Event

Syntax-based language modeling for sentence segmentation

Benoit Favre

Monday, May 12, 2008
12:30

Many NLP techniques, such as machine translation, have been developed on written text and work at a sentence level. An accurate sentence segmentation is necessary to bring those technique to automatically transcribed speech. While traditional approaches define the problem at the word level and take a decision for each boundary between two consecutive words, we aim at extending this framework to take advantage of sentence-level information. More specifically, the work presented focuses on sentence grammaticality, as measured by a probabilistic parser. The talk will focus on generating a reasonable sentence hypothesis lattice using the pointwise model, feed it to a parser, and combine the grammaticality output with the baseline system.