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

Xueliang Zhao, Wei Wu+4 more|arXiv (Cornell University)|Feb 24, 2020

Topic Modeling46 references84 citations

TL;DR

The paper presents a disentangled, pre-trained decoder for knowledge-grounded dialogue generation that achieves state-of-the-art results with only a fraction of training data by separating language, context, and knowledge grounding components and pre-training them on ungrounded data and plain text.

ABSTRACT

Responding with knowledge has been recognized as an important capability for an intelligent conversational agent. Yet knowledge-grounded dialogues, as training data for learning such a response generation model, are difficult to obtain. Motivated by the challenge in practice, we consider knowledge-grounded dialogue generation under a natural assumption that only limited training examples are available. In such a low-resource setting, we devise a disentangled response decoder in order to isolate parameters that depend on knowledge-grounded dialogues from the entire generation model. By this means, the major part of the model can be learned from a large number of ungrounded dialogues and unstructured documents, while the remaining small parameters can be well fitted using the limited training examples. Evaluation results on two benchmarks indicate that with only 1/8 training data, our model can achieve the state-of-the-art performance and generalize well on out-of-domain knowledge.

Motivation & Objective

• Motivate knowledge-grounded dialogue systems that work in low-resource settings where limited knowledge-grounded training data is available.
• Propose a disentangled response decoder that isolates knowledge-grounded parameters from the rest of the model so the majority can be learned from ungrounded dialogues and plain text.
• Show that pre-training most parameters on large ungrounded data enables strong performance with as little as 1/8 of the labeled data across benchmarks and generalizes to out-of-domain knowledge.

Proposed method

• Encode context and knowledge with two separate encoders (context encoder and knowledge encoder).
• Decompose the decoder into three independently trained components: a language model, a context processor, and a knowledge processor, coordinated by a decoding manager.
• Use a Gumbel-softmax-based decoding manager to select which component predicts each word during training and inference.
• Ground the knowledge processor with a hierarchical attention mechanism over document sentences and words.
• Pre-train: language model/context encoder on Reddit data; knowledge encoder on Wikipedia; fine-tune remaining components on small, domain-specific grounded data.
• Train with maximum likelihood, fixing pre-trained parameters to simulate a low-resource adaptation scenario.

Experimental results

Research questions

• RQ1 Can knowledge-grounded dialogue generation be effectively learned when only a small amount of grounded training data is available?
• RQ2 Does a disentangled decoder facilitate leveraging large amounts of ungrounded dialogue and plain text for pre-training to improve low-resource grounding performance?
• RQ3 How does pre-training different components affect performance on in-domain vs. out-of-domain knowledge?

Key findings

• The proposed model achieves state-of-the-art performance with as little as 1/8 of the training data on Wizard of Wikipedia and CMU DoG benchmarks.
• On Wizard, the model markedly outperforms baselines on out-of-domain documents despite using far less training data; it remains competitive in-domain.
• The model demonstrates strong generalization to out-of-domain knowledge while maintaining robust performance as training data decreases.
• Pre-training the majority of parameters is crucial for low-resource grounding; removing pre-training degrades performance, especially on out-of-domain data.
• Fine-tuning pre-trained parameters helps when data is plentiful but can hurt in very low-resource settings, supporting fixing pre-trained parameters and adapting only a small subset.