CoLV: A Collaborative Latent Variable Model for Knowledge-Grounded Dialogue Generation

Haolan Zhan, Lei Shen, Hongshen Chen, Hainan ZhangEMNLP 2021 · OpenReview

Abstract

Haolan Zhan, Lei Shen, Hongshen Chen, Hainan Zhang. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.

2021.

Introduction

Knowledge-grounded dialogue generation (KGDG) aims to enhance conversational agents by enabling them to leverage external information sources, such as Wikipedia, to provide informative and contextually relevant responses. Traditional approaches typically treat this as a disjoint two-step pipeline: first, selecting a relevant piece of knowledge (Knowledge Selection), and subsequently generating a response based on that selection (Response Generation).

However, this independent pipeline approach suffers from two primary limitations. First, it ignores the inherent coupling between knowledge selection and response generation; in human discourse, the choice of facts and the manner of expression are deeply intertwined. Second, deterministic models often lack diversity, tending toward "safe" or generic responses that fail to explore the varied ways a single piece of knowledge can be integrated into a conversation.

The CoLV Framework

To address these challenges, we propose the Collaborative Latent Variable (CoLV) model. CoLV utilizes a Variational Auto-Encoder (VAE) framework to model the dialogue process through two interconnected latent spaces:

  1. Knowledge Latent Space ($z_k$): A discrete space following a Categorical distribution that governs the selection of external knowledge.
  2. Response Latent Space ($z_r$): A continuous space following a Gaussian distribution that captures the stylistic and intentional variations of the response.
The CoLV architecture featuring BERT encoders and collaborative latent modules for knowledge selection and response generation.

The defining feature of CoLV is the collaborative dependency: the response latent space is conditioned on the knowledge latent space. This ensures that the generated response is not merely generic chatter but is specifically grounded in the selected factual content.

Furthermore, CoLV introduces a Heuristic-based Knowledge Selection module. Unlike traditional models that select discrete sentences, CoLV treats the knowledge pool as a continuous text paragraph and employs a pointer-style mechanism to identify specific spans. This fine-grained approach allows for more precise grounding than sentence-level selection.

Experimental Results

The model was evaluated on the Wizard of Wikipedia (WoW) and Holl-E benchmarks. CoLV consistently outperformed state-of-the-art baselines, including DukeNet and PIPM, across multiple metrics.

Performance comparison of CoLV against baselines across WoW (Seen/Unseen) and Holl-E datasets, showing accuracy relative to knowledge span percentage.

Key findings include:

  • Accuracy: On the WoW "Seen" test set, CoLV achieved a knowledge selection accuracy of 30.12%, surpassing the previous state-of-the-art of 27.75%.
  • Diversity and Quality: The model demonstrated superior performance in BLEU-4 and ROUGE scores, alongside higher Dist-2 scores, indicating more varied and human-like language generation.
  • Human Preference: In side-by-side evaluations, human annotators preferred CoLV's responses significantly more often than those from baseline models (e.g., a 67% win rate against a standard S2SA model).

Conclusion and Future Work

CoLV demonstrates that treating knowledge selection and response generation as collaborative tasks significantly improves the quality and diversity of grounded dialogues. While the model shows robust performance, challenges remain in generalizing to "Unseen" topics, where performance drops compared to "Seen" datasets. Future work will explore extending this collaborative framework to multi-turn reasoning, where the model maintains a "memory" of latent variables across a conversation.

Summary updated

BibTeX
@inproceedings{zhan2021colv,
  author = {Haolan Zhan and Lei Shen and Hongshen Chen and Hainan Zhang},
  title = {{CoLV: A Collaborative Latent Variable Model for Knowledge-Grounded Dialogue Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  url = {https://openreview.net/forum?id=Cs1mmBCG5H0},
}
Figure