Software to measure rambling, cognitive difficulty and degree expression in schizophrenic speech

This thesis develops new software to help distinguish schizophrenic speech from healthy speech. I used the Natural Language Analysis Tools (Covington 2002) and developed the Natural Language Feature Extraction program to collect linguistic features from forty four speech samples. Decision trees and neural networks use these linguistic features to classify the speech samples. The following linguistic features significantly distingui sh schizophrenic speech from healthy speech: 1) mean sentence length, 2) degree expressions, and 3) difficulty words. Decision trees classify 85.7% of the speech sample s correctly; neural networks classify 86% of the speech samples correctly. Schizophrenics longer sentences may be a manifestation of lack of discourse planning, rambling, and running together of thoughts and sentences. The smaller number of degree expressions may be a manifesta tion of lack of theory of mind. The increased amount of difficulty words may be a sign of cognitive difficulty with interpreting information. Limitations and future research directions are discussed.