Music is an extremely important and prevalent part of one’s life. According to the Nielsen Music 360 Report, in 2019, the average United States consumer spent 26.9 hours listening to music per week (Department). This translates to over 3.8 hours of music listened to per day by the average consumer. This is a large amount of space every day dedicated to a hobby which, excluding instrumental genres such as classical music and EDM, essentially involves someone speaking to us. The words this person says varies; some people prefer to listen to positive music with uplifting lyrics, some prefer catchy songs with meaningless lyrics, some prefer songs with more depressing tones and sadder lyrics. Regardless of what message one prefers, the lyrics still are being heard and absorbed into the brain, and one might be interested in figuring out how positive or negative these lyrics are.
The classic saying “you are what you eat” essentially means that your personal health can be quite easily controlled, to an extent, by what foods you decide to eat. It’s often told to kids who are quick to exchange healthier foods for foods higher in sugar content and fat in their diet. I believe however that another similar expression can be made: you are what you listen to. Because we take in these lyrics that we listen to, one could argue that it would be good to take an evaluation of the emotions portrayed and expressed in the songs one spends large amounts of time listening to.
I decided to take a deeper look into the music I listen to, with the goal of viewing how positive or negative the emotions expressed in my most-listened to songs are. Music has had a large influence in my life, with my parents even warning me at an earlier age that negative music would lead me down a darker path. I decided to attempt to figure out exactly how negative the music I listened to was, from a completely unbiased perspective. Studying this is important for several reasons.
For one, music has taken a darker tone recently. The emotions of “anger, disgust, fear, sadness, and conscientiousness have increased significantly, while joy, confidence, and openness expressed in pop song lyrics have declined” (Napier). I would consider myself to be a fairly average music-listener, and thus it would be interesting to see whether the music I listen to really reflects that cultural shift; whether there is truly more negative than positive.
Another reason this study is important is because the emotions that music expresses is a very “basic feature of Western music” (Fritz). In order to truly understand the music one listens to, one has to understand the basic components of it, and emotion just happens to be one of those components. Perhaps through this study I could find a greater understanding and appreciation of the music that I already listen to. Or perhaps I could gain new insight provided on how the language used leans more positively or negatively.
The final and most important reason why I would like to dive deeper into the emotions portrayed in my music is backed by scientific research. Music can be viewed as an art form; a form of expressing emotions, but music can also be viewed as a tool for managing emotions. Music has been found to regulate emotions among adolescents, both males and females (Alan Wells). There is a potential problem with this, however.
If an individual is finding that only negative music will give them any sort of joy, this can cause them to rely heavily on songs with lyrics that may put them in a worse mental position. It has been discovered that activation of any nodes in the brain, apart from those in the displeasure center, will cause the individual to sense pleasure (Schubert). Listeners experiencing a difficult season in life might become attracted to darker and more negative music, and since they can still sense pleasure from it, they might eventually replace all of their music with more negative tracks. This is a problem not only because this means there are less positive influences in the individual's life, but also because negative music has been shown to possibly worsen one’s mood (Stewart). This could perhaps create a downward spiral, where the individual’s mood worsens because of life and it’s challenges, and then the individual turns to negative music for a sense of comfort. This negative music, however, points the individual in a darker direction and then worsens the individual's mood, resulting in them relying on this music, which in turn worsens their mood further. If I find out that there’s more negativity expressed than positivity in the songs I listen to, perhaps it’s time to make a change. Possibly introducing new, positive music would help.
This project is one that would not per say reveal important information to others, and wouldn’t necessarily show others how to live their lives. This is one that is reflective for me, that shows me how my music listening habits might have to change. The main tool I’m using for this project, VADER sentiment analysis, views things purely from an objective perspective. Instead of interpreting sentences in their full context, which would be a more human way of sentiment analysis, the tool analyzes the individual words in a sentence, determining whether the words lean in a more positive or negative direction. Therefore this is a process that is purely objective, and reveals a very different side of sentiment analysis that could not be done easily or quickly without the use of programming.
If one were to perform sentiment analysis on these songs manually, one might skip over and fail to understand the meaning of negativity. While a phrase such as “I’m killing it” might sound positive to us as humans, the software would interpret this as negative, as this uses the word “kill”. Because of this individual word-based approach, we get deeper insight into the pure negativity and positivity, regardless of context or overall meaning. This is advantageous, as it might be possible that negative words in a positive sentence could still affect us negatively.
Of course this could also be a disadvantage to the strength of these results as context is very important. The reflection of negativity in this study is accurate based on how the listener listens to a song. If the listener listens passively, hearing the words but not understanding the meaning, the measures of negativity and positivity applies. If the listener actively listens and tries to understand the meaning, the measures of positivity and negativity do not apply as much. That being said, whatever style the listener would listen in, the measures still apply to an extent.
