Authorship Identification Using Stylometry and Document Fingerprinting

Abstract

Authorship identification focuses on finding the particular author of an anonymous or unknown document by extracting various predictive features related to that document. It helps to predict the most probable author of articles, messages, codes, or news. This task is generally viewed as a multi-class, single labeled categorization of text. This topic is important and among the most interesting topics in the field of Natural Language Processing. There are many applications in which this can be applied such as identifying anonymous author, supporting crime investigation and its security, also detecting plagiarism, or finding ghostwriters. Till now, most of the existing works are based on the character n-grams of fixed length and/or variable length to classify authorship. In this work, we tackle this problem at different levels with increasing feature engineering using various text-based models and machine learning algorithms. The propose work analyses various types of stylometric features and define individual features that are high in performance for better model understanding. We evaluate the proposed methodology on a part of Reuters news corpus. It consists of texts of 50 different authors on the same topic. The experimental results suggest that, using document finger printing features enhance the accuracy of classifier. Moreover, PCA (Principal Component Analysis) further improves the results. In addition, we compare the results with other works related to the authorship identification domain.