Zero-Resource Knowledge-Grounded Dialogue Generation

Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, Chongyang Tao

Abstract

While neural conversation models have shown great potentials towards generating informative and engaging responses via introducing external knowledge, learning such a model often requires knowledge-grounded dialogues that are difficult to obtain. To overcome the data challenge and reduce the cost of building a knowledge-grounded dialogue system, we explore the problem under a zero-resource setting by assuming no context-knowledge-response triples are needed for training. To this end, we propose representing the knowledge that bridges a context and a response and the way that the knowledge is expressed as latent variables, and devise a variational approach that can effectively estimate a generation model from a dialogue corpus and a knowledge corpus that are independent with each other. Evaluation results on three benchmarks of knowledge-grounded dialogue generation indicate that our model can achieve comparable performance with state-of-the-art methods that rely on knowledge-grounded dialogues for training, and exhibits a good generalization ability over different topics and different datasets.

Introduction

Open-domain chatbots often struggle with "hallucinations" or generic, repetitive responses due to a lack of access to external facts. Knowledge-Grounded Dialogue (KGD) systems aim to solve this by integrating external information. However, traditional training requires massive datasets of (Conversation Context, Knowledge Snippet, Human Response) triples. Collecting these is prohibitively expensive, as humans must manually link specific facts to every conversational turn.

This research introduces ZRKGC (Zero-Resource Knowledge-Grounded Conversation), a framework designed to build KGD systems using only independent piles of random conversations (e.g., Reddit) and factual data (e.g., Wikipedia), without any explicit links between them. This "Zero-Resource" approach significantly lowers the barrier to creating informative dialogue agents.

Methodology

The core of ZRKGC is a probabilistic framework that treats the relationship between context and response as a latent process. The model utilizes two primary latent variables:

  1. $Z_k$ (Latent Knowledge): Represents the specific piece of knowledge bridging the context and the response.
  2. $Z_\alpha$ (Grounding Rate): A variable that determines how much knowledge is expressed. This allows the model to distinguish between social, low-grounding replies and informative, fact-heavy ones.
The graphical model of ZRKGC illustrating the relationships between context (C), response (R), and the latent variables for knowledge (Zk) and grounding rate (Zα).

Variational Learning and Stability

To estimate the model without ground-truth triples, the researchers employed Generalized Expectation-Maximization (EM). A critical innovation is the use of a retrieval-based posterior. Instead of attempting to generate a latent fact—which often leads to training instability and "posterior collapse"—the model uses a retrieval mechanism to select candidates from the knowledge base. The system is built upon the UNILM backbone, leveraging its pre-trained linguistic capabilities for robust grammar and logic.

Experimental Setup

The model was trained using a dialogue corpus from Reddit and a knowledge corpus consisting of over 5.9 million Wikipedia articles. Performance was evaluated across three major KGD benchmarks:

  • Wizard of Wikipedia (WoW): Wikipedia-based conversations.
  • Topical-Chat (TC): Multi-domain conversations (sports, politics, etc.).
  • CMU Document Grounded Conversations (CMU_DoG): Movie-focused dialogues.

Results and Analysis

Despite the lack of labeled training triples, ZRKGC demonstrated performance comparable to supervised state-of-the-art models. On the Wizard of Wikipedia "Seen" test set, ZRKGC achieved an F1 score of 18.7, nearly matching the 19.3 score of the supervised DRD model.

F1 score comparison across multiple benchmarks, showing ZRKGC achieving performance comparable to supervised baselines like DRD.

Key Insights

  • Robustness of Grounding Rate: Modeling $Z_\alpha$ makes the system resilient to noise. It learns that not every sentence requires factual grounding, allowing for more natural social transitions.
  • Superior Generalization: Unlike supervised models that often overfit to their specific training topics, ZRKGC maintains stable performance on "Unseen" topics. This is because it learns the fundamental relationship between context and knowledge rather than memorizing specific triples.
  • Human Preference: In blind evaluations, human judges rated ZRKGC as more fluent and contextually coherent than supervised baselines, likely due to its effective use of the pre-trained UNILM backbone.

Conclusion and Future Work

ZRKGC proves that high-quality knowledge-grounded dialogue systems can be trained without expensive labeled data. While the model is highly fluent, future work will focus on improving the naturalness of fact integration to avoid "awkward" phrasing when weaving knowledge into prose. Additionally, researchers aim to address computational constraints related to token limits in large language models and implement more robust safety filters to mitigate biases inherent in social media and open-web training data.

Summary updated

Abstract

While neural conversation models have shown great potentials towards generating informative and engaging responses via introducing external knowledge, learning such a model often requires knowledge-grounded dialogues that are difficult to obtain. To overcome the data challenge and reduce the cost of building a knowledge-grounded dialogue system, we explore the problem under a zero-resource setting by assuming no context-knowledge-response triples are needed for training. To this end, we propose representing the knowledge that bridges a context and a response and the way that the knowledge is expressed as latent variables, and devise a variational approach that can effectively estimate a generation model from a dialogue corpus and a knowledge corpus that are independent with each other. Evaluation results on three benchmarks of knowledge-grounded dialogue generation indicate that our model can achieve comparable performance with state-of-the-art methods that rely on knowledge-grounded dialogues for training, and exhibits a good generalization ability over different topics and different datasets.
BibTeX
@inproceedings{li2020zero,
  author = {Linxiao Li and Can Xu and Wei Wu and Yufan Zhao and Xueliang Zhao and Chongyang Tao},
  title = {{Zero-Resource Knowledge-Grounded Dialogue Generation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=dAmcGZT1Z2m},
}
Figure