The Conversation Model

Turn-Based Models

• Definition of a turn (linguistics)
  • Period during which one speaker holds the floor
  • Varies from interruption to several minutes
• Logical chunking for conversations
  • Like turns in a game
• Accumulate features within each turn
  • Average, range, and max

State Dependencies

• Previous turn in conversation
  • Effect of what the last person said or did
  • Models reaction to others
• Previous turn by the given speaker
  • Captures the state of the speaker so far
    
  • Example: if a speaker is angry, will remain so

Parsing a Conversation

• Learning interaction grammars
  • By labeling turns, we can learn transitions
  • Patterns of these turns: "conversation type"
• Decoding conversation progress
  • Info from multiple turns => the individuals' roles
  • More accurate than attempting to classify individual turns