Automatic authorship attribution in Albanian texts.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Machine Learning

Background:

• Authorship attribution research is hindered by a lack of corpora for under-resourced languages.
• Albanian authorship attribution lacks sufficient datasets and methodological exploration.

Purpose of the Study:

• To investigate authorship attribution in Albanian, a low-resource language.
• To develop and evaluate machine learning and deep learning models for Albanian authorship identification.
• To identify optimal feature sets and classification strategies for this task.

Main Methods:

• Compiled a new corpus of Albanian newsroom columns and literary works.
• Engineered hand-crafted linguistic features (lexical, morphological, structural).
• Applied and compared various machine learning classifiers (e.g., XGBoost) and deep learning models (fastText, BERT-multilingual).

Main Results:

• Lexical features significantly improved algorithm performance in authorship attribution.
• XGBoost achieved the highest performance, with F1 scores of 0.982 (literary works) and 0.905 (newsroom columns).
• Deep learning models like fastText and BERT-multilingual showed promising results for specific Albanian text types.

Conclusions:

• Lexical features are highly effective for Albanian authorship attribution.
• XGBoost offers a robust machine learning approach for this task.
• The study provides a framework and valuable insights for authorship attribution in low-resource languages.