# Sentiment analysis and topic modeling in pop artists

## | Complex Systems | Data Science | Music Evolution |

In Natural Language Processing (NLP), sentiment analysis is a methodology to identify and study affective states in language (text, audio) data. It is usually helpful to analyze ratings or reviews of products but it can be applied to other types of analysis including comments on forums, tweets, etc.

In this particular case we are going to explore song lyrics, and we are going to try to go further and see if we can identify some of the topics associated with particular sentiments.

### How does sentiment analysis work?

The basic idea of sentiment analysis is to assign a label or ranking of a sentiment to a piece of text. The most simple case is to label the text as positive or negative, some approaches to this problem are based on a sentiment lexicon, a list of lexical features (words) labeled as positive or negative:

CategoryExample
Positive emotionLove, nice, good, great

Depending on how many good or bad words and taking into account their context in the sample, we can assign a ranking or number to the sentiment. Building an effective sentiment analyzer sounds (and is) complicated, fortunately for us this is a widely explored problem with different approaches in computational linguistics.

We are going to use the Valence Aware Dictionary for sEntiment Reasoning (VADER), a rule-based model that supposedly outperforms individual human raters. We are going to use Python in this case, but VADER has been implemented for different languages.

First, we are going to load the VADER model that is included in NLTK (Natural Language Tool Kit):

#for DataFrames
import pandas as pd

#better array management
import numpy as np

#Natural language toolkit
import nltk
#sentiment analyzer
sid = SentimentIntensityAnalyzer()

Now we can try out the analyzer with an example:

#computing polarity scores
sid.polarity_scores("The last season of game of thrones is the worst of the series")
{'neg': 0.255, 'neu': 0.745, 'pos': 0.0, 'compound': -0.6249}

The analyzer returns a dictionary with four different quantities: negative, neutral, positive and compound. The first tree are related to the sentiments and the fourth one is a compound score that takes into account all the scores and normalizes the value to be between $[-1,1]$. As we can see, the statement "The last season of game of thrones is the worst of the series" is accurately described as negative with a compound score of $-0.625$.

Now we need the songs from the artists we are going to analyze, I have downloaded this Kaggle dataset of lyrics from several pop artists and saved it in a folder named Data/KaggleDS/ and load them as dataframes with pandas:

import os
name_dir = 'Data/KaggleDS/' #directory where the data is stored
name_artists = []
df_artists = []
for file in os.listdir(name_dir):
if file.endswith('.csv'):
df_artists.append(clean_df(df)) #saving the clean dataframes
name_artists.append(file.split('.')[0]) #saving the names

The function clean_df() is used to clean the data, you can see more details in the notebook for this code.

Now we check the artists and the number of songs for each artist

n_songs = 0
for i, artist in enumerate(name_artists):
art_songs = df_artists[i].shape[0]
print(f'{artist}, {art_songs} lyrics.')
n_songs += art_songs

print(f'\n\n Total: {n_songs} songs from {len(name_artists)} Artists')
TaylorSwift, 256 lyrics.
Drake, 307 lyrics.
NickiMinaj, 174 lyrics.
ColdPlay, 159 lyrics.
PostMalone, 95 lyrics.
SelenaGomez, 85 lyrics.
CardiB, 53 lyrics.
CharliePuth, 49 lyrics.
Rihanna, 215 lyrics.
Maroon5, 97 lyrics.
JustinBieber, 158 lyrics.
Eminem, 366 lyrics.
KatyPerry, 131 lyrics.
EdSheeran, 126 lyrics.
BillieEilish, 52 lyrics.
DuaLipa, 85 lyrics.
ArianaGrande, 164 lyrics.
Khalid, 48 lyrics.
BTS, 233 lyrics.
Beyonce, 163 lyrics.

Total: 3172 songs from 21 Artists

and we use the analyzer to compute the compound score and assign it to a new column sentiment_score in the dataframe

for df in df_artists:
df['sentiment_score'] = df['Lyric'].apply(lambda song: sid.polarity_scores(song)['compound'])

Now we can explore, for example how a specific artist changes their sentiment over time

import matplotlib.pyplot as plt
import seaborn as sns

plt.figure(figsize=(12, 6))
sns.boxplot(x='Year', y='sentiment_score',data=df_artists[3])

the first thing we can notice about the plot is that the sentiment values seem to be very disperse in some of the years, this could be due to few amount of data or the shape of the distribution itself.

