Now, Where Was I? Resumption Strategies for an In-Vehicle Dialogue System

Jessica VillingACL 2010

Introduction

In-vehicle dialogue systems often fail to handle interruptions gracefully. When a driver pauses to navigate a complex intersection or respond to a safety-critical event, current AI systems frequently "reset," losing the context of the conversation and forcing the user to restart the interaction. This lack of continuity is not only frustrating but can also increase cognitive load and impact driving safety. This paper investigates how humans naturally manage interruptions and resumptions in a vehicle to inform the design of more fluid and context-aware AI dialogue managers.

Methodology

The study utilized the DICO corpus, an observational dataset featuring pairs of participants (driver and passenger) engaged in intense dialogue while driving. To simulate natural interruptions, the passenger interviewed the driver with unseen questions while simultaneously providing navigation instructions.

To monitor the driver's mental state, the researchers employed two metrics: the Tactile Detection Task (TDT), a direct measure of cognitive load via a vibrating wrist buzzer, and the Intelligent Driver Information System (IDIS), which infers workload from vehicle data like steering and braking. The resulting dialogues were annotated to track topic shifts, interruptions, and the linguistic strategies used to resume dropped threads.

Key Findings

The Redundancy Threshold

A primary insight from the study is the correlation between interruption length and the amount of information repeated upon resumption. The data reveals a clear "redundancy threshold" around the 10-turn mark.

Figure 3: Frequency of redundant vs. non-redundant resumptions based on interruption length.

As shown in Figure 3, for short interruptions (4 turns or fewer), humans rarely repeat information, with an average of 3.5 turns for non-redundant resumptions. However, for long interruptions (10 turns or more), redundant information is almost always used to "re-ground" the conversation, averaging 11.5 turns. This suggests that AI systems should adapt their resumption strategy based on the duration of the pause.

Linguistic Strategies for Resumption

Contrary to the common AI approach of asking "Where were we?", humans typically use declarative phrases to signal a return to a previous topic. For drivers, declarative utterances (e.g., "As I was saying...") were the most frequent resumption category across both interview and navigation domains.

Figure 4: Distribution of utterance types used by drivers when resuming a topic.

Figure 4 illustrates that nearly 40% of driver resumptions are declarative (DEC). Passengers showed a similar preference for declaratives in social contexts, though they frequently used imperatives (IMP) when resuming navigation tasks. Additionally, humans often use "sequencing moves"—extra-linguistic sounds like lipsmacks or filler words like "Alright"—to signal a domain switch before resuming the main topic.

Social Workload Management

The study highlighted a sophisticated "co-pilot" behavior in human passengers. Passengers rarely interrupted the driver during periods of high driving-induced stress. Conversely, drivers tended to resume conversations only when their workload dropped to "low" or "possibly driving-induced" levels. This suggests that future in-vehicle AI must be sensitive to the driver's primary task workload to determine the appropriate timing for both interruptions and resumptions.

Conclusion and Future Work

By identifying the 10-turn redundancy rule and the preference for declarative resumptions, this research provides a blueprint for more naturalistic in-vehicle AI. Current limitations include a small sample size (8 participants) and the lack of inter-coder reliability testing for the annotation schema.

Future work will focus on implementing these strategies into a functional dialogue manager. The author proposes using a Lane Change Task (LCT) to quantitatively measure whether human-like resumption strategies actually reduce cognitive load and improve driving performance compared to standard "reset-to-menu" systems.

Summary updated

BibTeX
@inproceedings{villing2010now,
  author = {Jessica Villing},
  title = {{Now, Where Was I? Resumption Strategies for an In-Vehicle Dialogue System}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
}
Figure