Introduction: The One-to-Many Challenge
In natural human conversation, a single prompt often warrants multiple valid responses. For instance, if a user expresses hunger and asks for a place to eat in the city center, a helpful assistant might suggest a specific restaurant, ask for a preferred cuisine, or inquire about a price range. This inherent flexibility is known as the one-to-many property of dialogue.
Traditional task-oriented dialog systems (TODS) typically fail to capture this diversity. Most models are trained to map a specific conversation context to a single "ground truth" response provided in a dataset. This approach leads to several issues:
- Data Bias: Datasets usually only record one of many possible valid responses for any given turn.
- Skewed Policies: Models tend to over-optimize for the most frequent actions, ignoring valid but less common alternatives.
- Lack of Diversity: The resulting bots often feel repetitive and "boring," lacking the adaptability required for real-world interactions.
Proposed Framework: MADA and DAMD
To address these limitations, this research introduces a novel framework consisting of Multi-Action Data Augmentation (MADA) and the Domain Aware Multi-Decoder (DAMD).
Multi-Action Data Augmentation (MADA)
MADA shifts the training paradigm from learning a single response to learning the entire set of valid actions for a given state. Instead of relying solely on the transcript's specific response, MADA scans the training corpus to identify all instances where a similar dialogue state (comprising conversation history, belief state, and database results) occurred. By aggregating the actions taken across these similar states, the framework creates a "state-to-actions" map. During training, the model is incentivized to learn this entire distribution, effectively de-biasing the dataset and encouraging diversity.
Domain Aware Multi-Decoder (DAMD)
The DAMD architecture implements this multi-action approach through an end-to-end neural network. It utilizes a single context encoder paired with three specialized decoders:
- Belief Span Decoder: Tracks user constraints (e.g., "area=west").
- Action Span Decoder: Determines the system's strategy (e.g., "inform" or "request").
- Response Decoder: Generates the final natural language output.
A key innovation within DAMD is Adaptive Delexicalization. By using generic placeholders (e.g., <v.name>) instead of domain-specific ones, the model can share knowledge across different tasks, such as applying booking logic from hotels to taxis.
Experimental Evaluation
The framework was evaluated on the MultiWOZ dataset, a complex multi-domain benchmark. The DAMD + MADA model achieved state-of-the-art results, reaching a combined score of 102.2, outperforming previous benchmarks like HDSA (99.5) and SFN+RL (93.7).
Beyond automated metrics, human evaluations confirmed the model's superiority in generating diverse and appropriate responses. On Likert scales, the augmented model significantly improved scores for Diversity (3.65 vs. 3.12) and Appropriateness (2.53 vs. 2.50) compared to baselines.
Conclusion
By explicitly modeling the one-to-many nature of dialogue through MADA and DAMD, this work provides a path toward more flexible and human-like task-oriented agents. While future work remains in addressing slot-level diversity and refining label quality, the shift toward structured, multi-action learning represents a significant advancement in dialogue system design.
Summary updated