Find the topics of lyrics in the top 100 songs and analyze how they change over time by clustering the topics of each song and visualizing the clustering results.
conda create -n nlp_project7 python=3.9
conda activate nlp_project7
pip install -r requirements.txt
- Genius API key: https://docs.genius.com/
- Spotify API key: https://developer.spotify.com/documentation/web-api
- How to apply for Spotify API keys: https://medium.com/@maxtingle/getting-started-with-spotifys-api-spotipy-197c3dc6353b
In the demo you can try the functions by changing address to a specific year, or the ID of a Spotify playlist.
Datasets are gathered from the APIs. Some playlists and year hot 100 songs are already generated and checked(manually check if the lyrics are right) in dataset folder.
The success of a pop song depends on many factors. Although the music charts are constantly changing, the keywords and underlying meanings included in the songs are limited, such as emotions, social issues, self and cultural identity, etc. We want to design a model that analyzes lyric themes to classify the theme of popular songs. We are curious whether the distribution of these topics in the list is immutable or changes in some way.
- Get the hottest 100 songs for a year and their lyrics(optional: a playlist from user input)
- Get a topic for each song (TF-IDF, KeyBERT)
- Cluster the topics for 100 songs this year (word2vec, k-means)
- Visualize the clustering results of each year (PCA, plot)
C_umass: a combined metric of combines document co-occurrence counts, one-preceding segmentation, and logarithmic conditional probability (Mifrah and Benlahmar, 2020). Cosine Similarity: a score based on the sentence embeddings generated from the ’all-mpnet-base-v2’ model, employing the SBERT SentenceTransformers library (Reimers and Gurevych, 2019) to encode both the title and relevant keywords into their respective embeddings.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent