Dataset Card for "sentiment140"

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Homepage: http://help.sentiment140.com/home
• Repository: Sentiment140 at Stanford
• Paper: Twitter Sentiment Classification using Distant Supervision (Go et al., 2009)
• Point of Contact: Alec Go, Richa Bhayani, Lei Huang (Stanford University)
• Size of downloaded dataset files: 81.36 MB
• Size of the generated dataset: 225.82 MB
• Total amount of disk used: 307.18 MB

Dataset Summary

Sentiment140 is a large-scale Twitter sentiment classification dataset containing 1.6 million tweets automatically labelled using emoticons as a form of distant supervision. Tweets containing positive emoticons (e.g., :), :-)) are labelled positive (sentiment=4), and tweets containing negative emoticons (e.g., :(, :-( ) are labelled negative (sentiment=0). The dataset is widely used as a benchmark for binary and multi-class sentiment classification in NLP research.

Supported Tasks and Leaderboards

• Text Classification / Sentiment Analysis: The dataset is designed for binary sentiment classification (negative vs. positive). The sentiment field serves as the label. Models are typically evaluated on the 498-example test split using accuracy and F1 score.
• Leaderboard and benchmark results can be found on Papers With Code — Sentiment140.

Languages

The dataset contains English-language (en) tweets collected from the Twitter platform.

Dataset Structure

Data Instances

sentiment140

• Size of downloaded dataset files: 81.36 MB
• Size of the generated dataset: 225.82 MB
• Total amount of disk used: 307.18 MB

An example of 'train' looks as follows.

{
    "date": "23-04-2010",
    "query": "NO_QUERY",
    "sentiment": 3,
    "text": "train message",
    "user": "train user"
}

Data Fields

The data fields are the same among all splits.

sentiment140

• text: a string feature. The raw text content of the tweet.
• date: a string feature. The date and time the tweet was posted.
• user: a string feature. The Twitter username of the author.
• sentiment: a int32 feature. Sentiment label: 0 = negative, 2 = neutral, 4 = positive.
• query: a string feature. The query keyword used to retrieve the tweet, or NO_QUERY if none.

Data Splits

Dataset Creation

Curation Rationale

The dataset was created to address the lack of large-scale labelled data for Twitter sentiment analysis. Rather than relying on manual annotation, the authors used a distant supervision approach: emoticons in tweets serve as noisy but scalable sentiment labels. This allowed the collection of 1.6 million labelled examples without human annotators.

Source Data

Initial Data Collection and Normalization

Tweets were collected using the Twitter API by querying for tweets containing positive or negative emoticons. Tweets with both positive and negative emoticons were discarded. Usernames and URLs were replaced with consistent placeholders (@user, URL) to reduce noise. The resulting dataset was split into a large training set (1.6M tweets) and a manually-labelled test set (498 tweets).

Who are the source language producers?

The text was produced by general Twitter users writing in English. The dataset was collected and processed by researchers at Stanford University: Alec Go, Richa Bhayani, and Lei Huang as part of the CS224N course project.

Annotations

Annotation process

Training labels were assigned automatically using emoticons as proxies for sentiment (distant supervision) — no human annotation was used for the training set. The test set of 498 tweets was manually annotated by the authors for evaluation purposes.

Who are the annotators?

Training set: no human annotators — labels derived automatically from emoticons. Test set: manually labelled by the paper's authors (Go, Bhayani, Huang) at Stanford University.

Personal and Sensitive Information

The dataset contains real Twitter usernames and tweet content from public accounts. While the data was publicly available at the time of collection, users may not have been aware their tweets would be used for research. Researchers using this dataset should be mindful of potential privacy implications, particularly when working with the user field.

Considerations for Using the Data

Social Impact of Dataset

Sentiment140 has been widely used to train and benchmark NLP models for social media sentiment analysis. Models trained on this data have applications in brand monitoring, public opinion research, and crisis detection. However, since labels are derived from emoticons, the dataset may not capture nuanced or ambiguous sentiment, which could affect the reliability of downstream models.

Discussion of Biases

• Emoticon bias: Labels are derived from emoticons, which are more commonly used by certain demographics, potentially underrepresenting users who express sentiment through language alone.
• Topic bias: Tweets were collected via keyword queries, meaning certain topics and events are overrepresented.
• Temporal bias: Tweets were collected in 2009–2010; language use, slang, and topics on Twitter have evolved significantly since then.
• Class imbalance in test set: The 498-example test set is small and may not be representative of real-world distributions.

Other Known Limitations

• The training set contains no neutral tweets (only positive and negative via emoticons); the sentiment=2 (neutral) class only appears in the test set.
• Tweet text may contain noise including misspellings, abbreviations, and platform-specific formatting.
• The dataset is from 2009–2010 and may not generalise well to contemporary social media text.

Additional Information

Dataset Curators

The dataset was created by Alec Go, Richa Bhayani, and Lei Huang at Stanford University as part of the CS224N Natural Language Processing course. It is hosted on the HuggingFace Hub by @patrickvonplaten and @thomwolf.

Licensing Information

The dataset was made freely available by the authors for research purposes. No explicit open-source license is attached. Users should refer to Twitter's Developer Policy regarding use of tweet content.

Citation Information

@article{go2009twitter,
  title={Twitter sentiment classification using distant supervision},
  author={Go, Alec and Bhayani, Richa and Huang, Lei},
  journal={CS224N project report, Stanford},
  volume={1},
  number={12},
  pages={2009},
  year={2009}
}

Contributions

Thanks to @patrickvonplaten, @thomwolf for adding this dataset.