HumanFirst’s workflows helps teams expand their training data using data-driven approaches.
By selecting an intent and clicking on
Get Suggestions, HumanFirst will run a nearest neighbour search across your data to surface additional training examples to be added to the intent. You can accept suggestions by clicking on them or clicking
Accept All if they are all good. Rejecting suggestions is done by clicking on
None of these look good for the suggestions that are left. Doing this creates a negative signal to not propose anything semantically similar to what has been rejected in the next turn of suggestions.
Similarity slider allows to broaden the search to find examples that might fit in the intent but that are not perfect matches. This helps mitigate over fitting the intent.
Explore by similarity allows to search your unlabeled data using intents/labels. If you’re using HumanFirst NLU as the active NLU engine, you can choose the minimum match score of results to narrow the results further.
We recommend lowering the minimum match score and ordering the data with increasing match scores. Starting your analysis with lower match scores will allow you to be selective as to what new training examples to annotate and expand your training data without overfitting your intents.