For this project I decided to take data from Apple and their privacy section from Apple’s website. I chose this data source not only because it was extensive, but also because Apple has a habit of setting the industry standards for accuracy and consistency. As the data I wanted to collect would be statistics on my Apple Music account, the most reliable source would logically be the provider of the service themselves, Apple. The other main reason I decided to collect this data from Apple was because they are the exclusive provider of said data, thanks to their extremely rigid policies on user privacy.
This data was collected from the 16th of August 2020, the date I subscribed to Apple Music, to the 4th of February 2021, the date I requested my data from Apple. The data is extremely extensive, covering not only the exact time the song was started and ended, but also the reason the song was ended, where the song was played from, the track name and artist, and many others which altogether totalled 30 different variables. The most important variable for this study that was recorded is the play duration of the song in milliseconds, which displayed how long the song was played for. There were no missing cases, which is impressive for a dataset with over 35621 different observations. The data was generated as I selected songs to be played, which allowed the data to be highly accurate.
The first step to conducting analyses on this dataset was to look through the data. Understanding which variables needed to be focused on and manipulated was confusing, as there are two very similar variables: play duration in milliseconds, and media duration in milliseconds. The former is a measure of how long I played the content for during that instance of playing the song, and the latter is how long the song actually is.
I imported pandas, matplotlib, and matplotlib pyplot for the data. I also had to install VADER sentiment analyzer through pip, and then imported the sentiment intensity analyzer through VADER sentiment analyzer.
The first thing I had to do in order to perform sentiment analysis on the data was use .groupby for the variable “content name”. I would then sum the play duration in milliseconds, in order to collect the total millisecond duration of the time that each song was played. I then sorted this data by descending values, which ranked all of the songs I have played on my Apple Music account by how long I’ve played them for. This allowed me to gather my top 5 most played songs and perform sentiment analysis on them.
I then found the lyrics for each song through Musixmatch, and wrote them as a list, separating each line with a comma meaning that each line would be analyzed individually through VADER sentiment analysis. I created a dataframe to put data in for what percentage of the songs was positive, negative, and neutral.
I wrote a for loop which ran the sentiment intensity analyzer for every line of each song. This would output three different values, “pos”, “neg”, and “compound”. I decided to go by the compound VADER sentiment intensity analyzer score of each of the lines to determine what sort of emotion they expressed. This compound score is a score between -1 and 1 which summarizes how positive or negative the sentence analyzed is, the lower scores being negative, the higher scores being positive, and the scores near zero being neutral. I decided to use the same threshold as the github page for the tool (Cjhutto). This threshold dictated that sentences were determined to be positive if the compound score was greater than or equal to 0.05, negative if the compound score was lesser than or equal to -0.05, and neutral otherwise. This also means that entire sentences were either counted as neutral, negative, or positive, so even though certain sentences could contain both positive and negative sentiments, whichever was the greatest level of emotion was the one that the sentence was labeled as.
All of the scores were converted to percentages, and then entered into the dataframe. This dataframe was then plotted as a matplotlib bar graph, with each bar representing the makeup of the sentences of a song. These bars almost functioned as a pie chart where each bar was 100 units tall, each unit representing one percent. The bar chart was then divided into three sections: one representing what percentage of the song’s lines were positive, one representing what percentage were neutral, and one representing what percentage were negative. This presented the data in an easy-to-understand manner.
The results for some songs were anticipated. I expected the songs by the late Pop Smoke, an artist who was known for creating drill rap (a type of rap which is focused on appearing intimidating and threatening), to rank more negatively, and in fact Welcome to the Party was the most negative out of the top five songs. Over 45 percent of the 59 lines leaned negatively, which is not surprising given the highly explicit language used and the general subject matter. Given that Pop Smoke was part of a gang, hearing that he shoots “one in the head” with “ten in the clip” is not something that is immediately shocking to hear from him, but one can also get lost in his extremely gruff voice which can sometimes distract from the extreme subject matter. I wasn’t even sure of what the lyrics were until they were laid out in front of me, and once I saw them it was easy to see that this song would be seen by the sentiment analysis tool as very negative.
The other song by Pop Smoke which cracked my top 5 most played songs, Hello, while not nearly as aggressive as Welcome to the Party, still displayed themes of objectification and references murder on multiple occasions. Once again, I suspected that this would have high negative scores, which it did (as around 34% of the lines were negative). The other results were more interesting and unexpected, however.
The song with the highest percentage of positive lines was Soul Food II by Logic. Given that Logic’s motto is “peace, love, and positivity”, this wasn’t too much of a shock but the real shock was that it was only really positive in comparison to the other songs I studied. Compared to other songs outside of my top 5, or in comparison to other music, this would be considered a fairly negative song. For Logic being an artist that prides himself on his friendly and fairly positive outlook on life, there was still a higher percentage of negative lines than positive lines in Soul Food II, and while the difference was only around 2 percent, this still strikes me as somewhat strange. Granted, Soul Food II is not necessarily one of Logic’s most positive and encouraging songs, but it’s also not one that I would initially say is more negative than positive. The latter half of the lyrics is entirely dedicated to storytelling, and the first half is Logic dealing with his thoughts as he’s realizing his success as a rapper.
