Investigating The Efficiency Of Machine Learning Methods In Authorship Detection For Low-Resourced Languages: The Case Of Kurdish Authors

Abstract

Textual data continues to multiply with time, Alongside the exponential growth of textual information, an increase in anonymous material has also been seen. Authorship detection has significant potential for usage in numerous applications of authorship analysis, such as history and literary science, Forensic examination, or Plagiarism detection. We manually collected 2798 documents by 150 authors for this study in order to investigate how effectively existing machine learning algorithms can differentiate Kurdish authors from unidentified writings. The approach that has been developed uses a TF-IDF technique to calculate the weight of each token and extracts the token frequency of each token, ranging from 1 to 5 grams, as a feature to find a pattern in each author's text. We train SVM, CNB, MNB, and K-NN classifiers with a collection of available documents because an unknown document's essential tokens are similar to a known document's crucial tokens. Then we give it a mysterious document so it may assess how closely it resembles the known document. We achieved an accuracy of 80% by SVM with both O-V-O and O-V-R approaches for the token 1-gram, also a promising results in precision, recall, and F1-score measures. Furthermore, to our knowledge, this is the first study to investigate authorship detection for the Kurdish language.

Keywords

• authorship detection
• nlp
• authorship analysis
• klpt
• ml
• tf-idf