Final Project: A Simple, Memory-Efficient, Bayesian Part-of-Speech Tagger with constant time udpate

In this project we explore a Bayesian part-of-speech (POS) tagging technique with a focus on low memory profile and computational demands. We achieve this by representing our beliefs about a word and its corresponding part-of-speech as a probability density function (PDF) and a confidence value instead of a tag. By computing trigrams and bigrams as combinations of parts-of-speech instead of combinations of words, we reduce the memory requirement to the size of generating n-gram priors, and demonstrate a linear solution in the size of the vocabulary. By utilizing confidence metrics, we can achieve arbitrary accuracy by choosing to skip difficult words, and instead reducing search space for more sophisticated taggers by restricting possible options to a subset of possible tags in those cases.