Classifying Human and Machine Generated Text
Norrec Nieh, Frank Rossi, & Jason Zhang Khoury College of Computer Sciences Northeastern University Boston, MA 02115
nieh.c@northeastern.edu, rossi.f@northeastern.edu, zhang.haozhe1@northeastern.edu
Abstract
We investigated the ability of perplexity to classify texts as human or machine generated using two approaches, a single perplexity score and a sequence of word probabilities corresponding to the input text. The former was classified according to a single threshold and the latter was fed into a neural network. These perplexities and probabilities were generated using N-grams. Our best result for the single score, threshold classifier was 77% and our best result for the probability sequence, ANN classifier was 80%. Our work demonstrates that perplexity can be used as a feature to distinguish human and machine texts even with basic classifiers.