My chatbot’s expected utterances are pretty lengthy/wordy. The user’ question requires wordy description often. I mean each utterance might have up to 100 words and which is almost the common situation. As you may guess, when the utterances are getting longer the variation of wording increasing much. So the bot starts to confuse the intents. You can think of this one as a kind of FAQ bot. What do you recommend to deal with long utterances? What should be the best approach to handle those smoothly.
How does your training data look like? It should be similar to utterances you will find in the dialogue in the end.
As long as the utterance does not switch the topic in between, the model should be able to detect the correct intent. Can you maybe paste an example utterance here? That might help. Thanks.
Thank you very much @Tanja for your response.
The below utterance is coming from actual user:
"Hi my name is XXX XXX I have you as my main insurance I was going over my bill and it looks like it’s 400$? When I first got unified insurance my second months payment was rejected and I was never told so my bill kept adding up and nobody called and told me that it was rejected and I called myself and I was told I’d receive a call back the following week and never did. Now I have a cancellation notice on my insurance for no reason at all, nobody communicated that my payments were not going through I have even called! If I can get some feedback from this it would be appreciated because I can afford my insurance there’s just absolutely no communication, I’ve never had this problem with any insurance agency’s I’ve gone through and I got recommended to go through you and this is what happens?
What do you recommend to handle those kind of utterances?
What intent would that message be? How many intents do you have in total? Do you have similar intents? E.g. two intents that are related to insurance?
In general, the model should be able to detect the correct intent no matter how long the text is, as long as
the intents are not too similar and the data is labeled consistently
you have provided enough training data for all intents
the user is just talking about one thing (we only predict one thing at a time)
Did you evaluate your model (Evaluating Models)? It should help you to get a feeling on how well your model is doing. It should also tell you which intents the model will get most likely incorrect. You can then take a look at those intents and check if, for example, you maybe labeled them inconsistently or training data is missing.