With the phenomenon called emerging intent, we refer to an intent similar to the one that exists in the bot, expressed in phrases often used by customers, which we, however, did not include in the training set. This type of cases often get a score above 0.4 due to the similarity to existing phrases and lead to errors in phrase recognition by the model.
For example, in the bot there was an intent to "repay the debt", but there was no intent to include phrases regarding the "repayment schedule". The call logs showed that customers often asked this type of question, and the model confused them with the intent of repayment. Log analysis allowed us to create and add schedule intent to the bot.
What to do when we are not sure if we are dealing with the phenomenon of the emerging intent?
- Save these types of phrases on a separate list of logs and observe the frequency of their occurrence in subsequent conversations.
- Add them TEMPORARILY to the collective intent with the assumption that the intent will have to be broken down into target, correct intent based on the present action and will have to receive many more paraphrases (at least 20) than initially stated (2-3).
Ultimately, leaving this type of collective intents reduces the effectiveness of the bot due to the small number of paraphrases, i.e. a small number of examples for a given action. A collective intent with different actions, topics or constructions in it will never be as effective as a properly constructed intent and it is an incorrect training case for the model. Therefore, it is better to safeguard and strengthen existing correct intents than to build incorrect intents even temporarily.
Updated 3 months ago