TODO: Add YAML tags here.
name: TRBLLmaker
annotations_creators: found
language_creators: found
languages: en-US
licenses: Genius-Ventura-Toker
multilinguality: monolingual
source_datasets: original
task_categories: sequence-modeling
task_ids: sequence-modeling-seq2seq_generate
Dataset Card for TRBLLmaker Dataset
Table of Contents
- Table of Contents
- Dataset Description
- Dataset Structure
- Dataset Creation
- Considerations for Using the Data
- Additional Information
Dataset Description
- Repository: https://github.com/venturamor/TRBLLmaker-NLP
- Paper: in git
Dataset Summary
TRBLLmaker - To Read Between Lyrics Lines. Dataset used in order to train a model to get as an input - several lines of song's lyrics and generate optional interpretation / meaning of them or use the songs' metdata for various tasks such as classification.
This dataset is based on 'Genius' website's data, which contains global collection of songs lyrics and provides annotations and interpretations to songs lyrics and additional music knowledge. We used 'Genius' API, created private client and extracted the relevant raw data from Genius servers.
We extracted the songs by the most popular songs in each genre - pop, rap, rock, country and r&b. Afterwards, we created a varied pool of 150 artists that associated with different music styles and periods, and extracted maximum of 100 samples from each. We combined all the data, without repetitions, into one final database. After preforming a cleaning of non-English lyrics, we got our final corpus that contains 8,808 different songs with over all of 60,630 samples, while each sample is a specific sentence from the song's lyrics and its top rated annotation.
Supported Tasks and Leaderboards
Seq2Seq
Languages
[En] - English
Dataset Structure
Data Fields
We stored each sample in a 'SongInfo' structure with the following attributes: title, genre, annotations and song's meta data. The meta data contains the artist's name, song id in the server, lyrics and statistics such page views.
Data Splits
- songs
- samples
train [0.64 (0.8 * 0.8)], test[0.2], validation [0.16 (0.8 * 0.2)]
Dataset Creation
Source Data
Genius - https://genius.com/
Annotations
Who are the annotators?
top-ranked annotations by users in Genoius websites / Official Genius annotations
Considerations for Using the Data
Social Impact of Dataset
We are excited about the future of applying attention-based models on task such as meaning generation. We hope this dataset will encourage more NLP researchers to improve the way we understand and enjoy songs, since achieving artistic comprehension is another step that progress us to the goal of robust AI.
Other Known Limitations
The artists list can be found here.
Additional Information
Dataset Curators
This Dataset created by Mor Ventura and Michael Toker.
Licensing Information
All source of data belongs to Genius.
Contributions
Thanks to @venturamor, @tokeron for adding this dataset.