goemotions / README.md
mrm8488's picture
Create README.md


GoEmotions is a corpus of 58k carefully curated comments extracted from Reddit, with human annotations to 27 emotion categories or Neutral.

  • Number of examples: 58,009.
  • Number of labels: 27 + Neutral.
  • Maximum sequence length in training and evaluation datasets: 30.

On top of the raw data, we also include a version filtered based on reter-agreement, which contains a train/test/validation split:

  • Size of training dataset: 43,410.
  • Size of test dataset: 5,427.
  • Size of validation dataset: 5,426.

The emotion categories are: admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise.

For more details on the design and content of the dataset, please see our paper.


Our raw dataset can be retrieved by running:

wget -P data/full_dataset/ https://storage.googleapis.com/gresearch/goemotions/data/full_dataset/goemotions_1.csv
wget -P data/full_dataset/ https://storage.googleapis.com/gresearch/goemotions/data/full_dataset/goemotions_2.csv
wget -P data/full_dataset/ https://storage.googleapis.com/gresearch/goemotions/data/full_dataset/goemotions_3.csv

See the data folder for more detailed data information.

Data Format

Our raw dataset, split into three csv files, includes all annotations as well as metadata on the comments. Each row represents a single rater's annotation for a single example. This file includes the following columns:

  • text: The text of the comment (with masked tokens, as described in the paper).
  • id: The unique id of the comment.
  • author: The Reddit username of the comment's author.
  • subreddit: The subreddit that the comment belongs to.
  • link_id: The link id of the comment.
  • parent_id: The parent id of the comment.
  • created_utc: The timestamp of the comment.
  • rater_id: The unique id of the annotator.
  • example_very_unclear: Whether the annotator marked the example as being very unclear or difficult to label (in this case they did not choose any emotion labels).
  • separate columns representing each of the emotion categories, with binary labels (0 or 1)

The data we used for training the models includes examples where there is agreement between at least 2 raters. Our data includes 43,410 training examples (train.tsv), 5426 dev examples (dev.tsv) and 5427 test examples (test.tsv). These files have no header row and have the following columns:

  1. text
  2. comma-separated list of emotion ids (the ids are indexed based on the order of emotions in emotions.txt)
  3. id of the comment


Here you can view a TSNE projection showing a random sample of the data. The plot is generated using PPCA (see scripts below). Each point in the plot represents a single example and the text and the labels are shown on mouse-hover. The color of each point is the weighted average of the RGB values of the those emotions.

Data Analysis

See each script for more documentation and descriptive command line flags.

  • python3 -m analyze_data: get high-level statistics of the data and correlation among emotion ratings.
  • python3 -m extract_words: get the words that are significantly associated with each emotion, in contrast to the other emotions, based on their log odds ratio.
  • python3 -m ppca: run PPCA (Cowen et al., 2019) on the data and generate plots.


We released a detailed tutorial for training a neural emotion prediction model. In it, we work through training a model architecture available on TensorFlow Model Garden using GoEmotions and applying it for the task of suggesting emojis based on conversational text.


If you use this code for your publication, please cite the original paper:

 author = {Demszky, Dorottya and Movshovitz-Attias, Dana and Ko, Jeongwoo and Cowen, Alan and Nemade, Gaurav and Ravi, Sujith},
 booktitle = {58th Annual Meeting of the Association for Computational Linguistics (ACL)},
 title = {{GoEmotions: A Dataset of Fine-Grained Emotions}},
 year = {2020}


Dora Demszky


  • We are aware that the dataset contains biases and is not representative of global diversity.
  • We are aware that the dataset contains potentially problematic content.
  • Potential biases in the data include: Inherent biases in Reddit and user base biases, the offensive/vulgar word lists used for data filtering, inherent or unconscious bias in assessment of offensive identity labels, annotators were all native English speakers from India. All these likely affect labelling, precision, and recall for a trained model.
  • The emotion pilot model used for sentiment labeling, was trained on examples reviewed by the research team.
  • Anyone using this dataset should be aware of these limitations of the dataset.

Dataset Metadata

The following table is necessary for this dataset to be indexed by search engines such as Google Dataset Search.

property value
name GoEmotions
description GoEmotions contains 58k carefully curated Reddit comments labeled for 27 emotion categories or Neutral. The emotion categories are _admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise_.
sameAs https://github.com/google-research/google-research/tree/master/goemotions
citation https://identifiers.org/arxiv:2005.00547
property value
name Google
sameAs https://en.wikipedia.org/wiki/Google