jiangchengchengNLP/Enhanced_Emotion_Classification_Dataset

Enhanced Emotion Classification Dataset (v2.0)

Dataset Description

This dataset is an enhanced version of the emotion classification dataset, including multiple sources of emotion data with Ekman emotion mapping. It contains a total of 240,426 samples across 7 emotion categories, with each sample labeled with its original data source.

Dataset Structure

The dataset is split into three parts:

• Train: 186,619 samples (77.6%)
• Validation: 31,086 samples (12.9%)
• Test: 22,721 samples (9.4%)

Emotion Categories

The dataset includes 7 Ekman basic emotions:

• neutral
• joy
• sadness
• anger
• fear
• surprise
• disgust

Data Sources

The dataset combines data from 7 different sources:

1. Movies_Reviews

   • Movie review emotion data
   • Contains 7 emotion categories

2. DailyDialog

   • Real dialog data with multi-turn conversations
   • Contains 7 emotion categories

3. GoEmotions

   • Reddit comment data with colloquial expressions
   • Contains 7 emotion categories

4. ISEAR

   • International emotion research data with high-quality text
   • Contains 7 emotion categories

5. MELD

   • Multimodal emotion dialog data from the TV show "Friends"
   • Contains 7 emotion categories

6. mteb_emotion

   • Emotion analysis dataset with various emotion expressions
   • Contains 7 emotion categories

7. Tweet Emotions

   • Twitter tweet data including @mentions
   • Contains 7 emotion categories

Data Format

Each CSV file contains the following columns:

• text: Text content
• label_text: Emotion label (one of the 7 emotion categories)
• source: Data source identifier (e.g., dailydialog, goemotions, tweetemotions, etc.)

File Structure

dataset_huggingface_enhance/
├── train.csv              # Merged training set (186,619 samples)
├── val.csv                # Merged validation set (31,086 samples)
├── test.csv               # Merged test set (22,721 samples)
├── README.md              # Dataset documentation
├── dataset_infos.json     # Hugging Face dataset configuration
├── info.md                # Detailed dataset statistics
├── label_list.txt         # List of emotion labels
└── stats.py               # Dataset statistics script

Usage

To use this dataset with Hugging Face Datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("jiangchengchengNLP/Enhanced_Emotion_Classification_Dataset")

# Access specific splits
train_dataset = dataset["train"]
val_dataset = dataset["validation"]
test_dataset = dataset["test"]

Emotion Distribution

Training Set (186,619 samples)

• neutral: 51.11%
• joy: 21.64%
• sadness: 7.98%
• anger: 5.97%
• fear: 5.90%
• surprise: 5.11%
• disgust: 2.29%

Validation Set (31,086 samples)

• neutral: 40.96%
• joy: 24.15%
• sadness: 9.38%
• fear: 8.06%
• anger: 7.49%
• surprise: 7.06%
• disgust: 2.89%

Test Set (22,721 samples)

• neutral: 45.11%
• joy: 22.88%
• sadness: 9.13%
• anger: 7.61%
• fear: 6.67%
• surprise: 5.93%
• disgust: 2.67%

Notes

• The dataset has class imbalance, with neutral being the most common category (40-51%) and disgust being the least common (2-3%).
• Each sample includes a source field indicating its original data source, which allows for source-specific analysis.
• Different data sources have different text styles, which may affect model performance.
• The dataset uses Ekman emotion mapping, which maps various emotion labels to the 7 basic emotions.

License

The dataset is released under the MIT License.

Citation

If you use this dataset in your research, please cite the original datasets:

• DailyDialog: https://aclanthology.org/I17-1099/
• GoEmotions: https://arxiv.org/abs/2005.00547
• Tweet Emotions: https://www.aclweb.org/anthology/W18-6212/
• MELD: https://arxiv.org/abs/1810.02508
• ISEAR: https://link.springer.com/article/10.1007/BF02112196

Version History

• v2.0 (2026-01-07): Updated dataset with merged sources and source field

  • Total samples: 240,426
  • Added source field to all samples
  • Updated emotion distribution statistics
  • Improved data quality and consistency
  • Filtered all NaN values from the dataset

• v1.0: Initial release