How can we perform WSD using semantic relatedness measures?

Question

I am focusing on knowledge based on Word-sense disambiguation (WSD). I have come across algorithms like Lesk, extended Lesk, etc., which perform WSD based on gloss definitions of senses of each word using WordNet.

Recently I came across a paper in which they have said that WSD can also be performed using semantic relatedness measures like Resnik, Lin, Jiang-Cornath, etc.

I am unable to understand this: how exactly can WSD be performed using similar measures?

Answer 1

One possible approach would be the following:

Semantic relatedness can be derived from word-space models. With a word-space model you could also identify the different contexts in which an ambiguous word is used (e.g. through clustering different contexts based on their vectors). If you do this, let's say, for the word bass, you should end up with two clusters - one with bass occurring in fish-related contexts, one with music-related contexts.

Once you have done this, you might be able to disambiguate new words through comparing their context vectors, and see which cluster provides the best match.

Now, this is just a suggestion. Whether or not it would work very well is a different matter. :)

(A link to the paper you mentioned would be appreciated, by the way!)