACL 1994, NAACL 2007, Sentence Orparagraph