Ricerc@Sapienza

Natural Language Processing has long sought ways to reliably map a word or longer expression in context to some representation of its meaning. The most popular framing of this problem is so-called Word Sense Disambiguation (WSD), in which an algorithm has to classify the contextual target into a set of sense classes defined by a dictionary. However, WSD suffers from many intrinsic problems: excessively fine granularity of the inventories, extreme class imbalance, paucity of training data. In our research we tackle these issues by incorporating pre-existing knowledge into supervised WSD systems: glosses, graph relations, multimodal information. Moreover, we investigate alternative framings for computational lexical semantics, i.e. generating a gloss directly -- doing away with sense inventories completely; or embedding sense representation into graph-like representations of larger units like sentences.