From Birdwatch to Community Notes, from Twitter to X: four years of community-based content moderation

Dataset and Code Description

This repository contains the data and code used to analyse interactions within the Community Notes platform from January 23, 2021, to January 23, 2025. The files are organised as follows:

🧪 Code Notebooks

• Create_graphs.ipynb: Constructs full interaction networks and separate sub-networks (helpful, somewhat helpful, unhelpful) from the monthly raw rating files.

• Url_analysis.ipynb: Detects the language of each note and extracts any URLs or domain names mentioned.

• topic_modeling.py: The code for the topic detection protocol used for categorising the notes.

📄 Data Files

Notes Data

• notes_with_lang.csv: All Community Notes written between January 23, 2021, and January 23, 2025, with detected language, extracted URLs, and domain names.

• english_notes_topics.csv: Subset of English notes with their allocated topics.

Each note file contains the following variables:

• noteId: Unique ID of the note.

• noteAuthorParticipantId: Unique ID of the note's author.

• tweetId: ID of the tweet the note addresses.

• date: Date the note was written (YYYY-MM-DD).

• Timestamp: Time the note was written (HH:MM:SS).

• language: Detected language of the note.

• extracted_urls: List of URLs mentioned in the note.

• news_source: List of extracted domain names.

• label1: The main topic allocated to the note.

• label2: The secondary topic allocated to the note.
• confidence: The level of confidence LLM had for the categorisation

Rating Data

• Monthly rating files are stored in the rating monthly files/ directory with the naming format ratings_m_yyyy.csv.

• Each file includes:

  • noteId: ID of the rated note.

  • raterParticipantId: ID of the participant giving the rating.

  • helpfulnessLevel: Rating category (HELPFUL, SOMEWHAT_HELPFUL, NOT_HELPFUL).

  • helpful, notHelpful: Deprecated binary flags (use helpfulnessLevel instead).

🌐 Network Files

Each month’s ratings are used to construct interaction graphs with user-to-user edges based on rating behaviours.

• Whole Networks (whole_network_<month>_<year>.graphml): Full user interaction networks, with edges annotated by the number of helpful, unhelpful, and somewhat helpful ratings.

  • Each edge contains:

    • source: Rater’s participant ID.

    • target: Note author’s participant ID.

    • helpful, unhelpful, somewhathelpful: Count of ratings by type from rater to author.

• Helpful Networks (network_<month>_<year>_helpful.graphml): Subnetworks based on helpful ratings only.

• Somewhat Helpful Networks (network_<month>_<year>_somewhat.graphml): Subnetworks based on somewhat helpful ratings.

• Unhelpful Networks (network_<month>_<year>_unhelpful.graphml): Subnetworks based on unhelpful ratings.