问题
I am building a Microsoft ChatBot using LUIS for natural language processing. I would like LUIS to improve by learning new utterences for the intents identified.
For example, if my 'Greeting' intent has utterences 'Hi', 'Hello', 'Hello, how are you?', the next time it encounters 'How are you?', it may predict the intent as 'Greeting' with a low accuracy. If that utterance is learnt as part of the intent, then in future, this utterence will be predicted with better accuracy and also help us in recognizing utterences closer to this utterence.
The learning could be based on:
- All input for which intent was identified. (I understand this can cause wrong learnings).
- Inputs identified by LUIS and verified by user or agent real-time or later offline.
- Inputs identified right/wrong but verified and corrected by the agent or support team later via an easier UI.
I understand LUIS gives a 'Suggested Utterences' tab which takes care of point 3. I am trying to understand how we can automate this learning by minimal user intervention.
What are the various approaches used in projects?
回答1:
Interesting! I've done a project similar to this by getting aid of LUIS programmatic API. There's no approach that can do this task automatically. An agent should support the LUIS model to identify the new utterances and direct them for the prevailing intents or create new intents for them. Rather than going for the UI they providing, you can build your own application for this task using the programmatic API I mentioned. https://westus.dev.cognitive.microsoft.com/docs/services/5890b47c39e2bb17b84a55ff/operations/5890b47c39e2bb052c5b9c2f
来源：https://stackoverflow.com/questions/46886079/approaches-to-improve-microsoft-chatbot-with-each-user-conversation-by-learning