Abstract
Kun Qian, Zhou Yu. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.
2019.
Building task-oriented dialog systems typically requires large amounts of annotated data for each specific domain. This paper proposes a framework called DAML (Domain Adaptive Meta-Learning) that enables systems to "learn to learn," allowing rapid adaptation to new domains with minimal data. By applying Model-Agnostic Meta-Learning (MAML) to a sequence-to-sequence dialog model, the system finds a parameter initialization that is primed for quick fine-tuning.
Introduction
The development of task-oriented dialog systems—such as those used in virtual assistants—is hindered by high data requirements. Teaching a bot to handle a new domain (e.g., moving from restaurant booking to movie ticketing) usually necessitates thousands of domain-specific annotated conversations. Traditional transfer learning often fails to bridge the gap between distinct domains, as models can become overly specialized to the source domain's logic. This paper addresses the challenge of data scarcity by enabling one-shot or few-shot domain adaptation.
Methods: Domain Adaptive Meta-Learning (DAML)
The DAML framework integrates two primary components: a sequence-to-sequence base model and a meta-learning training strategy.
Base Model: Sequicity
DAML utilizes the Sequicity architecture, which simplifies dialog management by using a single end-to-end model. It introduces a belief span—a text string tracking user constraints (e.g., cuisine: Italian). This approach avoids rigid database schemas and facilitates easier cross-domain transfer. Additionally, a CopyNet mechanism allows the model to handle out-of-vocabulary words by copying them directly from the user input.
Meta-Learner: MAML
The core innovation is the application of Model-Agnostic Meta-Learning (MAML). Instead of optimizing for performance on a single task, MAML optimizes for adaptability.
As shown in the figure above, while classic gradient updates focus on a single training set, the meta-learning update involves a two-step process. During training, the model performs "mock" adaptations across multiple source domains (the inner loop) and then updates global parameters based on the collective performance of these adapted models (the outer loop). This results in an initialization that is highly sensitive to the unique features of new, unseen tasks.
Experimental Setup
The researchers evaluated DAML using the SimDial dataset, which is designed for cross-domain evaluation:
- Source Domains (Rich Resource): Restaurant, Weather, and Bus information search.
- Target Domains (Low Resource): Movie information search (completely unseen logic), and variations of the restaurant domain with unseen slots or different speaking styles (NLG).
Results and Analysis
DAML demonstrates significant improvements in both performance and efficiency compared to zero-shot (ZSDG) and standard transfer learning baselines.
Key findings from the experimental results include:
- Superior Task Completion: In the "New Domain" (Movie) scenario, DAML achieved an Entity F1 score of 66.2, representing a 25.8% relative improvement over the ZSDG baseline.
- One-Shot Capability: Even in a "one-shot" scenario (adapting using only a single dialogue), DAML outperformed the zero-shot baseline in almost every category.
- Training Efficiency: DAML converges much faster than traditional methods. For instance, in the "New Domain" test, DAML reached its optimum in 4.5 to 5.8 epochs, whereas transfer learning required 5.6 to 14.2 epochs.
Discussion and Future Work
While DAML effectively addresses data scarcity, certain challenges remain. The model occasionally struggles with "unknown" tokens not present in the training vocabulary and can fail to correctly update belief spans during complex, multi-turn user corrections. Future research directions include applying meta-learning to Reinforcement Learning-based dialog systems to further enhance robustness and handling more complex conversational dynamics.
Summary updated