Evaluation of summarising and paraphrasing techniques in text classification

In this thesis, the performances of summarization, paraphrasing, and semantic similarity measurement methods applied to news articles are comprehensively examined. Initially, the texts were preprocessed and represented using the Bag of Words (BoW) model. Subsequently, classification experiments were conducted on different machine learning models (Logistic Regression, Support Vector Machines, Random Forest, and Perceptron) using the stratified k-fold cross-validation method; the model with the highest performance was employed in the subsequent stages. In the summarization stage, only extractive methods were preferred, and TextRank, LexRank, TF-IDF-based summarization, and rule-based methods were applied. In the paraphrasing stage, two different methods—WordNet-based and Back-Translation—were used. The generated summaries and paraphrased texts were evaluated using both the Cosine Similarity and BERTScore metrics. The analyses were carried out by comparing raw texts with summarized texts; raw texts with summarised and paraphrased texts; and raw texts with directly paraphrased texts. The findings revealed that, in some cases, the semantic similarity scores of summarised-and-paraphrased texts could be higher than those of solely summarized texts; however, in general, directly paraphrased texts achieved the highest similarity scores. This study demonstrates the effects of using summarization and paraphrasing methods together on semantic similarity and provides guiding insights for method selection in text processing applications.