Abstract
Segmentation of spoken discourse into distinct conversational activities has been applied to broadcast news, meetings, monologs, and two-party dialogs. This paper considers the aspectual properties of discourse segments, meaning how they transpire in time. Classifiers were constructed to distinguish between segment boundaries and non-boundaries, where the sizes of utterance spans to represent data instances were varied, and the locations of segment boundaries relative to these instances. Classifier performance was better for representations that included the end of one discourse segment combined with the beginning of the next. In addition, classification accuracy was better for segments in which speakers accomplish goals with distinctive start and end points.
Introduction
In natural conversation, participants do not merely discuss topics; they engage in specific "activities" such as greetings, information requests, or problem resolution. Automatically identifying the boundaries between these activities—discourse segmentation—is a critical challenge for AI systems, with implications for automated summarization, meeting analysis, and interactive agents.
This research investigates the "aspectual properties" of these activities—their temporal "shape." Drawing on linguistic theories of verbal aspect, which distinguish between ongoing processes (e.g., "walking") and completed events (e.g., "walking to the store"), the authors explore whether the contrast between the conclusion of one task and the onset of another serves as the primary indicator for segmentation.
Methods
The researchers framed discourse segmentation as a binary classification task: determining if a segment boundary exists at a specific point within a sequence of utterances.
Feature Engineering
Three distinct feature sets were evaluated:
- Discourse Features: 35 features encompassing acoustic data (pitch, intensity), lexical data (word counts, named entities), and pause data (filled pauses, silence duration).
- Bag-of-Words (BOW): Standard vector representations of vocabulary.
- Combo: A combination of discourse and BOW features.
Experimental Setup
The study utilized a corpus of 82 transcribed telephone dialogues between patrons and librarians. These calls were annotated for "Dialogue Function Units" and grouped into seven activity types, such as Book-Requests and Conventional greetings. A key variable was the "context window" or span. The researchers tested spans of 1 to 4 utterances and varied the position of the potential boundary within that span.
J48 Decision Trees were primarily used for classification due to their effectiveness with non-linear data and internal feature selection.
Results and Discussion
The experiments revealed that context and boundary positioning are critical for accurate segmentation.
The Importance of Contrast
Classifiers using a 4-utterance span significantly outperformed those using single utterances. The most successful models employed "spanning" windows—data capturing both the end of a preceding segment and the start of a new one (e.g., S2P1). This suggests that boundaries are defined less by isolated "start" markers and more by the behavioral shift between activities.
Goal-Oriented Clarity
Performance varied significantly across activity types. Activities with well-defined endpoints, termed "accomplishments" in aspectual theory, were more easily detected than fluid "processes." For example, Book-Requests—a goal-oriented task—showed a recall of 32% using discourse features, far exceeding the overall average recall of 14%.
Conversely, activities like Request-Action were rarely identified correctly, likely due to a lack of distinctive temporal markers. While overall accuracy hovered around 89% (matching the majority class baseline), the models successfully identified up to 50% of actual boundaries, whereas the baseline identified none.
Conclusion
This study demonstrates that aspectual properties—specifically the contrast at transitions and the goal-oriented nature of tasks—are vital for discourse segmentation. Future work will explore larger datasets to mitigate class imbalance, refine BOW features to exclude common discourse cues, and investigate hierarchical activity structures where sub-tasks are nested within larger conversational goals.
Summary updated