Pre-annotation with ChatGPT

For text projects with Classification and Named Entity Recognition labeling jobs, it is possible to trigger annotation with ChatGPT.

📘

• Only project admins and project managers have access to this feature.
• To improve the quality of generated labels, we recommend annotating some assets manually before pre-annotating the dataset with ChatGPT, and avoiding categories that are too ambiguous like "other" or "miscellaneous".

To to trigger annotation with ChatGPT:

1. From the project queue, select the assets that you want to pre-annotate with ChatGPT.

📘

Only select assets that aren't labeled. The Kili algorithm will still be able to use the manually-labeled assets from the dataset to prepare the prompt.

1. Select Annotate with ChatGPT. The Kili algorithm takes class and entity names, job instructions, and some labeled examples (if they exist) and then automatically creates a ChatGPT prompt to perform the labeling job. In general, it takes less than 1 second per asset to finish.
2. To see the added labels, refresh the page.

You can check the performance of ChatGPT from Kili's Explore view and the Analytics page.