Combining Relational and Distributional Knowledge for Word Sense Disambiguation

We present a new approach to word sense disambiguation derived from recent ideas in distributional semantics. The input to the algorithm is a large unlabeled corpus and a graph describing how senses are related; no sense-annotated corpus is needed. The fundamental idea is to embed meaning representations of senses in the same continuous-valued vector space as the representations of words. In this way, the knowledge encoded in the lexical resource is combined with the information derived by the distributional methods. Once this step has been carried out, the sense representations can be plugged back into e.g. the skip-gram model, which allows us to compute scores for the different possible senses of a word in a given context. We evaluated the new word sense disambiguation system on two Swedish test sets annotated with senses defined by the SALDO lexical resource. In both evaluations, our system soundly outperformed random and first-sense baselines. Its accuracy was slightly above that of a wellknown graph-based system, while being computationally much more efficient.