Daily Note: Agent Conversational Behavior

These notes are a summary of concepts presented in “Let Me Finish First – The Effect of Interruption-Handling Strategy on the Perceived Personality of a Social Agent.”

Ronald Cumbal, Reshma Kantharaju, Maike Paetzel-Prüsmann, and James Kennedy. 2024. Let Me Finish First – The Effect of Interruption-Handling Strategy on the Perceived Personality of a Social Agent. In ACM International Conference on Intelligent Virtual Agents (IVA’24), September 16–19, 2024, GLASGOW, United Kingdom. ACM, New York, NY, USA, 10 pages. https: //doi.org/10.1145/3652988.3673916

Additional reference, The Handbook on Socially Interactive Agents

  1. Impact of Conversational Behavior on Perceptions
    • Changes in agent behavior affect assigned personality traits and user enjoyment
    • Designing interactive behavior should align with desired personality and social dynamics
    • Context-sensitive design and dynamic interruption handling are crucial for effective agent behavior
    • Conversational strategies significantly shape perceptions of agent personality and likability
  2. Speech Overlaps and Interruptions
    • Types of overlaps
      • Competitive: Floor-claiming interruptions, often with increased pitch, intensity, and speech rate
      • Cooperative: Feedback-oriented interruptions with lower pitch and no intent to claim the floor
    • Interruptions
      • Defined as attempts to grab the floor during a speaker’s turn
      • Characterized by audiovisual cues like higher pitch and glottal stops
  3. Strategies for Handling Interruptions
    • Ignore (IGNR): Agent ignores all interruptions
    • Accept (ACPT): Agent pauses and yields the floor, resuming after the user finishes
    • Acknowledge (ACKN): Agent pauses, acknowledges the interruption, finishes its turn, and then responds
  4. Effects of Interruption Strategies on Perception
    • IGNR and ACPT are preferred for perceived openness, agreeableness, and competence
    • ACKN is associated with dominance and emotional instability, often seen as rude
    • User preferences align with ease of extracting information
  5. Conversational Dynamics and User Behavior
    • Users produce fewer utterances with IGNR but longer utterances overall
    • Higher overlap in ACKN and ACPT but shorter overlap durations than IGNR
    • ACPT strategies promote user engagement and involvement
  6. Experimental Context and Design
    • Scenario
      • Interactive game setting with competing human-agent goals
    • Setup
      • Agent embodied as a non-human alien in a virtual spaceship
      • Long agent utterances encourage interruptions
  7. Evaluation
    • Mixed methods (Shapiro-Wilk, Kruskal-Wallis, Spearman, Dunn tests)
    • Focus on perceived agent personality, likability, and user interaction dynamics
  8. Findings and Implications
    • IGNR and ACPT conditions facilitate ease of information retrieval
    • ACKN strategy perceived as rude due to content prioritization and phrasing
    • Preferences depend on task competitiveness; ACPT generally favored
    • Current conversational engines predominantly use IGNR as the default