# Knowledge Graph-Augmented Language Models for Knowledge-Grounded Dialogue Generation

**Authors:** Minki Kang, Jin Myung Kwak, Jinheon Baek, Sung Ju Hwang  
**arXiv:** [2305.18846](https://arxiv.org/abs/2305.18846)  
**Date:** 2023-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.18846v1/

## Key points

- Language models have made advancements in dialogue generation tasks, but struggle with generating responses that require factual knowledge.
- Some methods leverage facts from Knowledge Graphs (KGs), but do not guarantee the utilization of relevant knowledge.
- The SURGE framework proposes a solution for generating context-relevant and knowledge-grounded dialogues using KGs.
- The framework consists of two main steps: subgraph retrieval and response generation.
- Subgraph retrieval involves retrieving the relevant subgraph from the KG containing necessary factual knowledge.
- Response generation enforces consistency across facts by perturbing their word embeddings conditioned on the retrieved subgraph.
- Contrastive learning is employed to ensure generated responses are faithful to the retrieved subgraphs.
- Experiments on OpendialKG and KOMODIS datasets validate the effectiveness of SURGE in generating high-quality responses reflecting knowledge from KGs.
- A human study involving 46 annotators shows that SURGE outperforms other approaches in terms of consistency, informativeness, and fluency.
- Visualization of graph-text embeddings demonstrates that SURGE generates distinct response embeddings for different subgraphs, emphasizing its effectiveness compared to versions without contrastive learning.
- The proposed SURGE framework addresses the challenge of knowledge-grounded dialogue generation by retrieving context-relevant subgraphs, encoding them with text, and generating natural and informative responses based on the retrieved subgraph.

## Summary

In recent years, language models have made significant advancements in dialogue generation tasks. However, when it comes to generating responses that require factual knowledge, these models still fall short due to a lack of mechanisms for retrieving, encoding, and reflecting knowledge in the generated responses. Some methods have attempted to address this issue by leveraging facts from Knowledge Graphs (KGs), but they do not guarantee that the model will utilize relevant knowledge from the KG. To overcome this limitation, we propose a framework called SUbgraph Retrieval-augmented GEneration (SURGE) for generating context-relevant and knowledge-grounded dialogues using KGs. <SURGE Framework>

The SURGE framework consists of two main steps: subgraph retrieval and response generation. In the subgraph retrieval step, we retrieve the relevant subgraph from the KG that contains the necessary factual knowledge for generating an informed response. In the response generation step, we enforce consistency across facts by perturbing their word embeddings conditioned on the retrieved subgraph. To ensure that our generated responses are faithful to the retrieved subgraphs, we employ contrastive learning. This helps us achieve high similarity between the generated texts and the retrieved subgraphs. <Subgraph Retrieval>

To validate our SURGE framework's effectiveness in generating context-relevant and knowledgeable dialogues, we conducted experiments on OpendialKG and KOMODIS datasets. The results demonstrate that our approach generates high-quality responses that accurately reflect knowledge from KGs. <Human Study>

In addition to quantitative evaluations, we also conducted a human study involving 46 annotators to evaluate our generated responses' quality compared to other approaches such as All Knowledge and Space Efficient. The study evaluated responses based on criteria like consistency, informativeness, and fluency using a Likert-like scale. Our SURGE framework received significantly higher scores in all criteria, providing further evidence of its ability to generate consistent, informative, and fluent responses. <Graph-Text Embeddings>

Furthermore, we visualized the latent space of graph-text embeddings learned by our framework. The visualization demonstrates that SURGE with graph-text contrastive learning generates distinct response embeddings for different subgraphs, unlike the version without contrastive learning which shows less variety in responses for the same dialogue. <Conclusion>

In conclusion, our proposed SURGE framework effectively addresses the challenge of knowledge-grounded dialogue generation by retrieving context-relevant subgraphs, encoding them with text, and generating natural and informative responses based on the retrieved subgraph. Our experiments highlight the contributions of each component in retrieval, encoding, and graph-text representation learning. This work opens up new possibilities for generating informative responses in knowledge graph-based dialogue tasks by emphasizing the importance of retrieving relevant subgraph knowledge rather than relying on all available knowledge graphs.