Exploring Prompt-based Few-shot Learning for Grounded Dialog Generation

TL;DR

This work tackles the scarcity of grounded dialog data by introducing a prompt-based few-shot framework for grounded dialog generation (GDG). It shows that distinguishing input constructs (grounding source vs. context) via continuous or discrete prompts significantly boosts performance, with discrete prompts offering robust, practical benefits. The study demonstrates that prompted language models (e.g., GPT-2, T5) outperform conversational models across three GDG tasks, and that model pre-training and size critically influence prompting effectiveness. These findings provide actionable guidance for building GDG systems with limited data and inform choices of backbone models and prompting strategies. Overall, the paper offers a prompt-based lens for GDG and highlights the importance of backbone selection in achieving grounded dialogue capabilities in low-resource settings.

Abstract

Dialog models can be greatly strengthened through grounding on various external information, but grounded dialog corpora are usually not naturally accessible. In this work, we focus on the few-shot learning for grounded dialog generation (GDG). We first propose a simple prompting method for GDG tasks, where different constructs of model input, such as the grounding source and the conversation context, are distinguished through continuous or discrete prompts. On three typical GDG tasks, we empirically demonstrate and analyze in-depth the effectiveness of our method. We then conduct extensive experiments to thoroughly investigate how our prompting method works with different pre-trained models. We show that prompted language models perform superiorly to conversational models, and further analyze various factors that influence the effects of prompting. Overall, our work introduces a prompt-based perspective to the few-shot learning for GDG tasks, and provides valuable findings and insights for future research.

Figures (8)

• Figure 1: Overview of the effects of prompting. Model performance is measured by BLEU-2. Full experimental results are shown in Table \ref{['tab:comparison']}. Detailed experimental setups are described in § \ref{['sec:setup']}.
• Figure 2: Examples of grounded dialogs from Wizard-of-Wikipedia dinan2018wizard, PersonaChat zhang-etal-2018-personalizing and ESConv liu-etal-2021-towards, which are grounded on the Wikipedia knowledge (left), the persona profile (middle) and the emotional support strategies (right) respectively. The parts of grounding sources that the utterances are engaged in are marked in blue.
• Figure 3: Illustration of our prompting method. Our method inserts indicative tokens (middle, Continuous Prompts) or textual descriptions (lower, Discrete Prompts) into input sequences to distinguish different input constructs. For the discrete prompts, we also prepend input sequences with task instructions. All the model parameters are trained with the negative log likelihood loss on the target responses (the end token, either "<eos>" or the line break "\\ n", is underlined).
• Figure 4: Applying our prompting method to PersonaChat.
• Figure 5: Perturbed discrete prompts.