What is sentiment labeling?

Sentiment labeling is marking whether a piece of text expresses a positive, negative or neutral attitude, and sometimes which emotion or how strongly. It trains models that read reviews, support tickets and social media at scale.

What the work looks like

You read short texts and choose a label for each, following the project's scale. Simple projects use three labels; others use five-point scales, emotions such as anger or joy, or sentiment toward a specific aspect, for example a review that likes a phone's camera but hates its battery.

The hard cases

• Sarcasm: "Great, another update that breaks everything."
• Mixed reviews that praise one thing and criticise another.
• Neutral facts that sound negative, such as "The package arrived on Tuesday."
• Slang and dialect, which is why language projects hire native speakers.

Because these cases are subjective, projects track inter-annotator agreement closely and update guidelines when labelers disagree.

Example

"Delivery was fast, shame the shoes fell apart in a week." On an aspect-based project, the annotator marks delivery as positive and product quality as negative, with an overall label of negative.