Imegrating Domain and Paradigmatic Similarity for unsupervised Sense Tagging

An unsupervised methodology for Word Sense Disambiguation, called Dynamic Domain Sense Tagging, is presented. It relies on the convergence of two very well known unsupervised approaches (i.e. Domain Driven Disambiguation and Conceptual Density). For each target word a domain is dynamically modeled by expanding the its topical context, i.e. a set of words evoking the underlying/implict domain where the word is located. The estimation of the paradigmatic similarity within such a specific lexicon is assumed as a disambiguation model. The Conceptual Density measure is here used to account for paradigmatic associations, and the top scored senses of the target word are selected accordingly. Results confirm the impact of domain based representation in capturing useful paradigmatic generalizations, especially when small text fragments are available. In addition, the precision/recall tradeoff of the resulting method can be tuned in a meaningful way, allowing us to achieve impressively high precision scores in a purely unsupervised setting.