Finding word senses in tagging system

Abstract

Tagging is popular in blog and social website because a tag can search and describe a file instead of predefined category. However, people tend to give the tags liberally resulting in many similar, obsolete and ambiguous tags within the system. Searching efficiency can be severely reduced. Some situations also affect searching accuracy, such as the polysemous or synonymous tags. Although latent semantic indexing (LSI) can disambiguate the word senses but in slow computation time as the large dataset. An approach of combining random projection with LSI is proposed to speed up the performance. However, the result of a context-based document system shows that the performance is only improved in a smaller dataset. For this observation, combining random projection with LSI in tag-based system should be able to gain improvement as the dataset in tag-based system should be smaller than a context-based system. Also, to the best of my knowledge, this approach has not been experimented in tag-based system. In this project, the results show that LSI with random projection is able to reduce 34% running time with about 70% accuracy. Also, I attempt to apply adaptive folding-up algorithm to update SVD dynamically, but it does not always retrieve high accuracy result.