Theoretically, it is possible, although rare. Currently, research is being conducted on whether such cases can give us reliable information about the occurrence of more than one intent assignment and whether it is worth allowing this type of classification as a special case of multi-intent, and whether it is worth lowering the threshold here, e.g. to 0.3.
Some typical examples of multi-intent can result from:
- When a user is providing additional context (one of the intents is the context for the other, i.e. the explanation, highlighting the problem), e.g. I applied for a loan and I want to know when the decision will be made.
- Several separate actions are provided by a user in one utterance, which they want the bot to perform at the same time, e.g. I want to take a loan, complete an application and get a new card.
We do not handle these types of cases at the moment.
How do you explain the “no_intent” situation, that has a score of 1, and the sentences seem to match the bot's intents?
As explained previously, "no_intent" always gets the score 1 in an imposed manner, so this is not a score to indicate a high degree of confidence in the intent assignment. Also, no_intent is not an intent from a technical point of view. Therefore, we are dealing here with a situation in which in fact these phrases received a score below the threshold of 0.4, i.e. they have not been trained effectively or the model varies greatly between similar cases.
In this case, we often deal with too few phrases or paraphrases of a given type in the training set.
For example, the phrase "I would like to unlock my account" receives the result no_intent and a score 1. In the set we see only 2 phrases containing the word "unlock" and 2 phrases containing the verb "unlock" in relation to the account. Apparently the interpretation is as follows:
- More paraphrases of this phrase should be added.
- Make sure that they contain parts of speech other than the verb "unlock" itself.
Updated 3 months ago