An author identification approach based on integration of multiple text representations using heuristic algorithms

Advancements in artificial intelligence (AI) and natural language processing (NLP) have made it possible to identify authorship from textual data. In this thesis study, an authorship attribution application was developed using texts belonging to different authors. In the study, the texts were vectorized separately using three different methods: TF-IDF, Word2Vec, and BERT models. The obtained vector representations were then combined through various heuristic (sezgisel) algorithms, and a hybrid representation system was proposed. The purpose of this hybrid structure is to enhance the overall performance of the model by jointly considering the frequency information of words, their semantic relationships, and the contextual meanings at the sentence level. In constructing the hybrid vectors, text representations obtained from both a fine-tuned BERT model and a non-fine-tuned BERT Mean Pooling-based model were utilized. The resulting hybrid representations were evaluated using different classification algorithms. Experimental results demonstrated that combining vector representations through heuristic algorithms significantly improved system performance. In particular, the strong contribution of word frequency information, the incorporation of semantic relationships, and the inclusion of sentence-level contextual analysis emerged as critical components for achieving optimal performance. As a result, the proposed hybrid vector representation approach achieved high accuracy in the authorship identification task and was found to be an effective and generalizable method in the field of text classification.