Abstract
In task-oriented dialogue systems, response generation from meaning representations (MRs) often suffers from limited training examples, due to the high cost of annotating MR-to-Text pairs. Previous works on self-training leverage fine-tuned conversational models to automatically generate pseudo-labeled MR-to-Text pairs for further fine-tuning. However, some self-augmented data may be noisy or uninformative for the model to learn from. In this work, we propose a two-phase self-augmentation procedure to generate high-quality pseudo-labeled MR-to-Text pairs: the first phase selects the most informative MRs based on model's prediction uncertainty; with the selected MRs, the second phase generates accurate responses by aggregating multiple perturbed latent representations from each MR. Empirical experiments on two benchmark datasets, FewShotWOZ and FewShotSGD, show that our method generally outperforms existing self-training methods on both automatic and human evaluations.
Introduction
Natural Language Generation (NLG) in task-oriented dialogue systems involves converting structured Meaning Representations (MRs)—such as inform(food=seafood)—into fluent, natural language responses. While effective, these systems typically require large volumes of annotated data, which is expensive and time-consuming to produce. In "few-shot" scenarios, where only 30-50 examples per domain are available, models often struggle to generalize.
Self-training is a common strategy to mitigate data scarcity by using a model to label its own data. However, standard self-training often introduces two types of problematic data: noisy data, where the model hallucinates facts, and uninformative data, where the model selects examples it already understands well, leading to redundant training. This research introduces SA² (Two-phase Self-augmentation), a framework designed to identify the "sweet spot" of informative yet reliable data for augmentation.
Methodology: The SA² Framework
The SA² framework operates in two distinct phases to ensure the quality and utility of pseudo-labeled data.
Phase I: Uncertainty-based Data Selection
Instead of selecting unlabeled data randomly, SA² utilizes Monte Carlo (MC) Dropout to estimate the model's confidence. By performing multiple forward passes with different dropout masks, the system calculates two key metrics:
- Predictive Mean ($\mathbb{E}[p_\theta]$): A high mean suggests the input is "in-domain" and familiar to the model.
- Predictive Variance ($Var[p_\theta]$): A high variance indicates the model is sensitive to the input, suggesting it has not yet mastered the specific pattern.
By selecting data with both high mean and high variance, SA² identifies examples that are relevant to the domain but still offer significant learning potential.
Phase II: Response Refinement
To ensure the generated labels are high-quality, SA² employs a refinement technique inspired by contrastive learning. The model generates multiple latent representations (internal vector states) using different dropout masks and aggregates them. This averaged representation is then decoded into a final response, acting as a denoiser to produce more stable and accurate pseudo-labels than single-pass generation.
Experimental Results
The researchers evaluated SA² on two major few-shot benchmarks: FewShotWOZ (7 domains, ~50 examples each) and FewShotSGD (16 domains, ~35 examples each). The method was compared against state-of-the-art baselines, including SC-GPT and AUG-NLG.
As shown in the results for FewShotWOZ, SA² consistently outperformed baselines. In the Restaurant domain, it achieved a BLEU score of 36.48, significantly higher than the 34.62 achieved by SC-GPT. Crucially, SA² also reduced the Slot Error Rate (ERR), indicating that the refined responses were more factually accurate. On the more complex FewShotSGD dataset, SA² outperformed baselines in 14 out of 16 domains.
Human evaluations further supported these findings, with judges rating SA² responses as more informative and natural, nearly reaching the quality of human-written ground truth.
Conclusion
SA² demonstrates that uncertainty-aware selection combined with latent representation aggregation can significantly boost the performance of few-shot dialogue generation. By focusing on the most informative data and refining the resulting pseudo-labels, the framework provides a streamlined, effective path for training robust NLG modules with minimal human intervention. Future work may explore more computationally efficient decoding methods to reduce the overhead of multiple forward passes during the augmentation phase.
Summary updated