What do you mean, you're uncertain?: the interpretation of cue words and rising intonation in dialogue

Catherine LaiInterspeech 2010

TL;DR

This paper investigates how rising intonation interacts with "cue words" (like "yeah" or "really") in dialogue to signal uncertainty. The authors found that rising intonation primarily signals that the current topic is unresolved, rather than indicating a lack of credibility in what was said. This distinction is crucial for building more sophisticated dialogue systems.

What Problem Does This Paper Address?

Dialogue systems need to understand not just the literal meaning of words, but also the speaker's attitude and the overall flow of the conversation. Speakers express uncertainty and other attitudes through both word choice (cue words) and prosody (like intonation). While it's known that rising intonation can signal uncertainty, it's not always clear what kind of uncertainty it conveys. Does it mean the speaker doubts the information, or simply that the topic is still open for discussion? This paper aims to disentangle these interpretations by examining how rising intonation interacts with the semantics of cue words, and how this combined signal affects our understanding of dialogue structure, specifically the "Questions Under Discussion" (QUD).

Methods

The researchers conducted a perception experiment where participants rated resynthesized audio stimuli. The stimuli consisted of written contexts followed by cue word responses. The key manipulation was the prosody of the cue word responses: they varied the pitch contours (rising, falling, and neutral) and pitch range. Participants rated these responses on three scales: Expectedness (how surprising the response was given prior beliefs), Credibility (how believable the response was), and Evidence (how likely it was that the speaker would provide more explanation, reflecting the resolution of the QUD). A multilevel model was used to analyze the ratings, accounting for variations across cue words, contours, contexts, and individual subjects.

Datasets

The study used the Switchboard I Release 2 corpus of telephone conversational speech to select original cue word tokens. These tokens were then resynthesized with various pitch contours. The contexts were also drawn from this corpus. No new datasets were introduced; rather, existing data was manipulated for experimental purposes.

Key Insights and Ideas

  • Rising Intonation Signals Unresolved QUD, Not Lack of Credibility: The most significant finding is that rising intonation primarily signals that the current "Question Under Discussion" (QUD) is unresolved. It doesn't seem to make the content less credible. This suggests that rises are a way for speakers to indicate they need more information or clarification, rather than expressing doubt about the truthfulness of what was said.
  • Cue Words Carry Semantic Weight, Intonation Carries Structural Weight: Cue words themselves carry strong semantic meaning that influences credibility and expectedness. Rising intonation, on the other hand, seems to primarily affect the structural aspect of the dialogue, specifically the resolution of the QUD.
  • Interaction Matters: Prosody Modulates Semantics: The interpretation of cue words is significantly modulated by intonation. For example, a rising "yeah" can signal conditional acceptance and a request for more information, whereas a falling "yeah" is a more straightforward agreement. This interaction is more pronounced for cue words with less specific semantics.
  • "Well" is a Stronger Disaffirmative: The cue word "well" appears to be a stronger disaffirmative than "really," indicating a more definitive inability to accept the utterance into the common ground.

Key Results

  • General Trends: Affirmative cue words with falling intonation generally received higher scores across all scales, indicating more certainty, expectedness, and resolution.
  • Intonation's Effect: Rising intonation consistently pushed scores towards the "uncertain" end for Expectedness and Evidence scales, but had less of an effect on Credibility. This supports the idea that rises signal an unresolved QUD.
  • Cue Word Semantics: Cue word semantics had a large effect on perception, with stronger agreement markers like "right" generally scoring higher than others.
  • Interaction Effects: The interaction between cue word and contour was significant. For instance, rising "yeah" and "really" led to lower scores on the EVIDENCE and EXPECTEDNESS scales, while rising "well" seemed to increase scores, suggesting a more complex interaction for this particular cue word.
  • Pitch Range vs. Slope: Contrary to expectations, greater pitch range in rising intonation did not necessarily correlate with greater surprise. The authors suggest that pitch slope might be more important than overall range for signaling surprise.

Limitations and Future Work

  • Contextual Nuance: The study acknowledges that the experiment did not differentiate between epistemic clashes (disagreement with content) and relevance issues (content not fitting the current conversational goal) as sources of uncertainty. Future work could explore these distinctions.
  • Pitch Slope: While the study manipulated pitch contours, the authors suggest further investigation into the specific effects of pitch slopes is warranted.
  • Dialogue System Implications: The authors highlight the importance of these findings for building dialogue systems that can better interpret speaker attitudes and manage conversational flow. Understanding how prosody and cue words signal unresolved topics is key for systems to know when to ask clarifying questions or provide more information.

Summary updated

BibTeX
@inproceedings{lai2010what,
  author = {Catherine Lai},
  title = {{What do you mean, you're uncertain?: the interpretation of cue words and rising intonation in dialogue}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure