Using Word Embeddings for Bilingual Unsupervised WSD

Unsupervised Word Sense Disambiguation (WSD) is one of the challenging problems in natural language processing. Recently, an unsupervised bilingual WSD approach has been proposed. This approach uses context aware EM formulation for estimating the sense distribution by using the co-occurrence counts of cross-linked words in comparable corpora. WordNetbased similarity measures are used for approximating the co-occurrence counts. In this paper, we explore the feasibility of the use of Word Embeddings for approximating these counts, which is an extension to the existing approach. We evaluated our approach for Hindi-Marathi language pair, on Health domain. On using the combination of Word Embeddings and WordNet-based similarity measures, we observed 8.5% and 2.5% improvement in the F-score of verbs and adjectives respectively for Marathi and 7% improvement in the F-score of adjectives for Hindi. The experiments show that the combination of Word Embeddings and WordNetbased similarity measures is a reasonable approximation for the bilingual WSD.