CHAT to Your Destination

Fuliang Weng, Baoshi Yan, Zhe Feng, Florin Ratiu, Madhuri Raya, Brian Lathrop, Annie Lien, Sebastian Varges, Rohit Mishra, Feng Lin, Matthew Purver, Harry Bratt, Yao Meng, Stanley Peters, Tobias Scheideck, Badri Raghunathan, Zhaoxia Zhang2007 · DOI

Introduction

In the field of Spoken Dialog Systems (SDS), the "ideal user" is often envisioned as a focused individual in a quiet environment. However, the reality of automotive settings presents a starkly different scenario. Drivers are frequently distracted, prone to disfluencies, and often have imperfect recall of information provided by the system. To address these challenges, the CHAT (Conversational Helper for Automotive Tasks) system was developed as a multi-domain, context-aware conversational interface designed to act as a collaborative partner for drivers.

System Architecture

The CHAT system utilizes a highly modular architecture designed to handle the complexities of real-world human-machine interaction. At its core is a Dialog Manager (DM) that employs an "information state-update" approach to maintain context across different domains.

The modular architecture of the CHAT system, highlighting the interplay between language understanding, dialog management, and domain-specific applications.

As shown in the architecture diagram, the system integrates several key components:

  • Spoken Language Understanding (SLU): This module combines a deep structural parser for complex queries with a shallow semantic parser as a robust backoff strategy. It also includes algorithms for edit region detection to handle hesitations and self-corrections.
  • Content Optimizer (CO): Positioned between the Knowledge Manager and the Dialog Manager, the CO prevents information overload by suggesting additional criteria to narrow down large search results.
  • Multi-Domain Integration: The system seamlessly manages transitions between different applications, such as Navigation, Restaurant finding, and MP3 playback, through an Application Manager.

Collaborative Dialog Strategies

A primary goal of CHAT is to move beyond simple voice commands toward a more natural, collaborative interaction. This is achieved through several strategies:

  1. Contextual Awareness: The system maintains a shared context across domains. For instance, after finding a restaurant, a user can simply say "Take me there," and the system understands "there" refers to the previously identified location, triggering a domain switch to Navigation.
  2. Handling Imperfect Input: By passing the top n hypotheses from the SLU to the DM, the system can use current dialog context to resolve ambiguities in messy or partial speech.
  3. Regulated Information Presentation: To minimize cognitive load, the system proactively offers to filter results rather than reading long lists, which is crucial for a distracted driver.

Evaluation and Results

The system's performance was evaluated in a simulated driving environment with 20 subjects performing navigation and restaurant-finding tasks. The results demonstrated high levels of effectiveness and user satisfaction:

  • Task Completion: The system achieved a 98% completion rate for navigation tasks.
  • Efficiency: Interaction was streamlined, with tasks completed in an average of 2.3 turns, a significant improvement over earlier iterations.
  • User Satisfaction: CHAT received a score of 1.98 on the CU-Communicator rating scale (where 1 is best), placing it among the top-performing systems of its time.
  • Robustness: Despite the presence of out-of-vocabulary words and disfluencies, word recognition accuracy remained stable at approximately 85-86%.

Conclusion

The CHAT system represents a significant step forward in developing conversational interfaces for high-cognitive-load environments. By prioritizing collaborative dialog and modular robustness, it successfully addresses the "human" elements of driving—distraction, forgetfulness, and imperfect speech. Future work aims to further refine these collaborative strategies and adapt the system to the acoustic challenges of real-world driving environments.

This summary is based on the paper "CHAT To Your Destination" presented at the 8th SIGdial Workshop on Discourse and Dialogue, 2007.

Summary updated

BibTeX
@misc{weng2007chat,
  author = {Fuliang Weng and Baoshi Yan and Zhe Feng and Florin Ratiu and Madhuri Raya and Brian Lathrop and Annie Lien and Sebastian Varges and Rohit Mishra and Feng Lin and Matthew Purver and Harry Bratt and Yao Meng and Stanley Peters and Tobias Scheideck and Badri Raghunathan and Zhaoxia Zhang},
  title = {{CHAT to Your Destination}},
  year = {2007},
  doi = {10.18653/v1/2007.sigdial-1.17},
}
Figure