Prosodic measurements and question types in the spontal corpus of Swedish dialogues

Sofia Strömbergsson, Jens Edlund, David HouseInterspeech 2012

TL;DR

This paper investigates how intonation patterns in spontaneous Swedish dialogues correlate with different types of questions. Moving beyond the simplistic "rising tone for questions" idea, the authors propose a more nuanced categorization of questions based on their function and elicit a richer understanding of how prosody signals meaning in conversation.

What Problem Does This Paper Address?

For decades, researchers have explored the relationship between intonation and sentence type, particularly how questions are signaled. A common assumption is that a rising intonation at the end of an utterance marks it as a question, contrasting with the falling intonation of declarative statements. However, studies have increasingly shown this to be an oversimplification. Different languages, and even different types of questions within a language (like yes/no vs. wh-questions), exhibit varied intonational patterns. Furthermore, spontaneous, natural speech is far more complex than the elicited speech often used in earlier research. This paper tackles the challenge of understanding the prosodic realization of questions in natural Swedish conversation, aiming to uncover whether traditional question typologies are sufficient or if a more functional approach is needed to capture the full picture of how intonation conveys meaning.

Methods

The researchers employed a multi-stage approach to analyze questions in spontaneous Swedish dialogues. First, they extracted a substantial number of question-like utterances from the Spontal corpus, a large collection of recorded Swedish conversations. Crucially, they developed a novel annotation scheme that went beyond simple syntactic categories. This scheme included:

  • Question Type (Q1): Standard categories like yes/no (Y/N), wh-questions (WH), and alternative questions (ALT).
  • Response Elicitation (Q2): Whether a response was required, optional, or prohibited.
  • Attitude/Directionality (Q3): A key innovation, categorizing questions as "forward-looking" (seeking new information) or "backward-looking" (verifying or showing attitude towards existing information). This was a refinement from earlier, less consistent "attitude" labels.
  • Reported Speech (Q4): Differentiating between direct and reported questions.

After initial annotation and data cleaning, a subset of 600 questions was selected, balanced for speaker characteristics. For these questions, three prosodic measures were extracted:

  • Duration (DUR): The length of the utterance.
  • Pitch Variation (VAR): The standard deviation of pitch within the utterance, indicating how much the pitch fluctuates.
  • Intonation Slope (DIFF): A measure of the overall pitch contour, calculated by comparing the average pitch of the first half of the utterance to the second half. This serves as a proxy for rising or falling intonation.

Statistical analyses, including one-way MANOVA and ANOVA with post-hoc tests, were used to determine the dependency of these prosodic measures on the different question categories.

Datasets

The primary dataset used is the Sontal corpus, a collection of 120 half-hour spontaneous Swedish dialogues, recorded with high-quality audio and video. This corpus contains over 60 hours of speech. The study specifically focused on a subset of 600 questions extracted and annotated from this corpus.

Key Insights and Ideas

  • Beyond the Rising Tone: The study strongly supports the idea that there isn't a simple one-to-one mapping between intonation and question type. The traditional "rising tone for questions" is not universally applicable, especially in spontaneous speech.
  • Functional Categorization Matters: The "forward-looking" vs. "backward-looking" distinction proved to be a prosodically salient feature, suggesting that the function of a question in conversation is as important as its syntactic form for understanding its intonation.
  • Swedish Question Intonation is Unique: Contrary to findings in some other languages, Y/N questions in the Spontal corpus generally exhibited falling intonation, while wh-questions often showed rising intonation. This is a significant departure from common assumptions.
  • Prosody Signals More Than Just Information: The findings suggest that intonation in questions can signal deeper conversational meanings, such as non-understanding or a lack of acceptance of the preceding context, particularly in backward-looking questions.
  • Interplay of Factors: The prosodic realization of a question is not determined by a single factor but by an interplay between its form (e.g., Y/N vs. WH), its response elicitation degree, and its functional role (forward/backward looking).

Key Results

  • Duration: Alternative questions (ALT'S) were significantly longer than other types. Backward-looking questions were significantly shorter than forward-looking questions.
  • Intonation Slope (DIFF):
    • Backward-looking questions generally showed rising intonation, especially when response was optional.
  • Forward-looking questions generally showed falling intonation.
  • A significant difference was found between WH questions (generally rising) and Y/N questions (generally falling) when grouped by Q1 label.
  • Pitch Variation (VAR): Not significantly dependent on the Q1 label, but showed some patterns when clustered by Q2 and Q3.

Limitations and Future Work

  • Rough Estimate of Slope: The authors acknowledge that their DIFF measure is a "rough estimate" of intonation slope. More sophisticated acoustic analysis could provide finer details.
  • Annotation Disagreements: While efforts were made to resolve annotation inconsistencies (e.g., for Q3), some level of annotator disagreement is inherent in subjective annotation tasks.
  • Future Work: The authors suggest that their findings highlight the need for prosodically relevant categorization schemes that incorporate functional aspects of questions. Further research could explore the perceptual implications of these prosodic variations and investigate how these patterns manifest in other languages or in different types of Swedish discourse. The project aims to expand the database to 600 questions, allowing for more robust statistical analysis.

Summary updated

BibTeX
@inproceedings{strombergsson2012prosodic,
  author = {Sofia Strömbergsson and Jens Edlund and David House},
  title = {{Prosodic measurements and question types in the spontal corpus of Swedish dialogues}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure