An Empirical Framework for Bangla Word Sense Disambiguation Using Statistical Approach

Abstract

Recently, word sense disambiguation has gained increased attention by NLP practitioner due to its various potential applications in language technology. This paper proposes a Naïve Bayes classifier for resolving lexical ambiguities of Bangla words with the help of a Bangla sense annotated corpus. At the initial stage, a Bangla sense annotated corpus is generated from a raw text corpus for serving as a training dataset. For a given input Bangla sentence, ambiguous words detection is done first and then Bayes probability theorem is applied to calculate the posterior probability that an ambiguous word belongs to a particular sense class. The values of posterior probability of several senses of the detected ambiguous word finally train the Naïve Bayes classifier to classify a closest sense of the ambiguous word. Experimental outcome reveals that the proposed method outdoes existing techniques by achieving the highest F1-score of $90\%$ on the test data.