Heterogeneous Relational Graph Neural Networks with Adaptive Objective for End-to-End Task-Oriented Dialogue

End-to-end task-oriented dialogue systems, which provide a natural and informative way for human- computer interaction, are gaining more and more attention. The main challenge of such dialogue systems is how to effectively incorporate external knowledge bases into the learning framework. However, existing approaches usually overlook the natural graph structure information in the knowledge base and the relevant information between the knowledge base and the dialogue history, which makes them deficient in handling the above challenge. Besides, existing methods ignore the entity imbalance problem and treat different entities in system responses indiscriminately, which limits the learning of hard target entities. To address the two challenges, we propose Heterogeneous Relational Graph Neural Networks with Adaptive Objective (HRGNN-AO) for end-to-end task-oriented dialogue systems. In the method, we explore effective heterogeneous relational graphs to jointly capture multi perspective graph structure information from the knowledge base and the dialogue history, which ultimately facilitates the generation of informative responses. Moreover, we design two components, shared-private parameterization and hierarchical attention mechanism, to solve the overfitting and confusion problems in the heterogeneous relational graph, respectively. To handle the entity imbalance problem, we propose an adaptive objective, which dynamically adjusts the weights of different target entities during the training process. The experimental results show that HRGNN-AO is effective in generating informative responses and outperforms state-of-the-art dialogue systems on the SMD and extended Multi-WOZ 2.1 datasets. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

End-to-end task-oriented dialogue systems face challenges in effectively incorporating external knowledge bases and addressing entity imbalance. The proposed HRGNN-AO method utilizes heterogeneous relational graphs, shared-private parameterization, and hierarchical attention mechanism to improve response generation, while introducing an adaptive objective to handle entity imbalance and outperform state-of-the-art dialogue systems.