I’m pretty new to AI/NLP/NLU stuff, and something I need to do is extracting key points from a conversation. For example, if it’s a conversion between a teacher and a student, and the teacher is telling the students why he failed this exam, I want to get the information
The reasons why he failed
What actions should the student take for the next exam
Can you upload a single/sample chat transcript. This sounds like something that is possible with NLU, but you’re likely going to need to label a certain subset of your chat logs before Rasa can take over in parsing intents/entities.
I don’t really have a sample, but, something on my mind would be like this
The teacher: Let’s look at question #5. Have you seen the question clearly?
The student: Yes
The teacher: Do you know which formula should be used to solve it?
The student: No, I don’t
I would be expecting the result is like: For question #5, the student understands the question but forgot which formula needs to be used. And, also, they may talk about something not useful for the final result, those information needs to be purged.
By “need to label a certain subset”, are you saying I need to categorize the subset of the conversation? How do I get the subset? By sentence and paragraph boundary detectors?
Keep in mind that rasa-core/nlu is made primarily for bot-human interaction. When you have a human-human interaction, you need to think things a bit differently.
If you had a bot having a discussion with the student, it would be easy for rasa to know that there is a need for a slot (variable) to be filled in at that step.
If you have a requirement to have a human-human interaction, you can always have the teacher sending intent for fixing the slots herself I guess (/choose{“my_slot”: “failed question 5”} , you need to implement the choose intent), but you’re not really using rasa-core/nlu at this point.
What if I just want to highlight something. For example, after I get the converted transcripts, I highlight the part that needs attention so that the transcript reader can quickly get the points?