Word Category Maps based on Emergent Features Created by ICA
Proceedings of the STeP'2004 Cognition + Cybernetics Symposium, page 173--185. Finnish Artificial Intelligence Society, (2004)Abstract
In this paper, we assume that word co-occurrence statistics can be used to extract meaningful features, exhibiting syntactic and semantic behavior, from text data. Independent component analysis (ICA), an unsupervised statistical method, is applied to word usage statistics, calculated from a natural language corpora, to extract a number of features. With a self-organizing map (SOM), we will demonstrate that the extracted vector representation for words can further be applied to other tasks. It is also demonstrated, that the ICA-based encoding scheme is a good alternative to random projection (RP), a method commonly used in text analysis.