Structured language models for speech recognition have been shown to remedythe weaknesses of n-gram models. All current structured language models are,however, limited in that they do not take into account dependencies betweennon-headwords. We show that non-headword dependencies contribute tosignificantly improved word error rate, and that a data-oriented parsing modeltrained on semantically and syntactically annotated data can exploit thesedependencies. This paper also contains the first DOP model trained by means ofa maximum likelihood reestimation procedure, which solves some of thetheoretical shortcomings of previous DOP models.
展开▼
