Processing incoming utterances

Model maintenance

New data is consistently flowing in as users talk to your bot. Maintaining your model means processing incoming utterances by correcting wrong NLU interpretations, validating correct ones, and deciding whether or not to keep this new data.

Correct and validate

Intents and entities can be corrected in one click, allowing to quickly correct dozens of incoming utterances. Usual annotation tools are also available.

Once an utterance is correct, it will automatically mark it as validated. When you have a bunch of new validated examples, you can use them as an evaluation set or add them to the training data.

Validated data gives you the opportunity to evaluate your model on a regular basis with recent data. Then you can use it to augment your training data.

Controlling growth

New data is good as long as it teaches your model something new, but systematically adding everything will make your model very large and longer to train. Clai can help you deal with those challenges in an efficient way.

In the example below, “Oui” was interpreted as “basics.yes” with a very high score, the model won’t learn anything from this example so Clai recommends you to delete it. Clai looks at your training data before making those suggestions and makes sure, for example, not to suggest to delete an utterance where an entity might be missing.