Does someone know how retrieval intent confidence works? Consider the following example:
Grocery Store Nutrition FAQ
faq/ask_info_fruits
- How many apples a day?
- What vitamins do apples have?
faq/ask_info_vegetables
- Are there any vitamins in tomatoes?
- Can you get food poisoning from potatoes?
faq/ask_fruit_types
- How many types of apples are there?
- What is the difference between green and red apples?
Chitchat
chitchat/food_preferance
- Do you like apples?
- What kind of apples do you prefer?
chichat/ask_weather
- What is the weather like over there?
- What do you think about the weather?
- Is it going to rain?
chitchat/ask _name
- What is your name?
- What should I call you?
Lets say the user asks the bot: “What apples do you like?” From what I have understood the bot will decide what the intent is by firstly grouping the examples like below:
faq
- How many apples a day?
- What vitamins do apples have?
- Are there any vitamins in tomatoes?
- Can you get food poisoning from potatoes?
- How many types of apples are there?
- What is the difference between green and red apples?
chitchat
- Do you like apples?
- What kind of apples do you prefer?
- What is the weather like over there?
- What do you think about the weather?
- Is it going to rain?
- What is your name?
- What should I call you?
and then after it decides (let’s say faq) the intent response selector will decide if it is faq/ask_fruit_types or something else. Is this correct?
The problem with this is that if we have a very detailed faq knowledge base but only a few chitchat cases, on overlapping subjects (like food and fruits) the bot will most likely decide the intent is faq, because chitchat is grouping irrelevant subjects and will have a lower condifence. If I had not used retrieval intents and all these intents were not grouped under chitchat / faq the bot would most certainly chose the correct intent with very big confidence which is food_preference.
How do you go about solving this imbalance because of the grouping of intents?