Using Natural Language Processing Algorithms to Find Musical Similarity
Name
Mait Lindpere
Abstract
In this master's thesis, an experimental study was conducted to research whether it is feasible to recommend music according to it’s semantic description (in the context of this work, album descriptions). In order to solve the set task, NLP methods are used to extract the similarity of music from metadata and compare it to assessments of people who have rated the music in question. The music similarity found by audio methods is used for comparison to assess the performance of NLP methods. In addition, the research combines NLP and audio methods. The human expert’s similarity assessments of the music clips collected from the TagATune game were used as a benchmark (ground truth) in evaluating the results of NLP and audio methods. As a result of the research, it was found that the musical similarity extracted with NLP methods alone is not accurate to use in music recommendation, audio methods showed better results. By combining the best-performing methods of both approaches, the most accurate result where obtained.
Graduation Thesis language
Estonian
Graduation Thesis type
Master - Conversion Master in IT
Supervisor(s)
Anna Aljanaki (PhD)
Defence year
2022