Improving the Robustness of Knowledge-Grounded Dialogue via Contrastive Learning

TL;DR

The paper tackles robustness in knowledge-grounded dialogue (KGD) when faced with real-world perturbations in dialogue context and knowledge graphs. It introduces EnCo, an entity-guided contrastive learning framework that creates positive samples via entity-guided paraphrasing and negative samples via entity perturbations, optimizing a mixed objective $\mathcal{L} = \mathcal{L}_{ce} + \alpha \mathcal{L}_{ctr}$. Through experiments on three public KGD benchmarks, EnCo achieves state-of-the-art results, demonstrates strong robustness under diverse noises, and shows favorable few-shot and human-evaluated performance. The work highlights a practical pathway to deploying more reliable KGD systems in real-world settings by leveraging entity information to model perturbations explicitly.

Abstract

Knowledge-grounded dialogue (KGD) learns to generate an informative response based on a given dialogue context and external knowledge (\emph{e.g.}, knowledge graphs; KGs). Recently, the emergence of large language models (LLMs) and pre-training techniques has brought great success to knowledge-grounded dialogue. However, when building KGD systems in real applications, there are various real-world noises that are inevitable to face. For example, the dialogue context might involve perturbations such as misspellings and abbreviations. In addition, KGs typically suffer from incompletion and also might contain erroneous and outdated facts. Such real-world noises pose a challenge to the robustness of KGD systems and hinder their applications in the real world. In this paper, we propose an entity-based contrastive learning framework for improving the robustness of KGD. Specifically, we make use of the entity information in a KGD sample to create both its positive and negative samples which involve semantic-irrelevant and semantic-relevant perturbations, respectively. The contrastive learning framework ensures the KGD model is aware of these two types of perturbations, thus generating informative responses with the potentially noisy inputs in real applications. Experimental results on three benchmark datasets show that our method achieves new state-of-the-art performance in terms of automatic evaluation scores, verifying its effectiveness and potentiality. Furthermore, we show that our method can generate better responses than comparison models in both the noisy and the few-shot settings.

Figures (4)

• Figure 1: Illustrations of robustness in knowledge-grounded dialogue. The response in the first dialogue satisfies humans, while those in the second and the third dialogues do not due to misspellings and erroneous facts in the KG, respectively.
• Figure 2: Illustrations of positive samples and negative samples in our entity-based contrastive learning framework.
• Figure 3: Few-shot results on DuConv.
• Figure 4: Results on human study.