Albanian Authorship Attribution Model

Abstract

Authorship attribution (AA) is a subfield of NLP that analyzes the author's prior works to determine who wrote a text based on its features. Each natural language has its characteristics, just like every author's unique writing style. This study aims to conduct an in-depth comparison of several AA machine-learning techniques. The specially created Albanian corpus (A3C) and the English dataset (Reuters C50) have been used in the experiments. Using n-grams, we perform character-level and word-level analyses of the text to represent the author's writing style. We use five different classification algorithms to train the AA models. The TF-IDF feature vector is used as input to the models. Various experiments were conducted on the corpora. The most accurate results were obtained using word n-grams after stopword removal. The SVM algorithm performed best on the A3C dataset (Albanian). We get a 95% F1 score using SVM. On the C50 dataset (English), the SVM classifier achieved an 83% F1 score. Experiments have provided evidence of the models' robust performance on the AA corpora.