Previous attempts to a utomatically determine m ulti-words as the basic unit for language modeling h ave been successful for extending bigram models 10, 9, 2, 8 to improve t he perplexity o f t he language model and or the w ord accuracy of the speech d ecoder. However, none o f t hese techniques gave improvements over the trigram model so far, except for the rather controlled ATIS task 8 . We t herefore propose an algorithm, that minimizes the perplexity improvement o f a bigram model directly. The n ew algorithm is able to reduce the trigram perplexity a n d also achieves word accuracy improvements in the V erbmobil task. It is the n atural counterpart of successful word classi cation algorithms for language modeling 4, 7 that minimize the leaving-one-out bigram perplexity. W e also give some d etails on the usage of class nding t echniques and m-gram models, which can be crucial to s u ccessful applications of this technique.
Due t o r o b u s t n e s s , l e a r n a b ilitya n d e a s e o f i n t e g r a t iono f e r e n t i n f o r m a t ionsources, c o n n e c t ionist p a r s ings y s t e m s r p a r s ing s p o k e n l a n g u a g e ,
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.