HumanFirst makes the labeling of data (from both utterances and conversation logs) easy and efficient — let’s explore this workflow!
Full Text Search
You can use full text search to start your exploration and labeling if you are looking for a specific type of data to label (our pipeline automatically annotates all uploaded data to make full-text search instantaneous).
Filter by date
You can restrict search to unlabeled data from a specific time period (for example, if you want to focus on the last conversations your users had with your bot).
You can click on “Show similar suggestions” at any time in order to find utterances semantically similar to the ones already selected. This allows you to accelerate the labeling process by efficiently grouping and labeling similar data in one shot:
Add training examples to an existing intent’s labeled data
After selecting utterances, you can immediately add them to an existing intent’s labeled data by clicking on “Label selected data”, hovering over the desired intent and clicking “Move here”.
You’ll be able to view the destination intent’s existing training data to make sure you’re labeling it in the right intent before confirming.
Real-time intent recommendation
When dealing with many different intents, it can become difficult to find the intent you wish to move utterances to — HumanFirst makes finding the right intent possible in just a few clicks.
Simply click on the flat list button in the upper right of the intent panel and intents will be ranked in a flat list based on their similarity to the selected utterances. All that’s left to do is click “Move here” and confirm the move as you would normally do when adding utterances to an existing intent.
Creating a New Intent
HumanFirst makes it easy to create new intents/labels on the fly, thanks to our flexible hierarchy — including creating sub-intents under existing intents.
Simply navigate to where you want to create a new intent, and click “Create here”. After selecting a name, the new intent will be created, with the selected utterances added to its labeled training data. You can later re-organize the hierarchy and re-organize your intents if you want to improve the structure of your data.
HumanFirst aids the labeling process in a variety of ways. From accelerating the discovery of similar utterances to simplifying the intent creation and management process, labeling data has never been easier — check out this use-case that shows how LimeChat went from 19s+ to 2s per utterance labeled (8.5x improvement in efficiency!)