[Paper Review] Zero-Resource Knowledge-Grounded Dialogue Generation

Linxiao Li, Can Xu+4 more|arXiv (Cornell University)|Aug 29, 2020

Topic Modeling76 references50 citations

TL;DR

The paper proposes ZRKGC, a zero-resource, variational double-latent model that grounds dialogue with external knowledge using retrieval-based latent knowledge and a grounding rate, achieving competitive results without crowd-sourced knowledge-grounded training.

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.

Motivation & Objective

• Motivate knowledge-grounded dialogue generation without requiring context-knowledge-response triples for training.
• Introduce a double latent-variable framework (latent knowledge Zk and grounding rate Za) to bridge context and response.
• Develop a variational learning approach with retrieval-based posterior for Zk to enable efficient training.
• Incorporate knowledge selection and mutual information losses to improve grounding expressiveness and stability.
• Demonstrate generalization across topics and datasets on three benchmarks.

Proposed method

• Formulate p(R|C,K) with two latent variables Zk (knowledge) and Za (grounding rate) in a probabilistic framework.
• Use a retrieval-based posterior q(Zk|C,R) that selects from top-l knowledge candidates retrieved by a relevance model.
• Backbone generation uses UNILM to model p(R|C,Zk,Za).
• Introduce a knowledge selection model to constrain input size under model capacity.
• Incorporate a mutual information loss to encourage Za to capture knowledge expression.
• Optimize with Generalized EM (E-step with q, M-step with p) and use Gumbel-softmax for differentiable token sampling.

Experimental results

Research questions

• RQ1 Can knowledge-grounded dialogue generation be learned in a zero-resource setting without context-knowledge-response training triples?
• RQ2 Does a double latent-variable model (knowledge grounding and grounding rate) improve generation quality and control over knowledge use?
• RQ3 How well does retrieval-based posterior learning perform compared to fully generative posteriors in this task?
• RQ4 What is the impact of knowledge selection and mutual information losses on performance and grounding controllability?
• RQ5 How does ZRKGC generalize across topics and datasets compared to state-of-the-art methods?

Key findings

• ZRKGC achieves competitive F1 scores across Wizard Seen, Wizard Unseen, Topical-Freq, Topical-Rare, and CMU_DoG benchmarks, comparable to or better than several baselines.
• ZRKGC shows strong generalization with little performance drop between seen and unseen topics.
• Retrieval-posterior learning yields tighter ELBO and better F1 than generative posterior variants in ablations.
• Knowledge selection and mutual information losses contribute to controllability and stability of grounding expression.
• Human judgments indicate ZRKGC produces more fluent and coherent responses than a competitive baseline, though knowledge integration remains challenging.