Distributional Similarity in the Varied Order of Syntactic Spaces

The single valued decomposition (SVD) has been used widely in information retrieval and other language engineering projects. Another less-widely used technique is independent component analysis (ICA), which postulates much stricter conditions than SVD and aims to represent the meaningful structure behind data. We compare distributional word similarity estimates using raw syntactic relationships frequencies extracted from a large corpus, as well as SVD-reduced and ICA-extracted attributes. The results show that ICA does not help SVD in the syntactic space