GoEmotions: 58,000 Reddit Comments with 27 Fine-Grained Emotion Labels

Google Research released GoEmotions in 2020, a collection of 58,000 Reddit comments manually labeled for fine-grained sentiment analysis. The dataset features 27 distinct emotion categories plus a neutral class, provided in both raw and simplified formats with predefined train, validation, and test splits.

Use Cases

• Multi-label classification to identify overlapping emotions like 'joy' and 'gratitude' within a single comment
• Fine-grained sentiment analysis using 27 specific emotion labels rather than binary positive/negative polarities
• Distinguishing between emotional and 'Neutral' text in social media monitoring

Strengths

• 58,000 manually annotated records
• 27 fine-grained emotion categories plus a neutral label
• Includes predefined train, validation, and test splits for benchmarking

Limitations

• Subjective label noise inherent in crowdsourced emotion annotation
• Linguistic bias specific to Reddit's user base and platform culture
• Monolingual English focus limits cross-cultural emotion analysis

Provenance

Source

Google Research (Arxiv 2005.00547)

Collection Method

Crowdsourced annotation of scraped Reddit comments

Freshness

Last updated January 2024; based on 2020 research publication.