Discurse Marker Characterisation Via Clustering: Extrapolation from Supervised to Unsupervised Corpora

Abstract

In this paper we will show how clustering techniques provide empirical evidence for a characterisation of Discourse Markers (DMs) that helps in overcoming the lack of consensus and reduces the cost of building NLP resources based on DMs. By comparison of classifications from hand-tagged and unsupervised corpora we are capable of grounding a notion of DM prototypicality, from which reliable classifications can be obtained from fully unsupervised corpora.