Authorship Attribution Using Capsule-Based Fusion Approach

Authorship attribution is an important task, as it identifies the author of a written text from a set of suspect authors. Different methodologies of anonymous writing, have been discovered with the rising usage of social media. Authorship attribution helps to find the writer of a suspect text from a set of suspects. Different social media platforms such as Twitter, Facebook, Instagram, etc. are used regularly by the users for sharing their daily life activities. Finding the writer of micro-texts is considered the toughest task, due to the shorter length of the suspect piece of text. We present a fusion based convolutional Neural Network model, which works in two parts i) feature extraction, and ii) classification. Firstly, three different types of features are extracted from the input tweet samples. Three different deep-learning based techniques, namely capsule, LSTM, and GRU are used to extract different sets of features. These learnt features are combined together to represent the latent features for the authorship attribution task. Finally the softmax is used for predicting the class labels. Heat-maps for different models, illustrate the relevant text fragments for the prediction task. This enhances the explain-ability of the developed system. A standard Twitter dataset is used for evaluating the performance of the developed systems. The experimental evaluation shows that proposed fusion based network is able to outperform previous methods. The source codes are available at https://​github.​com/​chanchalIITP/​AuthorIdentifica​tionFusion.