One would think that Soul Food II would score significantly higher positive scores than Lemonade, a song entirely dedicated to prescription cough syrup and bragging, but the scores are very similar for both. This was somewhat strange as well, that a song which has far less encouragement and inspirational lines had a very similar percentage of positive lines to Logic’s song.
Fast Lane is the outlier here, with more negative lines than positive, but more than 53% of the lines were neutral. One of the two rappers in Bad Meets Evil, the artists of the song, is Eminem who is notorious for including negative connotations within his lyrics, but this is not the most negative out of the five songs. Nor is it the most positive, this is simply the song with the most neutral lines.
This reveals that I essentially couldn’t accurately predict the scores of three out of the five songs. This is interesting for several reasons.
One of the reasons that I wasn’t able to predict the emotions of the music I listened to could be that my usage of the VADER Sentiment Analysis tool has its flaws. The system for collecting scores on songs was flawed, but also somewhat difficult to improve. This is because instead of the songs being analyzed for how negative the lines are, they’re simply being analyzed as to whether the line is more on the positive or negative side, given that I used the compound score. This means that an extremely negative line will have the same impact as a line that’s slightly negative. So a line like “even when I’m winning, feel like I’m losing” on Soul Food II would be treated as equally negative as “I catch a body, next day I forget it” from Welcome to the Party. Because of this, I misjudged just how negatively a song like Soul Food II could be interpreted.
Secondly, it shows that I as a listener don’t truly comprehend how positive or negative the music I listen to is. The purpose of this study was to see what emotions were expressed in the songs I listen to most. I wondered whether I would be able to accurately guess whether the songs I listen to would be positive or negative, but unfortunately I wasn’t able to do that. I can’t accurately predict from an objective standpoint whether the music I listen to is positive or negative. This truly shows an advantage that technology could have over the individual.
This study mainly revealed to me that I can’t make accurate judgements on the emotions expressed in my music, and that the music I listen to is significantly more negative than I would’ve expected. Rap music tends to get a bad reputation, but according to this study, the bad reputation is somewhat deserved. It was interesting to see what the emotions expressed were of the songs I studied, but it also led me to the realization that it’s something I might want to pay more attention to. I might need to consider changing my listening habits to an extent in order to protect myself from extreme negative influences.
This study also revealed to me that computers can’t solve and interpret everything. We as humans have created such complex languages that can’t be nearly as accurately interpreted and understood. There are flaws with the VADER sentiment analysis tool, and the way I used it was not nearly perfect. Changing the way I used it, however, would only solve half of the problem, and the root problem is the idea of using a computer to interpret language, especially lyrics. Lyrics aren’t meant to be read, they’re meant to be heard, and the only way they can truly be interpreted and understood is through listening to them the way the artist intended.
In conclusion, while I learned a lot from this study about the emotions expressed in the music I listen to, I also learned more unexpected things. I learned that the music I listen to is likely too negative, and that I should put more effort into listening to more positive songs. I also learned that the emotions expressed in the music I listen to reflects the more negative shift of the culture as a whole. Finally, I learned that emotions are difficult to be processed by a computer, which poses several questions to those in the artificial intelligence field. This study could never be made perfect, but it provided some insight regardless, and is something I could continue to improve in the future.
Works Cited
“APA PsycNet.” American Psychological Association , American Psychological Association, psycnet.apa.org/record/2011-05681-001.
Alan Wells, Ernest A. Hakanen. “The Emotional Use of Popular Music by Adolescents - Alan Wells, Ernest A. Hakanen, 1991.” SAGE Journals , journals.sagepub.com/doi/10.1177/107769909106800315.
Cjhutto. “Cjhutto/VaderSentiment.” GitHub , github.com/cjhutto/vaderSentiment.
Department, Published by Statista Research, and Jan 8. “Time Spent with Music in the United States 2019.” Statista , 8 Jan. 2021, www.statista.com/statistics/828195/time-spent-music/#:~:text=Data on the amount of,even higher at 32.1 hours.
Fritz, Thomas, et al. “Universal Recognition of Three Basic Emotions in Music.” Current Biology , Cell Press, 19 Mar. 2009, www.sciencedirect.com/science/article/pii/S0960982209008136.
Napier, Kathleen, and Lior Shamir. “Quantitative Sentiment Analysis of Lyrics in Popular Music.” University of California Press , University of California Press, 4 Dec. 2018, doi.org/10.1525/jpms.2018.300411.
Schubert, Emery. “Enjoyment of Negative Emotions in Music: An Associative Network Explanation - Emery Schubert, 1996.” SAGE Journals ,journals.sagepub.com/doi/10.1177/0305735696241003.