Author detection on a mobile phone

Abstract

Traditional author detection is conducted on powerful computers using documents such as books and articles. With the explosion of mobile phone computing use, modern author detection needs to be lean enough to operate on a resource restrained mobile phone and robust enough to handle the terse and non-standard wording in text messages, Tweets, and e-mails. By testing natural language and machine learning techniques for size and speed, not just effectiveness, this thesis identifies feature and technique combinations appropriate for author detection on a mobile phone. Specifically this thesis will examine effectiveness versus storage size for word grams of size 1, 2, and 5 as well as Gappy Bigrams and Orthogonal Sparse Bigrams. To deal with the robust nature of Tweets and text message, the Google Web1T corpus will be tested for size versus effectiveness in combination with the word grams. Once appropriate feature and technique combinations are found, those combinations will be tested on actual Android mobile phones to gauge how effective the chosen techniques are on a real mobile phone.