MMT (Multilingual and Multi‑Topic Twitter Language Identification Dataset)

MMT: A Multilingual and Multi‑Topic Indian Social Media Dataset is a large-scale language identification dataset derived from 1.7 million tweets collected from Indian Twitter/X, annotated with coarse and fine-grained language labels. It supports research on multilingual and code-mixed text in noisy, real-world social media settings.

📁 Files Included in This Release

MMT_LID_sample10.csv: A small illustrative sample of 10 tweets, allowed under Twitter/X’s non-automated sharing exemption.

Columns:

• tweets: Raw tweet text
• language_detected: Predicted language (automated)
• is_the_language_correct: Binary correctness flag (1 = correct)
• correct_language_tag: Human-corrected language tag

⚠️ Not for bulk distribution. For research use only.

MMT_distributable_without_tweets.csv: This is the main distributable dataset (≈1.35M tweets), compliant with Twitter Developer Policy.

Includes:

• Tweet IDs
• Derived metadata (e.g., language tags, annotations, splits)
• No tweet text, media, or JSON content

📥 Request Full Dataset Access

👉 Request Full Dataset Access via Google Form
Please fill this form to request access to the full version of the MMT LID Full Dataset.
Mail at: [email protected]

twitter_dataset_release_rules.txt: Compliance documentation outlining redistribution rules from Twitter/X's developer policy and privacy guidelines (GDPR/CCPA).

Dataset Usage Disclaimer ‼️

This dataset is derived from public social media content and may contain:

• Offensive, sensitive, or biased material
• Non-standard or noisy language

Please use it responsibly and in accordance with ethical guidelines and applicable data protection laws (GDPR, CPRA).

Note: “This release contains only Post IDs/User IDs and our derived labels/metadata. We do not distribute tweet text, media, or full tweet/user JSON. Requesters must hydrate IDs via the X API and must honor deletions, privatizations, and other changes; any valid takedown request will be processed within 24 hours. Distribution is subject to the X Developer Policy, including the 1,500,000-IDs limit, and to applicable privacy laws (e.g., GDPR/CPRA). As a result, we have a reduced version of the file: MMT_distributable_without_tweets, which is under 1.5M posts (i.e., in the ballpark of 1.35M).”

⚖️ Licensing & Attribution

• Curated by: Lingo Research Group at IIT Gandhinagar
• Languages: Multilingual (Hindi, English, Gujarati, Bengali, Tamil, etc.)
• License: cc-by-4.0 (for derived annotations only)
• Tweet Content License: Subject to Twitter Developer Policy

📚 Citation

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

@inproceedings{dalal-etal-2023-mmt,
    title = "{MMT}: A Multilingual and Multi-Topic {I}ndian Social Media Dataset",
    author = "Dalal, Dwip  and
      Srivastava, Vivek  and
      Singh, Mayank",
    booktitle = "Proceedings of the First Workshop on Cross-Cultural Considerations in NLP (C3NLP)",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.c3nlp-1.6/",
    doi = "10.18653/v1/2023.c3nlp-1.6",
    pages = "47--52"
}