taishi-i/awesome-japanese-nlp-classification-dataset

Dataset overview

This dataset identifies whether a GitHub repository description pertains to Japanese natural language processing (NLP). The labels are categorized as "Relevant (1)" and "Not Relevant (0)".

Problem Setting:

• Training Data: Repository descriptions from before 2022
• Test Data: Repository descriptions from 2023
• Objective: To detect repositories related to Japanese NLP

Data Collection:

• Positive Examples: Repositories listed in "awesome-japanese-nlp-resources" as of September 9, 2023
• Negative Examples: Collected from the GitHub API and visually confirmed
• Note: The annotation process is subjective

Dataset Features:

• Subjective labeling
• Mixed English and Japanese descriptions
• Imbalanced label distribution

These dataset features mirror real-world challenges and are ideal for evaluating models. Based on GitHub's terms of service, please use this dataset for research purposes only.

How to use this dataset

How to load in Python.

from datasets import load_dataset

dataset = load_dataset("taishi-i/awesome-japanese-nlp-classification-dataset")

Details of the dataset.

DatasetDict({
    train: Dataset({
        features: ['label', 'text', 'url', 'created_at'],
        num_rows: 5496
    })
    validation: Dataset({
        features: ['label', 'text', 'url', 'created_at'],
        num_rows: 400
    })
    test: Dataset({
        features: ['label', 'text', 'url', 'created_at'],
        num_rows: 856
    })
})

Baseline

Baseline trained with bert-base-multilingual-cased. Please use the baseline model from here. The F1-score for label 1 is important for this task.

Label	Precision	Recall	F1-Score	Support
0	0.98	0.99	0.98	796
1	0.79	0.70	0.74	60
Accuracy			0.97	856
Macro Avg	0.89	0.84	0.86	856
Weighted Avg	0.96	0.97	0.97	856

Dataset stats

Label distribution:

Dataset	Label 0 (%)	Label 1 (%)
Train	92.59	7.41
Validation	95.75	4.25
Test	92.99	7.01

Relevant sample:

{
    "label": 1,
    "text": "JGLUE: Japanese General Language Understanding Evaluation for huggingface datasets",
    "url": "https://github.com/shunk031/huggingface-datasets_JGLUE",
    "created_at": "2023-02-25T04:33:03Z"
}

Not Relevant sample:

{
    "label": 0,
    "text": "Official repository of FaceLit: Neural 3D Relightable Faces (CVPR 2023)",
    "url": "https://github.com/apple/ml-facelit",
    "created_at": "2023-04-03T22:47:29Z"
}

Number of texts, average number of characters per text, minimum number of characters, maximum number of characters:

Dataset	Text Count	Average Length	Min Length	Max Length
Train	5496	58.05	2.0	609.0
Validation	400	54.33	8.0	226.0
Test	856	58.85	3.0	341.0

Proportion of text languages:

Dataset	English (%)	Japanese (%)
Train	89.34	10.66
Validation	82.00	18.00
Test	83.18	16.82

Time range:

Dataset	Start Date	End Date
Train	2008-02-11 22:55:26+00:00	2022-09-30 19:45:09+00:00
Validation	2022-10-01 06:02:56+00:00	2022-12-31 12:12:41+00:00
Test	2023-01-01 06:15:03+00:00	2023-08-21 15:30:53+00:00

License

We collect and publish this dataset under GitHub Acceptable Use Policies - 7. Information Usage Restrictions and GitHub Terms of Service - H. API Terms for research purposes. This dataset should be used solely for research verification purposes. Adhering to GitHub's regulations is mandatory.