Previous model’s performance: High human-machine agreement (Cohen’s Kappa: .89); 88% for non-questions, 94% for option-posing questions, and 95% for wh-questions/invitations
New model’s performance: High human-machine agreement
82% for non-questions
82% for option-posing questions,
98% for wh-questions
96% for invitations
Future Work
Differentiate between subtypes of option-posing questions
Yes-no questions (“Did you go home that night?”)
Forced choice questions (“Did you go home that night or did you stay with your sister?”)
Suggestive questions (“You went home that night, didn’t you?”)
Categorize children’s responses
Unelaborated (Q: “Did you go home that night?” A: “No.”)
Elaborated (Q: “Did you go home that night?” A: “No, I stayed with my sister.”)