Next, we are going to plot the whole distribution (independent of the year) for every artist, and this time we are going to visualize it as a violin plot to have a better idea about the shape of the distributions

#concatenating all the artists dataframes
all_df = pd.concat(df_artists)

plt.figure(figsize=(18, 6))
plt.title('Sentiment score per artist')
#plotting all artists sentiment score distributions and visualizing as violinplots
sns.violinplot(x='Artist', y='sentiment_score',data=all_df)
plt.ylabel('Sentiment score (Compound)')
plt.xlabel('')
plt.xticks(rotation=45)
plt.show()

this plot confirms our suspicion about the shape of the distributions: almost every artist distribution shows bimodality. In this case the bimodality is represented by the two different sentiments indicating that there are few songs across artists that contain a neutral sentiment. This result makes sense, since pop lyrical music usually has strong emotional connotation, with the majority of artists in this sample being positive. The only 3 artists with a larger number of negative songs are Eminem, Nicki Minaj and Cardi B, with Eminem being the most negative one. This result is not surprising since Rap (or hip hop) music is well known to talk about social issues and often has violent or explicit content in the lyrics.

### Topic modeling

Now, let's have more fun, let's try to see if there is any relationship between the sentiment and the content (topic) the lyrics have. In other words, we want to see if their positive or negative sentiments are represented by the same or similar words the artist uses. To explore this question we are going to do topic modeling on each artist, so we can assign each of the songs to a particular topic and explore the most used words on each topic.

For the topic model we use the non-negative matrix factorization (NNMF) method (I will make another entry with details about other topic models). For this particular case, NNMF is more than enough and is easier to understand if you are familiar with linear algebra.
The basic idea behind NNMF is to factorize an initial matrix $\boldsymbol{V}_{m\times n}$ into a product of two smaller matrices $\boldsymbol{W}_{m \times p} , \boldsymbol{H}_{p \times n}$ :

$\boldsymbol{V} \approx \boldsymbol{W} \times \boldsymbol{H}$

where we can think about what the matrices represent in our problem, the initial matrix would be some type of normalized frequency of each word across all songs ($\boldsymbol{V}_{m\times n}$), where $m$ is the number of words and $n$ the number of songs. In the product one matrix would be the words distributed through topics ($\boldsymbol{W}_{m \times p}$) and the other would be the songs distributed through topics ($\boldsymbol{H}_{p \times n}$) with $p$ being the number of topics, a parameter that we control.

First we are going to construct our frequency matrix or $\boldsymbol{V}$, for this we are going to use a vectorizer from sklearn that implements a term-frequency inverse document frequency (TF-IDF) normalization:

from sklearn.feature_extraction.text import TfidfVectorizer
#define the vectorizer
#removing stop words
#max_df = max documents frequency, min_df = minimum documents frequency
tfidf_vec = TfidfVectorizer(max_df=0.9, min_df=2, stop_words='english')

the max_df and min_dfare variables to consider words that appear in at most and at least those fractions or number of documents, next we build our frequency matrix, in this example Selena Gomez's one:

# artist on index 6 corresponds to Selena Gomez
freq_mat = tfidf_vec.fit_transform(df_artists[6]['Lyric'])

and now we define our NNMF model, we are going to set the number of components (topics) equal to 3 to see if the topics are related to the sentiments. This selection is arbitrary and is under the hypothesis that artists talk about specific topic(s) in a negative or positive way, there are different ways to do this analysis but for simplicity and illustrative purposes we are going to assume that they have only one topic for each sentiment.

from sklearn.decomposition import NMF

nmf_model = NMF(n_components=3)

now we fit the model to the data

nmf_model.fit(freq_mat)
NMF(n_components=3)

and now we can obtain the $\boldsymbol{H}$ matrix that contains the song and the weights for the song on each topic and go further by assigning the number of topic to each of the songs

#get the H matrix
topic_results = nmf_model.transform(freq_mat)
#finding where the largest weight is on every song
#and assigning the song to a topic in the original dataframe of the artist
df_artists[6]['Topic'] = topic_results.argmax(axis=1)

now to see what words are the most used on each topic, if we inspect the nmf_model.components_ entry of our model we will see the matrix $\boldsymbol{W}$ containing the weights for the words on each topic:

nmf_model.components_
array([[0.00800022, 0.00368562, 0.01332492, ..., 0.00355086, 0.0046122 ,
0.00167504],
[0.        , 0.        , 0.        , ..., 0.        , 0.01988694,
0.00160906],
[0.        , 0.        , 0.        , ..., 0.        , 0.00050211,
0.        ]])

with this information we are going to save in the data frame the first 10 most used words on each topic

top_words = []
for ix, topic in enumerate(nmf_model.components_):
twords = [tfidf_vec.get_feature_names_out()[i]  # topic i
for i in topic.argsort()[-10:]]
top_words.append(twords)

