Hi, I'm Jen, and I'm a Swiftie. This will not be a surprise to those of you that know me, but I am here today to confess that I, a grown woman, regularly listen to and enjoy Taylor Swift's music. People like music for different reasons; some appreciate the complexity of arrangements and harmonies, some enjoy the music associated with their cultural heritage, and some just like a good beat. I am a lyrics girl, and my favorite songs are more like poems set to background music. I remember in high school poring over my REM tapes trying to figure out what Michael Stipe was trying to say, both figuratively and literally (there was a lot of mumbling on the early albums, and back then I couldn't just Google the lyrics). Taylor Swift is, in my opinion, a pretty talented songwriter, and while her music itself is not very complex, her lyrics can be.
Part of working at QIC is bearing the shame associated with working for someone whose ringtone is a Taylor Swift song, so when one of them posted on our Slack an article describing taking a machine learning approach to analyzing her discography, he knew it was instant brownie points. He also correctly guessed I would take issue with it. What he did not realize is that it's a great example of why technologies such as machine learning and artificial intelligence still need a human in the loop.
There's a lot of discussion these days about outsourcing our cognition, and there seem to be two philosophical camps: those wildly enthusiastic about the promise of AI to make the world a better place (the "Can I please have a driverless car now, I am an idiot" people) and those terrified that when machines are smarter than we are, the world will lose its humanity (the "I saw 2001: A Space Odyessy/The Matrix/The Avengers: Age of Ultron/any movie with AI in it ever, this is the way the world ends" people). While I am solidly one of the former, the reality is that there is a long way to go before any of this technology is either the savior or destruction of our society. Case in point: Taylor Swift.
In this article, the author has undertaken the ambitious task of analyzing Taylor Swift's discography to find themes that speak to her development as a person and as an artist. She is a good candidate for this, as her songs are almost exclusively autobiographical, much to the consternation of other artists such as Trent Reznor. To accomplish this, he developed a model based on the frequency of groups of words in her songs, applying some techniques to filter out common words that are not particularly meaningful. From this process, he was able to identify six clusters of words:
Now, here comes the opposite of science. Based on these clusters of words, he assigned "topics" to them. This process is similar to what we behavioral scientists do when we analyze the results of an exploratory factor analysis: look at the things that "hang" together and try to make sense of it. When we do this, our decisions are typically based on an understanding of previous research about the constructs we're investigating and a theoretical understanding of cognition, personality, and other aspects of psychology. It's an educated guess. In the case of this analysis, the author clearly did not conduct his Taylor Swift literature review or SME interviews. Here's what he came up with, mapped by number of songs that reflect each "theme."
If the author had done his homework, he would know that like any good songwriter, Taylor Swift writes largely in metaphor, more so with each new album. This is relevant when we're discussing her personal development because she started writing songs at a very young age - her first album came out when she was fifteen. Over her career, she's progressed from "He's the reason for the teardrops on my guitar," which you can take literally, to "Loving him was like driving a new Maserati down a dead-end street" which you can't, to "Feeling so Gatsby for that whole year" which you can't even understand unless you get the literary reference. A big red flag here? The theme of "Dancing." According to these results, there are three songs about dancing on her latest album Reputation. Not that I can sing that whole thing by heart or anything, but I know for a fact there are zero songs about dancing on that album.
On the Red album, the word "dancing" can be taken at face value (e.g., "Keep dancing like we're 22"). On Reputation, the only song on the album with that word in it is "Dancing With Our Hands Tied," which is a metaphor for a relationship that she knows is inherently doomed by circumstances beyond her control. It's also brilliant. But more importantly, it's not something that a machine learning algorithm would be able to understand, because understanding metaphor is beyond its capacity. While the frequency of words may be a good indicator of themes if we're speaking literally, if the words don't mean the same thing all the time, the analysis falls apart.
I’m not saying machine learning is worthless. For one of our Navy efforts, we're incorporating language processing into AR and VR training for pilots learning commands to use when they're communicating with the tower. This works because what the pilot is supposed to say is highly scripted, and there's a consistently "right" answer. But until these technologies can understand the nuances of human language, you still need a person in the loop to understand the data. On the one hand, I'm not getting my driverless car any time soon, but on the other, I'm not going to be a captive energy source for machines in the foreseeable future either.
These posts are written or shared by QIC team members. We find this stuff interesting, exciting, and totally awesome! We hope you do too!