Turn-Taking Rules

Function

Input(s)

• Event heirarchy with times of last occurence for each event type
• Context
• Current story
• Current sentence
• Current word
• Current time

Output(s)

• DoNothing

Take no action

• InterruptUser

Interrupt the user and take a turn

• SelfInterrupt

Stop speaking

• Backchannel

Provide backchannel feedback

• Assist

Provide a short audible or visual response to encourage the student to continue

• TakeTurn

Decide that the student's turn is over, based on silence, and take a full conversational turn to respond to the student's reading

• HeardEOS

Decide that the student's turn is over, based on the end of sentence heuristic, and take a full conversational turn to respond to the student's reading

• Prompt

Prompt the student to do something

Algorithm

The function consists of a set of linearly ordered first order rules; the first rule whose preconditions are met is triggered.

In principle, turn-taking situations are classified by asking four questions:

• Is the student taking a turn?
• Is the student speaking now?
• Is the tutor taking a turn?
• Is the tutor speaking now?

The preconditions for the rules, as actually implemented, check a number of things:

• Is the event enabled?
• Is the listener in an utterance or not?
• Is the tutor currently speaking?
• Is the student currently speaking?
• Is the discourse state appropriate for this event?
• Is the intervention level appropriate for this event? (The intervention level is a quick way to determine how many actions in a row the tutor has taken without intervening student actions.)
• How far behind is the speech recognizer?
• Are the user action and/or dead air event timers showing more time elapsed than the threshold for this event?
• Has the listener detected any special conditions? (e.g. end of sentence.)

The preconditions for the turn-taking rules are as follows (excluding "Event Enabled?" since it is included in every rule):

In Utterance?	Tutor Speaking?	Student Speaking?	Discourse State?	Intervention Level?	Decoder Caught Up?	Event Timer(s)?	Listener Condition(s)?	Event
Don't care	Yes	Yes	In sentence (1)	Don't care	Don't care	listener detects speech > tutor self-interrupts	Self-Interrupt (stop speaking)
Yes	No	Don't care	In sentence	Don't care	Don't care	User turn > interrupt user	None	Interrupt user
Yes	No	No	In sentence	0	Don't care	User action clock > backchannel; dead air clock > backchannel	None	Backchannel
Yes	No	No	In sentence	< 2	Yes	User action clock > unstick; dead air clock > unstick	None	Unstick
Yes	No	No	In sentence	Don't care	Don't care	User action clock > take turn; dead air clock > take turn	None	Take Turn
Yes	No	No	In sentence	Don't care	Yes	None	Listener heard end of sentence	HeardEOS
Don't care	No	No	Don't care	Don't care	Don't care	User action clock > prompt user; dead air clock > prompt user	None	Prompt User

All remaining conditions are assigned a "Do Nothing" event.

Status