Part-of-speech n-gram and word n-gram fused language model

In this paper, an accurate and com pact language m odel is proposed to cope robustly with data sparseness and task dependencies. This language m odel adopts new categories which are generated by continuously interpolating POS word-class categories and word categories using M AP estimation. Thenew categories can reflect word statistics efficiently without loosing accuracy and task-independent general word-characteristics (i.e. gram m atical constraints captured by POS statistics) are em bedded to prevent taskovertuning. This m odeling reduces the m odel size to 50% of the conventional m odels. The bidirectional word-cluster N-gram s generated by this m odeling have 3% lower perplexity m easured on a m atched dom ain and 15% lower on a m ism atched domain compared to a conventi onal word 2-gram.