Learning linguistic features from natural text data by independent component analysis
Helsinki University of Technology, (2005)Abstract
The analysis of natural language is an important field for language technology. The symbolic nature of written language can be encoded in numeric form and analyzed using statistical signal processing methods. In this thesis, it is assumed that word usage statistics, namely word frequencies in different contexts, contain linguistic information that can be extracted using statistical feature extraction methods. Independent component analysis, an unsupervised statistical method for blind source separation, is applied to extracting features for words using a text corpus. A study between the closeness of match between the emergent features and traditional syntactic word categories for words shows that independent component analysis extracts features that resemble more linguistic categories than features extracted with principal component analysis.