I’ve been an Elliott Smith kick this past month. Although I was one of the lucky ones to see him live (back in 1998 at the now-defunct Trocadero in Philly), his music really didn’t click for me until I was in my mid-30s. My key hypothesis: It’s not just sad bastard music. Smith had some dark themes, yes, but there was a lot of light that went through his work. But it was all kind of poetic with some intricate wordplay, so would sentiment analysis be able to pick up on the lyrics to something like this?
Drink up, baby, look at the stars
I’ll kiss you again, between the bars
Where I’m seeing you there, with your hands in the air
Waiting to finally be caughtBetween the Bars
First I pulled all the lyrics off of Genius using the genius R package. This included the posthumous compilation New Moon. According to Wikipedia, the songs were recorded between 1994 and 1997, so I just decided to assign them all a year value of 1996. I wanted to make sure I included this mainly because of my love of “Whatever (folk song in #C).”
To do the actual sentiment analysis, I compared two lexicons : bing and the newish vader. Let’s start with Bing. After all the standard pre-processing, I used a similar strategy to what I did with IHI’s Stories of COVID. The Bing lexicon assigns words to a positive or negative sentiment (Liu, 2015). I computed a net sentiment scale to identify the overall positive or negative content of a song. I then broke sentiment down by song and album.
So, looking across these, we can see some of the gaps in understanding. Behind the Bars shows up as positive, as does Twilight. But so does Waltz No. 2 (XO) so that’s good.
Now, let’s look at what vader came up with. The Valence Aware Dictionary for sEntiment Reasoning (Hutto and Gilbert, 2014) was developed specific for micro-blogging situations, like twitter. Would it perform well for lyrics? Let’s check it out.
Well, it’s clear that the metaphors in songs like Twilight and King’s Crossing are buried pretty deep, but otherwise this actually kind of looks okay.
Finally, I looked at sentiment compared over time. I had to scale both the bing and vader scores to make them more directly comparable, and whattayouknow? they were mostly in alignment. Actually, a lot better in alignment than I originally thought looking song by song. The biggest gap in in Either/Or, but it that doesn’t seem that major or meaningful.
So what, if anything does this have to do with community-based evaluation. Well, I’m glad you/I asked. I’ve gotten really interested in how we can synthesize community and cultural artifacts to learn about community-level conditions. I had a professor way back in undergrad, Dr. Steven Krauss who said something that stayed with me: to understand a culture, you have to be attuned to both the “high” and “low” aspects. I’m not quite sure if and how lyrics can be a window to that because you’d have to link things by geographic and timing. The simple nature of the recording process preventions that data from being super timely, but it could potentially give some insight into the general community of scene vibe. Like, if you were interested in Philly at the turn of the century, would The Roots and Man Man give any insight? Maybe?
But back to Elliott. Just on an individual-level it is interesting the sentiment parallels what we know of the trajectory of his life. And, while it’s really on the nose, this is great song:
Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text. Eighth International Conference on Weblogs and Social Media (ICWSM-14). Ann Arbor, MI, June 2014.
Liu, B. (2015). Sentiment Analysis: mining sentiments, opinions, and emotions. Cambridge University Press