TL;DR
This paper explores adapting the "centering theory" of discourse coherence to multi-party dialog. The authors propose and evaluate three naive models, finding that while centering principles seem to hold in dialog, current models struggle to predict local focus, suggesting room for improvement.
What Problem Does This Paper Address?
Understanding how conversations flow and how speakers maintain focus is crucial for building better dialogue systems, especially for tasks like reference resolution and generating coherent responses. Centering theory, a computational linguistics framework, has been successful in explaining local coherence in monologues by relating what a speaker is focusing on to the words they choose. However, applying this theory to multi-party conversations, where turns are more fluid and participants constantly shift focus, presents several challenges. This paper tackles the question of how to adapt centering theory for dialog by developing and testing initial models.
Methods
The authors developed three "naive" models to adapt centering theory for dialog, each addressing specific issues:
- Utterance Boundaries: They honored transcribed utterance boundaries, even for fragments, and split compound sentences into separate utterances.
- Cf Elements: They experimented with including or excluding first- and second-person pronouns (I/you) as discourse entities in the "forward-looking centers" (Cf) list, which represents what a speaker might focus on next.
- Backward-Looking Center (Cb): They explored different ways to determine the Cb (the entity that is the focus of the previous utterance and is also a current focus). This included considering only the immediately preceding utterance or looking at the previous utterance of either speaker. They also handled empty utterances by skipping them when determining the Cb.
- Cf Ordering: They used a standard ordering for Cf elements: Subject > Direct Object > Indirect Object > Other, with linear sentence position used for multiple "other" constituents.
Datasets
The data used for this study came from four randomly chosen dialogs from the CALLHOME-English corpus (LDC, 1997). These dialogs consisted of approximately 614 utterances, totaling about 30 minutes of speech. After annotation and processing, 664 non-empty utterances were used for evaluation.
Key Insights and Ideas
- Centering Principles Hold in Dialog: Surprisingly, the core centering rules, particularly the link between local focus and pronoun use, appear to hold well even in the dynamic environment of two-party dialog. Very few utterances violated these rules.
- Dialog Participants Matter: Including first- and second-person pronouns (I/you) as discourse entities in the forward-looking centers (Cf) list (Model 1) consistently led to better performance across evaluation metrics. This suggests that acknowledging the speakers themselves as entities of focus is important.
- Predicting Local Focus is Hard: Despite the success of centering principles, the models struggled significantly to predict the "real" topic of an utterance. The chosen backward-looking center (Cb) only matched the annotators' intuition of the "real" topic about 21% to 35% of the time.
- Transitions are Often "Expensive": The models produced a low percentage of "cheap" transitions (where the Cb is also the preferred center of the next utterance), indicating that the inferred inference load on the listener is often high, suggesting less coherent transitions than ideal.
Key Results
- Rule 1 Violations: Only 10 out of 664 utterances violated Centering Rule 1 (linking pronoun use to focus).
- Model 1 Performance: Model 1, which included dialog participants (1/2PPs) in Cf, consistently performed best.
- Cb vs. Real Topic: When a Cb could be selected, it matched the "real" topic in only 21% to 35% of cases.
- Transition Costs: All three models produced a low percentage of cheap transitions (ranging from 21% to 26%), significantly lower than the 80% reported in monolog studies.
Limitations and Future Work
- Model Naivety: The authors acknowledge that their models are "naive approximations" and that their utility is questionable. They believe revising Model 1 could lead to a more useful dialog centering model.
- Defining Utterance Boundaries: The determination of utterance boundaries in spoken dialog remains a challenge and significantly impacts the Cf lists.
- Cf Element Selection: The selection of items for Cf and their ranking order are active areas of research, and the current models' choices might be suboptimal.
- Future Directions: The authors suggest several promising avenues for future research:
- Evaluating models with other criteria, such as improved pronoun resolution.
- Experimenting with alternative Cf orderings and incorporating semantic theories that include entities referred to by personal pronouns, associations, and ellipsed entities.
- Modifying utterance boundary definitions, potentially by re-attaching interrupted utterances or using "center update units."
Summary updated