I wonder what is meant by a “turn” in Rasa terms.
Although I looked it up in both the internet and the rasa docs, I couldn’t get an unambigous definition.
In my understanding, in general conversational terms, turn could be:
- conversation handled over or taken over from speaker-1 to speaker-2
- Speaker jumps to other topic
- Speaker-1 utters new intent ( f.e. “I want some thee”, followed by “I’d like to watch a video on chatbot”, either with intermediate speaker-2 reaction or not)
So, if the rasa docs are about turns (f.e. in events and policy), what is meant?
Apart from helping me understand, I also suggest you to define it in the glossary.
A “turn” is a step in the conversation, e.g. a user message plus a bot response. If we say, that something like “every turn each policy defined in your configuration will predict a next action”, we basically mean after every incoming user message, each policy defined in your configuration will predict a next action.
In general you distinguish between a single-turn and a multi-turn conversation. In a single-turn conversation the bot response just depends on the last user message given, e.g. the last turn. Multi-turn conversation are more complex as the bot needs to consider things that happened multiple turns before, e.g. a couple of messages before.
Does that clarify your question?
I think adding it to the glossary seems like a good idea.
Yes, clear explanation.
Although I had some idea on this, now I know for sure.
I might not get the point here but: the way you explain it would imply that exchanges may only be initiated by humans, and that to each human action, there would be one and only one machine reaction, both of which are not what I see in the data. Instead, I see sequences of actions by either side and I see opening actions (or sequences therof) by bots. Is this maybe subject to configuration?
Coming closer to the original topic of the thread: I’m not a linguist but the way I read the literature means that a turn is any one action bei either of a set of conversation participants (“speakers”), where the rules of turn taking are squarely extra linguistic (pragmatic, at best).
Any comment on this view?