Word and Punctuation N-Gram Features in Romanian Authorship Attribution

This study addresses the problem of authorship attribution for Romanian texts, focusing on the use of N-gram features with an emphasis on semantic-independent representations. While character N-grams have been previously studied, this work extends the exploration to word and part-of-speech (POS) N-grams, as well as combinations involving punctuation, closed-class words, and filtered content words. Using the ROST corpus, we evaluate six supervised learning algorithms, with results averaged over multiple runs to ensure robustness. Our experiments show that Artificial Neural Networks (ANN) consistently achieve the highest performance, with word-based unigrams enhanced by punctuation reaching an average macro-accuracy of 0.93. Importantly, semantically independent features, such as closed-class words and POS replacements for nouns and verbs, yield small further improvements. These findings highlight the effectiveness of carefully designed N-gram features for Romanian AA and suggest that semantic-independent representations can complement traditional lexical approaches.