Statistical Models for Natural Language Processing
Kenneth Church · 2016
There is a considerable literature on applications of statistical methods in natural-language processing. This chapter focuses on two types of applications: (1) recognition/transduction applications based on Shannon’s Noisy Channel such as speech recognition, optical character recognition (OCR), spelling correction, part-of-speech (POS) tagging, and machine translation (MT); and (2) discrimination/ranking applications such as sentiment analysis, information retrieval, spam email filtering, author identification, and word sense disambiguation (WSD). Shannon’s Noisy-Channel model is often used for the first type, and linear separators such as Naive Bayes and logistic regression are often used for the second type. These techniques have produced successful products that are being used by large numbers of people every day: web search, spelling correction, translation, etc. Despite successes such as these, it should be mentioned that all approximations have their limitations. At some point, perhaps in the not-too-distant future, the next generation may discover that the low-hanging fruit has been pretty well picked over, and it may be necessary to revisit some of these classic limitations.