Music Sentiment Analysis

Assessing the affective quality of a piece of music is a difficult task, generally relying on expensive surveys to generate user annotations. Still, many such datasets exist doing just that, including AMG1608, PmEmo, and DEAM. In their 2018 paper, Deezer sought to automate this process, using NLP analysis of song lyrics and analysis of acoustic features to predict music valence and arousal. We look to expand on this work by using social media conversations to predict music affect.

In 2021, we presented preliminary findings at NCUR. Our initial approach at this learning task relied on parsing these comments using existing word affect and valence/arousal dictionaries like the work presented by Warriner et. al. and generating summary statistics based on the aggregate total of valence/arousal words in a comment to create a feature space. This yielded modest performance.

Future work will use pre-trained transformer models like BERT to perform end-to-end analysis and prediction of our comment data.

This work is ongoing in the SoundBendOR lab @ OSU-Cascades, and is advised by Dr. Patrick Donnelly.