Independent component analysis of word contexts and comparison with traditional categories
Jaakko J. Väyrynen, Timo Honkela, Aapo Hyvärinen · 2004
We study written language as if it were a multidimensional signal rather than a stream of symbols. We show that it is possible to find emergent features by independent component analysis from word contexts. The closeness of match between the learned features and traditional linguistic word categories is examined. It is shown that independent component analysis performs better than principle component analysis.