Lyric Emotion Estimation Using Word Embedding Learned from Lyric Corpus
Kazuyuki Matsumoto, Manabu Sasayama · 2018
In searching huge music databases, users typically use “title,” “artist” or “song writer's name” as their search query. When users have only vague search conditions, searching for a song using affective information - that is, emotional impressions of the piece such as “relief” or “sorrow” - can be an appealing possibility. For cases in which the song's audio information consists of vocal and instrumental performances, both words and music reflect the writer's thoughts and emotions and encourage the imagination of the listeners. While a number of studies have used audio and lyric features in the emotion estimation of songs, many of the studies that use only lyrics have been based on existing textbook emotion estimation methods. Because the way in which words used in song lyrics tends to differ from the way they are used in text sentences, an emotion estimation method specific to song lyrics would be more appropriate. In this study, we extract features from a lyric corpus to handle lyric-specific expressions and propose a method to estimate emotion from each lyric phrase. The results of our evaluation experiment confirmed the effectiveness of the proposed method (lyric corpus word vector + 3-NN), as it produced a higher F-measure than the other methods.