Exploiting the Semantic Fingerprint for Tagging "Unseen" Words

In this paper we want to investigate the use of external and ”orthogonal” semantic resources in building coarse-grained semantic taggers. Our aim is to reduce the degree of supervision for the learning phase by keeping small the set of words whose behaviour has to be manually studied throughout a corpus. We introduce the notion of semantic fingerprint in order to exploit these external semantic resources in both machine learning and statistical models. Semantic fingerprints allow a straightforward integration of hierarchical information in the feature vector model. We will study and experimentally compare the effect on coarse-grained semantic taggers of different kinds of semantic fingerprints based on different semantic resources.