Emotion Patterns in Music Playlists

Try our demo: http://data.doremus.org/emotion.

For more details read our Report or our project slides.

---

Music streaming services such as Spotify are revolutionizing the music world, enabling a transition from artist-created bundles of songs (CDs) to user-created playlists. Different logics may be applied in the generation of a playlist: they can contain songs of a similar genre (e.g. “Rock playlist”), fit to a particular occasion (e.g. “New year’s eve party”), to a particular context (e.g. “Gym”), to a particular mood (e.g. “Happy”) and so on.

The goal of the project is to unravel the emotion patterns underlying the sequences of songs in a playlist using automatic approaches of Emotion Detection on the lyrics. Emotion Detection is a novel and promising field of study of Natural Language Understanding, which is able to automatically infer what are the emotions expressed in a text.

In the context of this project we had to:

1. Select an appropriate playlist dataset among the ones available on the web
2. Collect the lyrics associated to each song through an automatic approach
3. Study the state-of-the-art in Emotion Detection
4. Extract the emotions present in each song using an Emotion Detection tool
5. Analyze the experimental results statistically to unravel emotion patterns in the playlists

Before running any script, please install dependencies

pip install -r requirements.txt

Emotion Classification

In order to provide emotion classification of songs lyrics we provided with this repository already trained models. We trained three different model, each one related to a different version of the dataset. Specifically we trained our model on the following input combinations:

• MoodyLyrics alone
• MoodyLyrics4Q alone
• MoodyLyrics + MoodyLyrics4Q

For more details about those models, please refer to what is written in the report for this repo contained in the Report folder.

Using the code in this repository, lyrics can be classified at three levels:

1. Lyric level: using the function classify(sid, artist, title) of the src/emoclassify.py script
2. Playlist level: using the function robust_classify(playlist_vect) from the src/playlist_classify.py script
3. Spotify's dataset slice level: using the functino classify_slice(slice_path) from the src/slice_classify.py script

For more information about those functions please refer to their respective scripts and to the report.

Silver Standard

In order to face the absence of a dataset of playlists labelled with emotion, a perfectly balanced silver standard dataset has been generated by manually picking 40 playlist from Spotify, chosen among the ones representing an emotion in the playlist title, e.g. "Sad songs" or "angry music".

Utility Scripts

lyrics_downloader.py

This script simply downloads the lyric of the songs contained in any CSV file formatted in a similar way to MoodyLyrics dataset files (Index,Artist,Song,Emotion). For a better understanding of the options wich can be used with this script, please run:

python3 lyrics_downloader.py -h

From the src folder. If you want to test it with the dataset provided in this repository, please run:

python3 lyrics_downloader.py -i datasets/moodylyrics_raw.csv -o LYRICS_DIR -s

The script will download the lyrics for the songs listed into the src/datasets/moodylyrics.csv into the src/LYRICS_DIR folder. Each file will follow this name convention: EMOTION_ARTIST_SONG-TITLE.

LyricsManager.py

Map Spotify track URIs to LyricsWiki, and download the lyrics in a folder.

• Download and unzip MoodyLyrics

  • Convert ml_raw.xlsx to ml_raw.csv (to be better documented)

• Download and unzip the MPD

• run it

    python src/dataset_parsers/LyricsManager.py -ml /Users/pasquale/Desktop/MoodyLyrics/ml_raw.csv -i /Users/pasquale/Desktop/mpd.v1/data/ -o ./tmp/
  

Demo

In order to run the demo web application you need to install the dependencies:

    pip install -r requirements.txt
    python -m spacy download en_core_web_lg

After that, the web app server can be executed by running:

    FLASK_APP=index.py flask run