Avoiding and Resolving Initiative Conflicts in Dialogue

Fan Yang, Peter A. HeemanNAACL 2007

TL;DR

This paper investigates "initiative conflicts" in human-human dialogue, where both participants try to take the conversational lead simultaneously. The authors found that humans actively avoid these conflicts and, when they do occur, resolve them quickly and efficiently, primarily through subtle changes in vocal volume.

What Problem Does This Paper Address?

Current dialogue systems often struggle with initiative, typically defaulting to system-initiative or offering limited mixed-initiative capabilities. To build more natural and interactive computer dialogue systems, researchers need to understand how humans manage initiative, especially when conflicts arise. An initiative conflict occurs when two people try to speak at the same time, vying for control of the conversation. This paper aims to empirically study these conflicts in human conversations to inform the design of more sophisticated dialogue systems.

Methods

The researchers analyzed two corpora of human-human task-oriented dialogues: the Trains corpus and the MTD corpus. They annotated these dialogues to identify initiative conflicts, defining them as overlapping utterances where at least one participant attempts a "forward function" (initiating a new topic or action) and the other fails to yield. To understand how conflicts are resolved, they analyzed the duration of these conflicts and the acoustic properties of the speech, specifically focusing on volume and pitch. They also examined the role of conversational context, such as the urgency of the task and the roles of the participants (e.g., system vs. user).

Datasets

  • Trains Corpus: A collection of human-human task-oriented dialogues where participants collaborate on planning manufacturing and transportation of goods. Eight dialogues, totaling approximately 45 minutes, were annotated.
  • MTD Corpus: Contains dialogues where participants play two games via conversation: an ongoing game and an interruption game. Six dialogues, totaling about 90 minutes, were annotated.

Key Insights and Ideas

  • Humans actively avoid initiative conflicts: The data suggests that people generally try to prevent these conflicts from happening, only initiating when there's an urgent need or a perceived opportune moment.
  • Conflicts are resolved quickly and efficiently: When conflicts do occur, they are typically resolved within a very short timeframe, often less than two syllables.
  • Volume is a key resolution device: The louder speaker in a conflict is significantly more likely to "win" and gain the turn. This isn't just about being loud, but about strategically increasing volume during the conflict.
  • Urgency drives conflict: Initiative conflicts are more frequent in situations with higher time pressure (e.g., a short interruption game timer), suggesting that urgency can override the general tendency to avoid conflict.
  • System initiative can be advantageous: In task-oriented dialogues, the system, with its complete domain knowledge, tends to win initiative conflicts more often than the user.

Key Results

  • Conflict avoidance: Most initiative conflicts occur at the beginning of utterances, suggesting unintentional collisions rather than deliberate interruptions. The frequency of conflicts decreases with less urgent tasks.
  • Conflict duration: Initiative conflicts are resolved quickly, with a mean duration of 328ms in the Trains corpus and 427ms in the MTD corpus, significantly shorter than typical forward utterances.
  • Volume as a predictor: The speaker with higher relative volume wins initiative conflicts with approximately 79% accuracy. This accuracy improves to 85% when considering volume increment during the conflict.
  • Pitch is less predictive: While winners tend to have slightly higher pitch, it's not a statistically significant predictor of conflict resolution.
  • System advantage: In the Trains corpus, the system won 16 out of 23 initiative conflicts, achieving 70% accuracy if predicted to always win.

Limitations and Future Work

  • Limited corpora: The study used only two specific corpora, and the findings might not generalize to all types of conversations or domains.
  • Face-to-face vs. remote: The current study focused on audio-only conversations. Future work will examine face-to-face interactions, where visual cues like eye contact and gestures play a role.
  • Multi-party conversations: The research is limited to two-party dialogues. Extending the analysis to multi-party conversations is a key future direction.
  • Building an initiative model: The authors plan to develop an initiative model that captures the negotiation process observed, incorporating devices like volume, pause, and other prosodic features to help systems manage initiative more effectively.

Summary updated

BibTeX
@inproceedings{yang2007avoiding,
  author = {Fan Yang and Peter A. Heeman},
  title = {{Avoiding and Resolving Initiative Conflicts in Dialogue}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
}
Figure