Advanced sentiment analysis of music analyzing the sentiment of lyrics and acoustic features.
Overview
Using the lyricsgenius API, searches for the song lyrics requested by the user via command line arguments and performs a sentiment analysis of those lyrics displaying the result.
Pre-processing
Using the Natural Language Toolkit (NLTK) the program first tokenizes the lyrics so pre-processing can begin. Then, stop words are filtered out of the tokenized lyrics to remove extra noise and unnecessary words that have little effect on overall sentiment. Next, each word is reduced to its base form via lemmatization. Finally, the lyrics are detokenized back into a string so sentiment analysis can begin.
Lexicon Based Sentiment Analysis
Using both NLTK's SentimentIntensityAnalyzer and textblob's sentiment polarity a sentiment analysis is performed on the pre-processed lyrics. An average is taken between the result of the textblob analysis and the SentimentIntensityAnalyzer. The average gives a good middle ground between the two analyzers as sometimes one provides a better assessment of sentiment depending on the lyrics.
Machine Learning Based Sentiment Analysis
Two models were created using the Yelp polarity reviews data set and the IMDB reviews data set from the provided data sets in Tensorflow. Both models used the same recurrent structure, see code for specific details. The "Yelp model" achieved 94.7% test accuracy after training while the "IMDB model" achieved 86% test accuracy. Despite the high accuracy achieved by the Yelp model and the moderately high accuracy achieved by the IMDB model both were inconsistent in determining the sentiment of lyrics during manual testing. Perhaps these are not the correct data sets to use for this problem. Movie or restaurant reviews are quite different from song lyrics. A data set containing lyrics and their associated sentiment may be more appropriate.
An attempt was made to create a model using the MER data set which contained lyrics associated with one of 4 quadrants from the Russell emotion model: Q1: Happy (can be excited or pleased) Q2: Tense Q3: Melancholy Q4 Serene Joy (peaceful, relaxed). However, the model achieved very low accuracy as well as a consistent loss of nan. Many different solutions were attempted to fix these issues to no avail. This model should be revisited in the future.
Notes:
After testing a plethora of songs, lyrical analysis is not always indicative of the general mood of the song. Though it gives a good baseline there are other factors impacting the sentiment of the song. Factors such as tempo, mode, loudness, and how reliant the song is on its lyrics will be also investigated.
The IMDB and Yelp models should be sufficient to continue research into the other factors impacting song sentiment however the MER model should be revisited to improve accuracy in the future.
Overview
Attempt to predict the sentiment of a song based on its average loudness and tempo.
Data Set
In order to associate these measures with sentiment, a dataset was compiled in order to study the relationship between song sentiment and tempo and loudness. Using Spotipy, a Python library for the Spotify API, a diverse dataset containing over 1,000 Spotify tracks from multiple genres was complied. Each track has various attributes including Spotify's assessment of the track's sentiment (valence), average loudness in dB, and tempo in BPM.
Data Analysis
The goal of this analysis was to find the average tempo/loudness of average songs, sad songs, and happy songs. The average tempo over all tracks in the dataset as well as the average loudness served as a middle ground for the sentiment prediction. Then, averages of the same measures were taken over tracks only above or below certain valences in order to assess the average tempo and loudness of happy and sad songs. A regression analysis was conducted on the resulting (valence, average tempo) and (valence, average loudness) pairs. The two regression equations obtained from this analysis each provide estimates of the sentiment of the song given its tempo or average loudness.
Notes:
Both regression lines extend beyond valence values of 1.0, however, for this application, the valence will be capped at 1.0 to keep a consistent 0-1 sentiment scale. This means that though a tempo of 160 may be projected to have a valence of 13.0 by the regression equation it will be represented as a 1.0 valence when being used in the final equation.
Overview
Something must tie together the various factors impacting song sentiment to be explored. An average of the lyrical and title sentiment between the IMDB model, Yelp model, Text blob, and NLTK's SentimentIntensityAnalyzer does not take into account the other factors contributing to a song's sentiment and must be expanded upon.
Equation Inputs: Mode, Lyric Sentiment, Title Sentiment, Loudness, Tempo
Equation Outputs: The Final Sentiment, a measure of how sad or happy the song is (0-1)
The equation is a weighted average of the sentiments predicted by each input:
Final Sentiment = ((Mode Weight * mode sentiment) +
(Text Weight * combined sentiment of the lyrics and title) +
(Loudness Weight * sentiment predicted by the loudness regression equation) +
(Tempo Weight * sentiment predicted by the tempo regression equation)) / Sum of the weights
The mode of the song is extracted using Spotipy and is represented as a 1 for major and 0 for minor. A major mode is usually associated with happier feelings while a minor often evokes a more somber mood. The combined sentiment of the lyrics and title is calculated by another weighted average of their sentiments. This average weighs the lyric sentiment by 0.75 and the title sentiment by 0.25. The loudness and tempo of the song are also extracted via Spotipy and used with the regression equations created in the tempo and loudness analysis to produce their predictions of the song's sentiment.
Equation Weights
The weights used in the final sentiment equation were taken from a study conducted in 2015 by Jamdar, Abraham, Khanna, and Dubey which proposed a method to detect the emotion of music using lyrical and audio features with a k-Nearest neighbors classifier. Jamdar et al. used weights for the features to better categorize the song into an emotion. These weights worked well for this analysis in the final sentiment equation as well.
Notes:
Though this equation proved to be slightly more accurate in initial testing than a purely lyrical analysis, the acoustic features of the song may vary widely depending on the genre of the song. This along with improved lyrical assessment will be investigated to increase the accuracy of this evolving system.
Inspiration: https://kvsingh.github.io/lyrics-sentiment-analysis.html
Guideline used for sentiment analysis: https://www.datacamp.com/community/tutorials/text-analytics-beginners-nltk
NLTK Vader: 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.
MER dataset: Ricardo Malheiro (2017). “Emotion-based Analysis and Classification of Music Lyrics“. Doctoral Program in Information Science and Technology. University of Coimbra.
TensorFlow Load Text/Text Classification tutorial: https://www.tensorflow.org/tutorials/load_data/text
Python library for Spotify API: https://github.com/plamere/spotipy
Regression calculator used: https://keisan.casio.com/exec/system/14059932387562
Study used for weights in final equation: https://arxiv.org/ftp/arxiv/papers/1506/1506.05012.pdf
Jamdar, A., Abraham, J., Khanna, K., & Dubey, R. (2015). Emotion Analysis of Songs Based on Lyrical
and Audio Features. International Journal of Artificial Intelligence & Applications, 6(3), 35-50.
doi:10.5121/ijaia.2015.6304