Have you ever accidentally stumbled upon a song whose lyrics mirrored exactly your situation at that moment. The serendipity of the occasion heightens the experience and often, one feels and connects more deeply with the song. This is not a co-incidence though. The best song-writers infact aim to write songs which will apply very specifically to your situation and yet be vague enough to apply equally well to a completely different situation. A bit like fortune-cookies. Bob Dylan was a master of this art. Dont believe me? Check this out.
A legitimate question to ask would be, can I create an algorithm which could identify songs which might fit my current situation. This task is incredibly difficult due to 2 reasons.
- Even with human intelligence, one person might not be able to completely understand the various meanings of a song. The realisation might come only after rehearing the song in the right circumstances. In short, song meanings are basically open to interpretation and very subjective.
- In the case when the song’s meaning is clear to a human mind, linguistically the meaning could be tough to identify due to non-standard construction of phrases and reliance on rhetorics. This is explained in detail in this excellent post analysing the ABBA song ‘The day before you came’.
This should not stop us however in this endeavour. If we can collaboratively understand the universe (ref: Wikipedia), what are a few songs? Some of you might be familiar with the site songmeanings, where one can find interpretation for popular songs of the contributors, with a voting system to build consensus. While these interpretations vary highly in size, grammatical structuring, accuracy of the opinion etc., it nonetheless opens up possibilities to understand the lyrical contents of a song by collective human intelligence. Social tags for songs generated by the ‘crowd’ has already been used to classify song topics, for ex. . However tags are usually too brief for deeper analysis and are probably apt only for topic detection
I need to come clean though. Using Songmeanings to tag songs is not my idea. I got it from , where they run clustering algorithms on song lyrics for topic identification. They do not however use the interpretation provided by listeners on the forum. My claim is, we can do better in understanding the songs if we find right ways of analysing user-provided interpretations.
But thats just one side of the story: ‘Understanding’ the song, in the best possible way we can. The other side is identifying the current life-situation of the listener. This could be achieved through two ways. Either the user explicits his current state (by choosing from a list/ entering a phrase summarising his situation), or it has to be deduced based on previous songs the user has heard. This gives us a lot of possibilities for linking up the two sides.
Explicit declaration by the user could be used to expand the search by including related keywords. It needs a suitable training set for this expansion, which I am currently unable to zero in to. One could then use supervised learning algorithms for training the algorithm to map the expanded ‘query’ to a song which connects with it. In addition, one can of course improve the entire system by using reinforcement learning algorithms for dynamic adaptation of the recommendations.
I find this an interesting exercise, and in any case, lyrics have always been an integral component of a song. Any information that can be extracted from it will only help make better recommendations. More importantly, creation of pertinent tags representing the ‘real’ song meanings is definitely worthwhile, even for other purposes. As a first step, I would like to create a small algorithm which would suggest a Bob Dylan/Beatles song corresponding to your life situation. There is always one.
 Hu, Xiao. Improving music mood classification using lyrics, audio and social tags. Diss. University of Arizona, 2010.
 Lukic, Alen. A Comparison of Topic Modeling Approaches for a Comprehensive Corpus of Song Lyrics. Carnegie Mellon University, 2015