An unsupervised EM method to infer time variation in sense probabilities.

The inventory of senses for a given word changes over time – tweet has gained the ‘Twitter post’ sense only relatively recently and this paper addresses the problem of the computational detection of such change. We propose a generative model which conditions context words on a target expression’s sense and conditions the sense choice on the time of writing. We develop an EM algorithm to estimate the parameters from raw time-stamped ngram data with no sense annotation. We are able to demonstrate the inference of parameters which plausibly reflect the objective dynamics of sense use frequencies, in particular the emergence of a new sense.