Google AI presents “GoEmotions”: a set of NLP data for a fine classification of emotions



The emotions we experience on a daily basis can motivate him to take action and influence the important and minor decisions he makes in his life. Hence, they greatly influence the way people socialize and bond.

Communication helps us express a wide range of delicate and complicated emotions with just a few words. With recent advances in NLP, several datasets for language-based categorization of emotions have been made accessible. The majority of them focus on specific genres (news headlines, movie subtitles, and even fairy tales) and the six main emotions (anger, surprise, disgust, joy, fear and sadness). There is therefore a need for a larger scale dataset covering a greater range of emotions to enable a wider range of possible future applications.

A recent Google study presents GoEmotions: A human-annotated data set of fine-grained emotions with 58,000 Reddit comments drawn from major English subtitles and 27 emotion categories identified. It has 12 categories of positive emotions, 11 negative, 4 categories of ambiguous emotions and 1 category of “neutral” emotions, which makes it widely useful for conversational interpretation tasks that require delicate discrimination between displays. of emotions. They also present a comprehensive tutorial that demonstrates how to use GoEmotions to train a neural model architecture and apply it to recommending conversational text-based emojis.

Dataset creation

Their goal was to compile a large dataset focused on conversational data, in which emotion plays a vital role in communication. The Reddit platform is an important resource for emotion research because it provides a vast volume of publicly accessible content that includes direct user-to-user dialogue. Researchers collected Reddit comments from subreddits with at least 10,000 comments, removing deleted, non-English comments.

They used data retention procedures to ensure that the dataset did not promote general or emotion-specific linguistic biases, allowing them to create broadly representative models of emotions. This was especially important because Reddit has a well-documented demographic tilt towards young male users, not representative of the global population. The platform also promotes the use of toxic and inflammatory words.

They recognized negative remarks using specified criteria for offensive / adult and vulgar content, as well as identification and religion. They used them to filter and hide data in order to address the above concerns. They also filtered the data to remove profanity, limit text length, and balance the emotions and opinions conveyed. They also balanced the data across subreddit communities to avoid over-representing important subreddits and to ensure that comments reflect less active subreddits as well.

They focused on three objectives when building the taxonomy:

  • Provide the most comprehensive coverage of emotions expressed in Reddit data
  • Provide the most comprehensive coverage of types of emotional expressions
  • Limit the total number of emotions and their overlap.

A taxonomy like this allows for fine-grained data-driven analysis of emotions while solving the dearth of data for specific emotions.

The categories of emotion tags were defined and refined through an iterative approach to taxonomy creation. They considered a total of 56 categories of emotions during the data labeling steps. They found and eliminated emotions that were seldom chosen by reviewers, had poor agreement among reviewers due to resemblance to other emotions, or were difficult to discern from the text in this sample. Emotions commonly suggested by reviewers and well reflected in the data have also been added. Finally, they revised the names of the categories of emotions to improve interpretability. This resulted in strong agreement among reviewers, with at least two reviewers agreeing on at least one emotion tag in 94% of the samples.

They use principle preserved component analysis (PPCA) to ensure that taxonomy choices reflect the underlying data. This helped them identify emotional aspects with high agreement among reviewers.


Each component is found to be significant, demonstrating that each emotion captures a distinct element of the data. Based on the correlations between the evaluators’ judgments, they study the clustering of defined emotions. When two emotions are often co-selected by raters, they come together using this method. Despite the fact that the taxonomy does not have a specified definition of feeling, they find that emotions that are connected in terms of feeling (negative, positive, and ambiguous) cluster together, demonstrating the quality and consistency of ratings.


Likewise, emotions of similar intensity, such as joy and excitement, discomfort and fear, sadness and grief, annoyance and rage, are related.


While GoEmotions has a large collection of human-annotated emotion data, the emotion datasets use heuristics to automatically categorize weak emotions. The most popular heuristic uses emotion-related Twitter tags as categories of emotions, making it easier to create large data sets at a lower cost. However, this method is limited for several reasons.

Emojis have fewer inconsistencies than Twitter tags because they are more standardized and minimalist. They suggest a more accessible heuristic in which emojis contained in user conversations serve as a substitute for categories of emotions. This approach can be used on any linguistic corpus with a reasonable number of emojis, including many conversational corpora.

This data can be useful for developing expressive chatbots as well as generating contextual emojis, and this is a promising area for future research.




Leave A Reply

Your email address will not be published.