Natural Language understanding (NLU) is the biggest challenge to have a good customer experience.
Training with a lot of positive comments and the gradual learning by modifying and correcting the wrong intent mapping is one approach to do it and mostly a reactive approach. This is the most widely used, trusted and correct approach unless you have a big initial set of data handy.
Another approach to make the understanding better ( not fully covered ), but reduces the risk of wrong intent mapping by certain level is using
"Negative phrases in the default fall back intent".
e.g. If you want to handle airplane tickets. All other requests for train, bus etc. related info should go to fall back. So enhance the fall back by adding these phrases for train, bus in the fallback intent. This will increase the probability based on the matching factor
This will basically do two purposes.
1. Increase the probability of new phrases to match the negative intent and return a fall back response to the end customer which is better than sending him some unexpected or stupid response.
2. Increase the probability factor to increase for the real request vs the intent.
Dialog flow ( From Google ) explains with some other working examples on their blog.
Training with a lot of positive comments and the gradual learning by modifying and correcting the wrong intent mapping is one approach to do it and mostly a reactive approach. This is the most widely used, trusted and correct approach unless you have a big initial set of data handy.
Another approach to make the understanding better ( not fully covered ), but reduces the risk of wrong intent mapping by certain level is using
"Negative phrases in the default fall back intent".
e.g. If you want to handle airplane tickets. All other requests for train, bus etc. related info should go to fall back. So enhance the fall back by adding these phrases for train, bus in the fallback intent. This will increase the probability based on the matching factor
This will basically do two purposes.
1. Increase the probability of new phrases to match the negative intent and return a fall back response to the end customer which is better than sending him some unexpected or stupid response.
2. Increase the probability factor to increase for the real request vs the intent.
Dialog flow ( From Google ) explains with some other working examples on their blog.
Great posting
ReplyDeleteEducation together with plenty of optimistic feedback as well as the progressive studying simply by adjusting and also Heroku Vs Aws repairing a bad purpose mapping will be a single way of taking action and also largely any reactive method. Here is the hottest, reliable and also appropriate method except if there is a huge original pair of info helpful.
ReplyDelete