Fast logistic regression for text categorization with variable-length n-grams