Portability of a class-based backoff language model
Angel Xuan Chang · 2000
In this thesis, we explore how we can combine the advantages of both word and class ngrams by backing off from a word n-gram to a class-based language model (LM). In particular, we are interested in whether the resulting LM will prove to be more robust to slight changes in domain. For our purposes, words are classified based on their part-ofspeech (POS) tags. We discuss the incorporation of our class based LM into the search phase of a speech recognition system. In particular, we discuss the creation of a simple online tagger and its performance as compared to the Brill POS tagger originally used to tag the training text. We also present a series of perplexity and recognition experiments to investigate the performance of our class based LM across a variety of broadcast shows. In addition, we examine the effect of out-of-vocabulary (OOV) words on the recognition accuracy and a feature-based LM that considers the separation of root and stem. Thesis Supervisor: James R. Glass Title: Principal Research Scientist