Novel Word-sense Identification

Automatic lexical acquisition has been an active area of research in computational linguistics for over two decades, but the automatic identification of new word-senses has received attention only very recently. Previous work on this topic has been limited by the availability of appropriate evaluation resources. In this paper we present the largest corpus-based dataset of diachronic sense differences to date, which we believe will encourage further work in this area. We then describe several extensions to a state-of-the-art topic modelling approach for identifying new word-senses. This adapted method shows superior performance on our dataset of two different corpus pairs to that of the original method for both: (a) types having taken on a novel sense over time; and (b) the token instances of such novel senses.