From Lyrics to Their Interpretations: Automated Reading between the Lines

Although musical metadata, such as genre, mood, and usage, have been widely studied, there is a less explored type of metadata, the subject of the lyric, which might be potentially useful in music digital libraries. In my previous research, I have explored the use of the subject metadata in the context of music classification based on it. The biggest challenge of the classification task is that lyrics are often hard to analyze automatically due to their poetic nature. What I discovered is that user-created interpretations such as those readily available on songmeanings.com are more suitable for this kind of task than the lyric text itself. A subsequent project was to apply Latent Dirichlet Allocation (LDA) topic modeling on this collection of interpretations, and then to automatically discover the underlying subjects. This unsupervised topic modeling technique is especially meaningful, because human annotation of the subject categories per each song can be difficult and expensive, while the interpretation can be seen as a verbal explanation of the subject categories. I proposed a topic refinement system that uses some intrinsic topic evaluation techniques. The system is based on Normalized Pointwise Mutual Information (NPMI) and topic weights that are available as the results of LDA, followed by an extrinsic evaluation method as well to demonstrate the quality of refinement. The next phase of this study is to learn the relationship between the lyric and the interpretation. Although I found the interpretation dataset to be especially useful for the subject discovery and prediction, it is true that the interpretation of lyrics has certain limitations such as its reliance on volunteer work and its scarcity. I believe that an elaborately designed intelligent system should be able to predict the subject from nothing but the lyric texts. However, the relationship between the lyrics and existing interpretations requires a development of an intelligent mapping technique. I plan to learn the mapping by using LDA as a dimensionality reduction technique, and then learning a regression model between the dimension-reduced lyrics and interpretations, which is a universal way to learn all the relationships that I need.