A corpus-based study of interruptions in spoken dialogue

Agustín Gravano, Julia HirschbergInterspeech 2012

TL;DR

This paper investigates when and how interruptions occur in spoken conversations, moving beyond simply identifying them. By analyzing a corpus of task-oriented dialogues, the researchers found that interruptions aren't random but tend to happen at specific points in a speaker's utterance, signaled by certain prosodic cues. They also identified distinct acoustic and linguistic features that differentiate interrupting speech from regular conversational turns.

What Problem Does This Paper Address?

Understanding and managing interruptions in spoken dialogue is crucial for building more natural and effective human-computer interaction systems, such as advanced interactive voice response (IVR) systems and automatic meeting processing tools. While previous research has explored the social and linguistic aspects of interruptions, often framing them as signs of dominance or rudeness, this paper aims to delve deeper into the predictability and characteristics of interruptions. Specifically, it asks: 1) Are there particular moments within a speaker's turn where an interruption is more likely to occur? and 2) What are the acoustic and linguistic features of the speech used during an interruption? This research seeks to bridge the gap between linguistic theories of turn-taking and the practical needs of computational dialogue systems.

Methods

The study employs a corpus-based approach, analyzing spontaneous task-oriented dialogues. The researchers meticulously annotated these dialogues for various types of turn transitions, including smooth switches (S), overlaps (O), pause interruptions (PI), and simple interruptions (SI). They then extracted a range of acoustic and prosodic features from the speech segments preceding and during these transitions. These features include:

  • Prosodic Features: Pitch, intensity, speaking rate, and voice quality measures (jitter, shimmer, noise-to-harmonics ratio).
  • Lexical and Syntactic Features: Analysis of speech completion (whether an utterance is finished or incomplete).
  • Intonation: Examination of phrase accents and boundary tones using the ToBI framework.

Statistical analyses, including ANOVA tests and chi-square tests, were used to compare these features across different transition types (e.g., comparing speech before a pause interruption vs. speech before a smooth switch). Machine learning classifiers were also utilized to automatically assess the completeness of utterances.

Datasets

The primary dataset used is the Columbia Games Corpus. This corpus consists of 12 spontaneous task-oriented dyadic conversations, totaling approximately 9 hours of dialogue, elicited from 13 native speakers of Standard American English. The speech was orthographically transcribed and time-aligned. A significant portion (roughly 6 hours) was also intonationally transcribed using the ToBI framework. The paper notes that this corpus has been previously used for turn-taking research and provides detailed annotation methodologies and inter-labeler agreement measures in a related publication [17].

Key Insights and Ideas

  • Interruptions are not random: The study provides strong evidence that interruptions are more likely to occur at specific points in a speaker's turn, particularly after segments that exhibit certain prosodic characteristics, rather than at arbitrary moments.
  • Turn-medial cues matter: Speech segments preceding pause interruptions (PIs) share characteristics with turn-medial segments (those followed by a hold, H) but are not identical. They tend to be longer, have a lower mean pitch, and a faster speaking rate, and are more likely to be lexically/syntactically incomplete.
  • Interrupting speech has distinct characteristics: The initial parts of interrupting turns (PIs and SIs) differ acoustically from non-interrupting turns (Ss and Os). Interrupting speech often features higher intensity and a faster speaking rate, aligning with previous findings on the role of these cues in resolving conversational conflicts.
  • Collaborative vs. Competitive nuances: The paper hints at differences between collaborative and competitive interruptions, suggesting that collaborative interruptions (where a speaker helps complete an utterance) might be more common in PIs and exhibit different prosodic patterns than competitive ones.

Key Results

  • Predicting interruption points: IPUs (Inter-Pausal Units) preceding pause interruptions (PIs) are significantly longer, have a lower mean pitch, and a faster speaking rate compared to IPUs preceding holds (H). They are also significantly more likely to be lexically/syntactically incomplete (75.6% for PI vs. 52.6% for H).
  • Intonation patterns: IPUs preceding PIs are more likely to end with an intermediate phrase or a plateau contour, and less likely to end with a falling or high-rising contour, which are typically considered turn-yielding cues.
  • Characteristics of interrupting speech: Interrupting IPUs (both PI and SI) tend to start with higher intensity and a faster speaking rate than non-interrupting turns. PIs tend to have a lower mean pitch than Ss, while SIs start with a higher mean pitch than Os.
  • Overlap duration: For simple interruptions (SIs), maximum intensity correlates with overlap duration, suggesting speakers raise their voices during prolonged overlaps to increase the chances of success.

Limitations and Future Work

Limitations:

  • The study focuses on task-oriented dialogues, and findings might differ in other conversational contexts.
  • While statistical significance is reported, some differences (e.g., between SI and O for pitch) only approach significance.
  • The paper acknowledges that the high proportion of incomplete IPUs preceding PIs is somewhat expected given the annotation criteria for PIs themselves.

Future Work:

  • Investigate the role of filled pauses (e.g., "um," "uh"), affirmative cue words (e.g., "yeah," "okay"), and other lexical classes in and before interruptions.
  • Characterize "butting-ins" (failed interruption attempts) and collaborative completions (CCs) to gain a more nuanced understanding of different interruption types.
  • Contrast successful and unsuccessful interruption attempts.

Summary updated

BibTeX
@inproceedings{gravano2012corpus,
  author = {Agustín Gravano and Julia Hirschberg},
  title = {{A corpus-based study of interruptions in spoken dialogue}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure