Finding diachronic sense changes by unsupervised methods

Abstract

An existing word in a language may acquire a new meaning as time passes, in addition to meanings it already possesses. Such a new word meaning pairing is called a semantic neologism. An example of a semantic neologism in English is 'tweet' referring to 'a post on twitter', which is a newly acquired sense, in addition to its long established sense of referring to 'a bird song'. This thesis addresses the problem of the computational detection of such changes from time-stamped raw text i.e., text without sense annotations. For this, a generative model is proposed with variables for time Y, sense S, and contexts w around a given target word (a potential semantic neologism). A target word will be treated as having is exhibiting K senses over a given time period. The model has senses dependent on times, expressed by P(S\Y), and context words dependent on senses, expressed by P(w\S). This reflects first a reasoning that sense ratios change over time. Secondly it reflects another reasoning that senses themselves, each seen as a probability distribution over words, can be regarded as more or less eternal, and subject to little change over time.