Machine Learning-Based Authorship Identification in Web Fictions (Student Presentation, Group 17)

This is a STAT 451 class project presentation by Fangying Zhan, Weijia Cao, and Yuan Tian This presentation is shared with the students' permission. Abstract: To tell the authors of online web fictions from analyzing the text data using snippets cut from those works, we use machine-learning algorithms to try and identify differ- ent authors’ writing styles. We collected data from Fan- Fiction.Net, the most popular online archive of fan fictions. Four authors’ chapters of fan fictions based off of Inception, Harry Potter, Avengers, and Naruto were cut into 851 snippets to constitute our dataset. All the text data were preprocessed using tokenization, lowercasing, and lemmatization. We compared two approaches of preprocessing the text data: LSM words + punctuations as features and stop word removal, two approaches of train-test splitting: a random 80%-20% split and a split based on themes, and two approaches for feature extraction: bag-of-words and tf-idf models. We employed multinomial naive Bayes classifier, which is empirically proven suitable and effective in its performance with tackling text data. We explored the performance of different approaches under 4 different scenarios to find the best setting for our model. Under Inception-non- Inception splitting, bag-of-words with multinomial naive Bayes classifier gave us the most reasonable and improved results to identify an author’s writing style.