---
dataset_info:
  features:
  - name: token
    dtype: string
  - name: content
    dtype: string
  - name: vector
    sequence: float32
  splits:
  - name: tag
    num_bytes: 2740380
    num_examples: 2227
  - name: artist
    num_bytes: 46025354
    num_examples: 37002
  - name: track
    num_bytes: 898952880
    num_examples: 697812
  download_size: 1387722409
  dataset_size: 947718614
configs:
- config_name: default
  data_files:
  - split: tag
    path: data/tag-*
  - split: artist
    path: data/artist-*
  - split: track
    path: data/track-*
tags:
- music
---


# Musical Word Embedding

> [**Musical Word Embedding for Music Tagging and Retrieval**](https://arxiv.org/abs/2404.13569)

> SeungHeon Doh, Jongpil Lee, Dasaem Jeong, Juhan Nam  
> To appear IEEE Transactions on Audio, Speech and Language Processing (submitted) 


<p align = "center">
<img src = "https://i.imgur.com/Yw4UPnM.png">
</p>

Word embedding has become an essential means for text-based information retrieval. Typically, word embeddings are learned from large quantities of general and unstructured text data. However, in the domain of music, the word embedding may have difficulty understanding musical contexts or recognizing music-related entities like artists and tracks. To address this issue, we propose a new approach called Musical Word Embedding (MWE), which involves learning from various types of texts, including both everyday and music-related vocabulary.

### Resources: Using Musical Word Embedding

- [Pre-trained Embedding Vector](https://huggingface.co/datasets/seungheondoh/musical-word-embedding) 
- [Paper](https://arxiv.org/abs/2404.13569) 
- [Blog](https://seungheondoh.github.io/musical_word_embedding_demo/)
- [**notebook**-query_recommendation](https://github.com/seungheondoh/musical-word-embedding/blob/main/notebook/query_recommendation.ipynb)
- [**notebook**-music_retrieval](https://github.com/seungheondoh/musical-word-embedding/blob/main/notebook/music_retrieval.ipynb)


### Run the download script for embedding vector:

Check our huggingface dataset: You can download important embedding vectors such as tag, artist, and track from the Hugging Face dataset.

```python
from datasets import load_dataset
dataset = load_dataset("seungheondoh/musical-word-embedding")
```

```
{
  "token": "happy", 
  "content": "happy",
  "vector": [0.011484057642519474, -0.07818693667650223, -0.02778349258005619, 0.052311971783638, -0.1324823945760727, 0.03757447376847267, 0.007125925272703171, ...]
  },{
  "token": "ARYZTJS1187B98C555", 
  "content": "Faster Pussycat",
  "vector": [-0.13004058599472046, -1.3509420156478882, -0.3012666404247284, -0.34076201915740967, -0.8142353296279907, 0.3902665972709656, -0.1903497576713562, 0.6163021922111511, ...]
}
```

For other general 10M word vectors, you can also download them using the script below.
```
bash scripts/download.sh
```

### Citation
If you find this work useful, please cite it as:

```
@article{doh2024musical,
  title={Musical Word Embedding for Music Tagging and Retrieval},
  author={Doh, SeungHeon and Lee, Jongpil and Jeong, Dasaem and Nam, Juhan},
  journal={update_soon},
  year={2024}
}

@inproceedings{doh2021million,
  title={Million song search: Web interface for semantic music search using musical word embedding},
  author={Doh, S and Lee, Jongpil and Nam, Juhan},
  booktitle={International Society for Music Information Retrieval Conference, ISMIR},
  year={2021}
}

@article{doh2020musical,
  title={Musical word embedding: Bridging the gap between listening contexts and music},
  author={Doh, Seungheon and Lee, Jongpil and Park, Tae Hong and Nam, Juhan},
  journal={arXiv preprint arXiv:2008.01190},
  year={2020}
}
```

Feel free to reach out for any questions or feedback!