📄 Common Ground and Perspective-taking in Real-time Language Processing

Summary

This one-page keynote abstract from SIGDIAL 2011 asks whether people engaged in dialogue actually track which knowledge they share with their partner (common ground) and which is theirs alone (privileged ground) while language unfolds in real time. One view says such bookkeeping is too costly in computation and memory to guide online processing. Experiments showing that speakers and listeners often behave egocentrically support that view, because their private knowledge leaks into what they say and how they interpret. Tanenhaus summarizes work from his Rochester lab that used unscripted conversation rather than scripted tasks. In that work, the way speakers phrase utterances reveals whether they consider information shared or private, and listeners pick up on these cues with great sensitivity. He proposes an integrative framework to reconcile the conflicting findings in the literature and suggests it may be useful to people building computational dialogue systems. The abstract gives no detailed numbers.

Key points

• The talk contrasts two kinds of knowledge: common ground, which is shared between interlocutors, and privileged ground, which is private to one of them.
• Earlier experimental results show strong egocentric intrusions. These have been taken as evidence that tracking common ground is too expensive for real-time use.
• The new evidence comes from unscripted conversation studies in Tanenhaus's lab (Brain and Cognitive Sciences, University of Rochester). It shows two things:
  • Speakers' utterances carry signals about whether they believe information is shared.
  • Addressees respond strongly to those signals.
• The talk offers an integrative framework to explain why different studies disagree.
• Venue: SIGDIAL 2011, the 12th Annual Meeting, Portland, Oregon, June 17–18, 2011. The abstract appears on page 67.

Interesting ideas

• I1. Keep separate stores for what a participant privately knows and what all participants share, instead of one global fact pool, because many interpretation errors come from mixing the two.
• I2. Treat the form of an utterance as evidence about the speaker's beliefs about shared knowledge, so phrasing itself reveals what the speaker assumes the listener already knows.
• I3. When studies disagree, look for a framework that explains when each result holds rather than declaring one side wrong, since task conditions such as scripted versus free conversation may drive the difference.
• I4. Allow egocentric interpretation as a cheap default and correct it when cues signal privileged knowledge, which balances processing cost against accuracy.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): The world of the text could tag each entity or state with who knows it (narrator, a given character, all parties), which makes questions about beliefs and deception in fables easier to answer explainably.
• I2 (frames): The prag crate could read assumed shared-versus-private status from linguistic cues and attach it to frames, which would feed implicature and coreference decisions.
• I3 (learn): When evaluating Mova, results on constructed test items could be checked against naturally occurring text, so the evaluation shows which behaviors depend on artificial test conditions.
• I4 (answer): Readers could first answer from the full text world, then filter the answer through what a specific character could know when the question concerns that character's perspective.