Introduction
Spoken dialog systems, such as automated customer service agents and virtual assistants, frequently operate in challenging acoustic environments. Spontaneous speech and background noise often lead to Word Error Rates (WER) of 20-30%. These recognition failures typically manifest as misunderstandings, where the system extracts the wrong semantic meaning, or non-understandings, where it fails to extract any meaning at all.
Historically, error recovery logic has been hardcoded into specific task domains, making systems difficult to scale and preventing the reuse of conversational behaviors. This paper introduces a modular, task-independent error-handling architecture for the RavenClaw framework. By decoupling recovery strategies from domain logic and employing distributed Markov Decision Processes (MDPs), the system provides a robust, plug-and-play solution for conversational agents.
The RavenClaw Architecture
The RavenClaw framework is built on a strict separation between domain-specific logic and domain-independent conversational skills. The Dialog Task Specification defines the hierarchical structure of the conversation (e.g., booking a room), while the Dialog Engine manages the runtime execution, including the dialog stack and expectation agenda.
This separation allows the Dialog Engine to handle generic conversational tasks—like error recovery—without requiring the developer to specify these behaviors for every new domain. When a task is specified, the engine automatically manages the flow and expectations of the interaction.
Distributed Error Handling
The proposed architecture addresses the scalability issues of traditional Reinforcement Learning (RL) in dialog management. Instead of a single, monolithic MDP that would suffer from state-space explosion, the authors introduce three core components:
- Strategy Library: A collection of reusable agents for recovery. Misunderstandings are handled via explicit or implicit confirmation. Non-understandings utilize a wider arsenal, including AskRepeat, AskRephrase, Yield, and FullHelp.
- Distributed Decision Processes: The system employs "local" MDPs tied to generic concepts. Each piece of information (a "concept") has its own Concept-MDP to manage misunderstandings, while each system request has a Request-MDP for non-understandings. This "divide and conquer" approach ensures that a more complex dialog task simply requires more small MDPs rather than an exponentially larger global state space.
- Gating Mechanism: At each turn, local MDPs evaluate the likelihood of an error. A central gating mechanism aggregates these decisions and dynamically injects the appropriate recovery strategy onto the dialog stack when necessary, pausing the main task until the error is resolved.
Evaluation and Results
The architecture was evaluated through deployment in three live systems: RoomLine (conference room booking), Let's Go Public (bus route information), and Vera (a taskable phone agent).
The results demonstrated high adaptability and robustness across diverse domains:
- RoomLine: Achieved a 75% task success rate despite a combined 30% error rate from misunderstandings and non-understandings.
- Let's Go Public: In a high-noise environment with error rates exceeding 50%, the system maintained a 52% success rate, with error handling engaging in over half of all conversational turns.
- Ease of Authoring: The framework proved accessible to developers with no prior RavenClaw experience, who successfully configured error handling for the Vera system by simply selecting policies and writing prompts.
Conclusion
The RavenClaw error-handling architecture successfully decouples recovery logic from task-specific code. By using distributed MDPs and model tying, the system remains scalable and allows for the reuse of sophisticated error-handling behaviors across diverse domains. While current implementations rely on heuristic policies, the distributed architecture sets the stage for future work in automatically learning optimal recovery behaviors through Reinforcement Learning.
Summary updated