specific-AI/hebrew_ner_song_titles

Hebrew Song Titles NER Dataset

Overview

this is a snapshot of https://huggingface.co/datasets/NHLOCAL/SingNER

This dataset is designed for Named Entity Recognition (NER) tasks focused on identifying key entities within Hebrew song titles. The dataset has been built using real-world music databases and enriched with synthetic data generated by language models. The annotations were performed using the Gemini model to ensure high-quality entity labeling.

Dataset Description

The dataset consists of over 32,000 song titles, primarily in Hebrew, with entity annotations covering multiple categories:

• SINGER – Recognizing singer names.
• SONG – Identifying song titles.
• ALBUM – Detecting album names.
• MISC – Capturing other relevant named entities.
• GENRE – Identifying musical genres.

Example Record

{
  "text": "מרום שטיינברג אילי מנצל רון קירשנבוים - עברתי בחושך (קולולחן)",
  "entities": [
    {"start": 24, "end": 37, "label": "SINGER"},
    {"start": 0, "end": 13, "label": "SINGER"},
    {"start": 14, "end": 23, "label": "SINGER"},
    {"start": 40, "end": 55, "label": "SONG"}
  ]
}

Data Format

• Format: JSON Lines (JSONL) – Each line in the file is a valid JSON object.
• Encoding: UTF-8 (for full support of Hebrew characters).
• Files:
  • train.jsonl – The main dataset.
  • test.jsonl – A separate validation set.

Model and Purpose

This dataset is designed to enhance NER models in Hebrew, particularly in the music domain. Its extensive coverage and synthetic augmentation make it well-suited for training models that generalize well to unseen data.

Usage

To load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Replace USERNAME with your Hugging Face username or organization
dataset = load_dataset("NHLOCAL/SingNER")
print(dataset["train"][0])

Licensing

• License: MIT
• Attribution: If using this dataset, please cite appropriately.

Data Collection & Processing

• The dataset was sourced from real-world music databases and expanded with synthetic data using LLMS.
• Annotation was performed using Gemini, ensuring consistent and high-quality entity recognition.

Future Improvements & Contributions

• Expand dataset coverage with more entity types if needed.
• Improve entity disambiguation and multi-word entity recognition.
• Open contributions for refining annotations and expanding the dataset.

Contact & Contribution

Contributions, feedback, and enhancements are welcome! If you have suggestions, feel free to reach out via GitHub or Hugging Face.