Analyzing the Impact of Text Preprocessing on Thesis Title Similarity Using TF-IDF and Cosine Similarity

Abstract

The similarity of undergraduate thesis titles may lead to the repetition of research topics and reduce the diversity of student research. This study aims to analyze the effect of preprocessing stages on thesis title similarity using the TF-IDF and Cosine Similarity methods. The dataset consists of 480 information technology thesis titles used as reference data for similarity computation. System evaluation was conducted using 50 testing titles outside the reference dataset, comprising 25 thesis titles from previous academic periods and 25 newly proposed titles submitted in the 2025 even semester. The preprocessing stages evaluated include raw text, cleaning, stopword removal, and stemming. The results indicate that each preprocessing stage produces variations in similarity values. For the previous thesis title group, the average similarity values were 36.89%, 37.07%, 36.61%, and 38.28%, respectively, while the corresponding values for the newly proposed title group were 39.18%, 39.44%, 39.07%, and 39.62%. Among the evaluated preprocessing stages, stemming produced the highest average similarity values for both testing groups. However, the improvement was relatively small, indicating that the effect of preprocessing on similarity values was limited for the dataset used in this study. In addition, this research developed a web-based system that can support a faster and more objective evaluation of undergraduate thesis title submissions.