Datasets:

MisileLab
/

youtube-bot-comments-v2

dataset_info:
  features:
    - name: comment_id
      dtype: string
    - name: content
      dtype: string
    - name: author_name
      dtype: string
    - name: author_image_url
      dtype: string
    - name: video_id
      dtype: string
    - name: video_title
      dtype: string
    - name: video_author
      dtype: string
    - name: parent_id
      dtype: string
    - name: is_bot_comment
      dtype: bool
  splits:
    - name: train
      num_bytes: 4165725
      num_examples: 9381
    - name: test
      num_bytes: 1172862
      num_examples: 2680
    - name: validation
      num_bytes: 591155
      num_examples: 1341
  download_size: 2812002
  dataset_size: 5929742
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: validation
        path: data/validation-*
license: mit
task_categories:
  - text-classification
language:
  - ko
size_categories:
  - 10K<n<100K

Dataset Card for youtube-bot-comment-v2

This dataset contains Korean YouTube comments labeled for bot detection, focusing on identifying automated comments that promote adult content or gambling websites.

Dataset Details

Dataset Description

This dataset consists of Korean YouTube comments collected from top South Korean videos, with binary classification labels indicating whether each comment is generated by a bot or a human user. The dataset specifically targets the identification of spam bots that promote adult content and gambling websites using repetitive patterns and specific keywords.

Curated by: Misile
Language(s) (NLP): Korean (ko)
License: MIT
Task: Binary text classification (bot detection)

Dataset Sources

Data Source: YouTube comments from top 200 South Korean videos
Collection Method: Random sampling of comments from selected videos

Uses

Direct Use

This dataset is suitable for:

Training Korean text classifiers for bot detection
Research on spam detection in Korean social media content
Developing content moderation systems for Korean online platforms
Academic research on automated content generation detection

Out-of-Scope Use

This dataset should not be used for:

General Korean language modeling without considering the specific domain bias
Classification tasks unrelated to bot/spam detection
Applications requiring detection of sophisticated or evolving bot patterns beyond the specific types captured

Dataset Structure

The dataset contains the following fields:

comment_id: Unique identifier for each comment
content: The actual comment text in Korean
author_name: Username of the comment author
author_image_url: URL of the author's profile image
video_id: YouTube video identifier
video_title: Title of the video where the comment was posted
video_author: Creator of the video
parent_id: ID of parent comment (for replies)
is_bot_comment: Boolean label indicating if the comment is from a bot (True) or human (False)

Dataset Creation

Curation Rationale

This dataset was created to address the need for Korean-language bot detection in social media comments, particularly focusing on spam bots that promote inappropriate content on YouTube. The prevalence of such automated comments on popular Korean videos motivated the creation of this specialized dataset.

Source Data

Data Collection and Processing

Comments were collected from the top 200 most popular South Korean YouTube videos through random sampling. The dataset focuses on a specific time period and represents a cross-section of popular Korean video content.

The classification process involved:

Regex-based initial filtering: Using common spam keywords and patterns such as '19금' (adult content indicator) and '19금' variations
Manual verification: Human review and labeling to ensure accuracy
Pattern identification: Focus on repetitive promotional patterns for adult and gambling websites

Who are the source data producers?

The source data consists of comments from YouTube users on popular Korean videos. The bot comments were generated by automated systems designed to promote adult content and gambling websites.

Annotations

Annotation process

The annotation process combined automated and manual approaches:

Initial filtering using regex patterns targeting common spam keywords
Manual classification and verification of edge cases
Focus on identifying repetitive promotional patterns

Who are the annotators?

Annotations were performed by the dataset curator (Misile) using a combination of pattern matching and manual review.

Personal and Sensitive Information

The dataset contains YouTube usernames and profile image URLs as they appeared publicly. No additional personal information has been added, but users should be aware that this data was originally public on YouTube.

Bias, Risks, and Limitations

Technical Limitations

Temporal bias: Comments represent a specific time period and may not capture evolving bot patterns
Platform specificity: Patterns are specific to YouTube and may not generalize to other platforms
Language specificity: Optimized for Korean language patterns and may not work for other languages
Bot type limitation: Focuses specifically on adult/gambling promotion bots, not general spam or other bot types

Content and Sampling Bias

Video selection bias: Limited to top 200 South Korean videos, which may not represent the full spectrum of Korean YouTube content
Geographic bias: Focused on South Korean content and user patterns
Topic bias: Popular videos may have different comment patterns than niche content

Classification Limitations

Pattern dependency: Relies heavily on specific keywords and patterns that may become outdated
False positive risk: Legitimate comments containing flagged keywords may be misclassified
Evolving threats: Bot patterns evolve over time, potentially reducing long-term effectiveness

Recommendations

Users should be made aware of the following:

Regular updates needed: Bot detection patterns require frequent updating as spam techniques evolve
Platform adaptation: Models trained on this data may need adjustment for other platforms
Balanced evaluation: Consider both precision and recall when evaluating bot detection performance
Ethical considerations: Ensure privacy protection when deploying models trained on this data
Complementary approaches: Combine with other detection methods for robust spam filtering

Citation

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

BibTeX:

@dataset{misile2024youtube_bot_comment_v2,
  title={youtube-bot-comment-v2: Korean YouTube Bot Comment Detection Dataset},
  author={Misile},
  year={2025},
  license={MIT}
}

Dataset Card Contact

For questions or issues regarding this dataset, please contact Misile(misile@duck.com).