This study aims to measure the similarity of melodies objectively using natural language processing (NLP) techniques. We utilize Mel2word which is a melody tokenization method based on byte-pair encoding to facilitate the semantic analysis of melodies. In addition, we apply two word weighting methods: the modified Tversky measure for word salience and the TF-IDF method for word importance and uniqueness, to better understand the characteristics of each melodic element. We validate our approach by comparing song vectors calculated from an average of Mel2Word vectors to the ground truth in 108 cases of music copyright infringement, sourced from an extensive review of legal documents from law archives. The results demonstrate that the proposed approach is more in accordance with court rulings and perceptual similarity.