Using graphical models for mixed-initiative dialog management systems with realtime Policies

Stefan Schwärzler, Stefan Maier, Joachim Schenk, Frank Wallhoff, Gerhard RigollInterspeech 2009

TL;DR

This paper introduces a novel graphical model (GM) for mixed-initiative dialog management that significantly improves real-time policy calculation and error handling compared to traditional POMDPs. The GM-based system achieves shorter dialogs and better error recognition, making dialog systems more flexible and robust.

What Problem Does This Paper Address?

Modern dialog systems, especially those involving speech recognition, face a significant challenge: imperfect speech recognition leads to errors that can derail the conversation. Traditional approaches like Partially Observable Markov Decision Processes (POMDPs) can model these systems but often struggle with real-time policy updates and efficient error correction. This paper addresses the need for a dialog management system that is not only flexible enough to adapt to user behavior in real-time but also robust in handling speech recognition errors, leading to more natural and efficient human-computer interactions.

Methods

The authors propose a novel dialog management system based on a Graphical Model (GM). This GM extends the concepts of POMDPs by allowing for real-time calculation of dialog policies. The core idea is to represent the dialog as a sequence of semantic slots, which are binary representations of user intents or information. These slots form dialog states, and transitions between states are modeled as semantic pairs.

Key methodological aspects include:

  • Semantic Slots and States: User utterances are decoded into semantic slots, and combinations of these slots define the dialog state. This allows for a structured representation of the dialog's progress.
  • Trellis-based Dialog Model: The dialog is visualized as a trellis, where states evolve over time. The Viterbi algorithm is employed to find the most likely path through this trellis, effectively determining the optimal sequence of system actions (questions) to reach the user's goal.
  • Real-time Policy Calculation: Unlike POMDPs that might require extensive pre-computation, the GM approach allows for dynamic policy adjustments during runtime, increasing system flexibility.
  • Error Recognition and Correction: The Viterbi algorithm, combined with semantic slot re-estimation, enables the system to detect and potentially correct errors introduced by the speech recognizer. By re-evaluating past states based on new observations, the system can revise its understanding of the dialog.

Datasets

  • ATIS (Air Traveling Information System) Corpus: This is the primary dataset used for both training and evaluation. The authors extracted 21,650 natural spoken questions and answers from ATIS log files.
  • Training Data: 1,585 spoken dialogs from the ATIS corpus were used to learn the parameters of the dialog model.
  • Evaluation Data: The experiments involved 1,990 dialogs for the GM-based system and 2,000 dialogs for the POMDP reference system.

Key Insights and Ideas

  • Flexibility through Real-time Policies: The ability to calculate dialog policies in real-time is a significant advantage, allowing the system to adapt to user behavior and potentially change its strategy mid-dialog.
  • GM as a Robust Alternative to POMDPs: The GM approach offers a compelling alternative to POMDPs, demonstrating comparable or superior performance in key metrics while being more computationally efficient for real-time applications.
  • Viterbi for Error Correction: The application of the Viterbi algorithm is not just for finding the best path but also for implicitly correcting errors by re-evaluating past states based on current observations. This makes the system more resilient to noisy speech recognition.
  • Structured Dialog Representation: The use of semantic slots and a trellis structure provides a clear and manageable way to model complex dialogs, making it easier to understand and debug system behavior.

Key Results

  • Reduced Average Dialog Length: The GM-based system achieved a 31.6% reduction in average dialog length compared to the POMDP reference system, indicating more efficient goal achievement.
  • Improved Error Recognition: The GM-based system demonstrated a 64.4% relative enhancement in sensitivity rate for error recognition capabilities, while maintaining the same specificity rate as the POMDP system.
  • Quantitative Comparison: Table 1 provides a detailed comparison, showing the GM system's higher sensitivity (0.39 vs. 0.18) and relevance (0.50 vs. 0.32) in error detection, while matching the specificity (0.93).

Limitations and Future Work

  • Limitations:
    • The paper notes that the GM's classifier is not based on a variable threshold, which limits its ability to generate a full ROC curve.
  • While the GM system shows improved error recognition, the paper acknowledges that the POMDP model can be optimized further.
  • Future Work:
    • The authors plan to integrate the proposed GM-based approach into a multi-agent framework.
  • They aim to further leverage the system's error recognition capabilities to correct speech recognition errors more effectively.

Summary updated

BibTeX
@inproceedings{schwarzler2009using,
  author = {Stefan Schwärzler and Stefan Maier and Joachim Schenk and Frank Wallhoff and Gerhard Rigoll},
  title = {{Using graphical models for mixed-initiative dialog management systems with realtime Policies}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure