Aspect Based Sentiment Analysis in Music: a case study with Spotify

Nowadays, there are more and more social networks and Web platforms that give their users the opportunity to share their opinions and tastes on items of different types. This inevitably led to a growth of data relating to the subjective sphere of each individual. This information is extremely useful for several purposes, such as providing personalized recommendation services or understanding opinions conveyed through text. Sentiment Analysis provides helpful methods to analyze these textual opinions (e.g. reviews) from a global point of view. In case we want a more detailed representation of the opinion represented in a text, Aspect-based Sentiment Analysis identifies a valuable option thanks to its fine-grained level of text analysis. In this paper, we have designed a processing pipeline aimed to extracting domain-related aspects from text by means of an unsupervised approach. We formally define Aspect Terms and Aspect Categories as well as Aspect-based Sentiment Embedding, an approach of representing documents by computing aggregated sentiment scores for each aspect. We perform experimental evaluations on the Spotify dataset to prove the utility of our technique in predicting elements strictly related to emotions and feelings. Our results show improvements on the regression task for sentiment-related features compared to the classical semantic-based representations.