Extract message intent instead of setting the node with a dict when using visualization

[MetcalfeTom]

Fixes RasaHQ/rasa#2995

Under certain circumstances we were passing the message parse data to the node label during interactive learning visualization, which (since this is a dictionary) was causing fatal errors.

Proposed changes:

• Use message intent name instead of the parse data as the graph node label

Status (please check what you already did):

• made PR ready for code review
• updated the changelog

[codeclimate]

Code Climate has analyzed commit 1386fa8 and detected 0 issues on this pull request.

View more on Code Climate.

[paulaWesselmann]

Why is message set to parse_data here?

            if isinstance(el, UserUttered):
                if not el.intent:
                    message = interpreter.parse(el.text)
                else:
                    message = el.parse_data

[MetcalfeTom]

@paulaWesselmann

Just a convenient way of passing both the text and the intent; keeping the text makes the graph readable and keeping the intent allows us to compare equivalent nodes. If the same intent has appeared twice, we overwrite the node label with the subsequent message text.

[paulaWesselmann]

Approved because it error proofs this from happening. But I'm still confused on how it comes to this sequence of events that it is able to happen at all.