#saving the words in the artist df
df_artists[6]['top_words'] = df_artists[6]['Topic'].apply(lambda topic: top_words[topic])

and now we can plot the sentiment distributions for each topic

plt.figure(figsize=(10, 6))
plt.title(name_artists[6])
sns.violinplot(x='Topic', y='sentiment_score', data=df_artists[6])
plt.ylabel('Sentiment score (Compound)')
plt.xlabel('Topic Number')
plt.show()

and get the top 10 words for each topic

n_artist = 6
wordsptop = [df_artists[n_artist]
[df_artists[n_artist]['Topic'] == topic]['top_words'].iloc[0] for topic in range(3)
]
print(f'Topics for artist {name_artists[n_artist]} \n')
for i in range(3):
print(f'Top 10 words for topic {i}: {wordsptop[i]} \n')
Topics for artist SelenaGomez

Top 10 words for topic 0: ['selena', 'good', 'got', 'want', 'just', 'don', 'know', 'yeah', 'love', 'like']

Top 10 words for topic 1: ['gun', 'war', 'blow', 'youre', 'lies', 'head', 'ahead', 'kindness', 'kill', 'em']

Top 10 words for topic 2: ['doesn', 'think', 'need', 'baby', 'know', 'magic', 'believe', 'disappear', 'night', 'oh']

From this result we can inspect what are the words commonly used in positive sentiment songs, which topic 0 is the one with the most positive median value, words like 'love', 'good' and 'like' with 'selena' makes us think that here she is probably singing about love and relationships or things she likes. The negative sentiment songs seem to talk about 'war' and 'lies'.

Now we do the same for every artist and plot the results

def get_topics(df_artist):
tfidf_vec = TfidfVectorizer(max_df=0.9, min_df=2, stop_words='english')
freq_mat = tfidf_vec.fit_transform(df_artist['Lyric'])
nmf_model = NMF(n_components=3)
nmf_model.fit(freq_mat)
top_words = []

for ix, topic in enumerate(nmf_model.components_):
twords = [tfidf_vec.get_feature_names_out()[i]  # topic i
for i in topic.argsort()[-10:]]
top_words.append(twords)

# returns the vector of weights for each song on every topic
topic_results = nmf_model.transform(freq_mat)

df_artist['Topic'] = topic_results.argmax(axis=1)
df_artist['top_words'] = df_artist['Topic'].apply(lambda topic: top_words[topic])

return df_artist

for df in df_artists:
df = get_topics(df)

fig, axs = plt.subplots(3,7, figsize=(32,14))

for i,ax in enumerate(axs.flatten()):
sns.violinplot(ax=ax, x='Topic', y='sentiment_score', data=df_artists[i])
ax.set_title(name_artists[i])
ax.set_xlabel(' ')
ax.set_ylabel(' ')

and we can inspect the top words for any artist in the same fashion we did for Selena Gomez, for example for Billie Eilish:

n_artist = 15
wordsptop = [df_artists[n_artist]
[df_artists[n_artist]['Topic'] == topic]['top_words'].iloc[0] for topic in range(3)
]
print(f'Topics for artist {name_artists[n_artist]} \n')
for i in range(3):
print(f'Top 10 words for topic {i}: {wordsptop[i]} \n')
Topics for artist BillieEilish

Top 10 words for topic 0: ['bullshit', 'bad', 'wanna', 'leave', 'just', 'way', 'need', 'want', 'know', 'don']

Top 10 words for topic 1: ['help', 'say', 'home', 'gonna', 'know', 'just', 'lie', 'love', 'let', 'like']

Top 10 words for topic 2: ['sound', 'run', 'town', 'silence', 'favorite', 'make', 'em', 'crown', 'bow', 'watch']

where the negative topic would be represented by the words on topic 0, the positive is topic 1 and neutral topic is represented by topic 2.

### Conclusion:

Sentiment analysis and topic modeling are two of the most common techniques in natural language processing, here we have combined both to give some insights about the lyrics of some of the top pop artists in the last years, we found that some of the artist seem to sing consistently more positive (or negative) about specific subjects.

Remember to take a look to the notebook for this example if you want to see more details.

CC BY-SA 4.0 Alfredo González-Espinoza. Last modified: March 02, 2023. Website built with Franklin.jl and the Julia programming language.