Amazon Mechanical Turk closed on 2026-09-30. Here is how to keep your work running →

Services / Text annotation

Text annotation

We label text for your models and your operations: intent, topic, sentiment, entities and your own categories. Trained annotators apply your taxonomy the same way every time. The team lead brings you the edge cases your guidelines do not cover yet, so the taxonomy gets sharper each week.

Request a pilot plan →See pricing

Supported tasks

  • ·Classify support tickets by topic, product and urgency
  • ·Label intent and outcome in chat transcripts
  • ·Tag sentiment and product issue in reviews
  • ·Mark entities, such as product names, companies and places
  • ·Label training and evaluation data for text models
  • ·Relabel an existing dataset after a taxonomy change

Sample input and output

Illustrative example. Invented data, not from a customer.

Sample input

  • Your taxonomy: Topic (Billing, Shipping, Account, Product) · Urgency (Low, Normal, High) · Sentiment (Positive, Neutral, Negative)
  • Ticket: I was charged twice for my March order and the tracking link doesn't work.

Sample output

  • Topic: Billing
  • Urgency: High (the customer was charged twice)
  • Sentiment: Negative
  • Entity: "March order" (order reference)
  • Flag for your contact: The ticket also raises a Shipping issue. Your guidelines allow one topic. Should tickets like this take two?

How it works

  1. 1

    You share the taxonomy. Send the label set, guidelines and labeled examples.

  2. 2

    We train and build a gold set. Annotators train on your examples. We build a set of items with known-correct labels and confirm it with you.

  3. 3

    Annotators label live items. They work in your tool. Edge cases go to the team lead.

  4. 4

    You get a weekly report. Items labeled, QA sample results, the labels most often confused, and guideline questions.

Quality checks

The team lead checks: edge cases, and keeps a running list of decisions. Guideline updates go to the whole team at once. The team lead also tracks each annotator's QA results.

QA reviewers sample: a share of finished items, relabeled and compared with the annotator’s label and the gold set. The report shows which label pairs get confused most, for example Billing and Account.

What this means: every item is labeled by one trained person. QA re-checks a sample, not every item. The weekly report shows the sample size and the acceptance rate.

How we check quality →

Pricing scope

  • $5 per person-hour (standard tasks): English text classification, tagging, entity marking and sentiment, to your taxonomy.
  • Quoted after our ops team reviews the task: text that needs domain expertise (for example clinical notes or legal contracts), languages other than English, or annotation in a tool we have to set up for you.

FAQ

Can you work in our labeling tool?

Yes, if it runs in a browser and you can create accounts for us. That includes SageMaker Ground Truth, open-source labeling tools and spreadsheets.

What happens when our taxonomy changes?

Tell your team lead. We update the guidelines, retrain the team and note the date in the weekly report, so you know which items used which version.

Is every item labeled twice?

No. One trained annotator labels each item. QA reviewers relabel a sample.

Can you label text that contains personal data?

Yes, if you tell us first. We agree in writing who can see it and what gets masked. See the security page.

Need image, video or lidar annotation? Specialized annotation for computer vision, automotive and robotics is handled by the Deepen AI enterprise team.

See Deepen AI Annotate →

Try it on your own data for two weeks.

Send your taxonomy and a sample. We email a pilot plan with a quote.

Request a pilot plan →