APSIPA Transactions on Signal and Information Processing > Vol 12 > Issue 1

Advances and Challenges in Multi-Domain Task-Oriented Dialogue Policy Optimization

Mahdin Rohmatillah, EECS International Graduate Program, National Yang Ming Chiao Tung University, Taiwan, Jen-Tzung Chien, Institute of Electrical and Computer Engineering, National Yang Ming Chiao Tung University, Taiwan, jtchien@nycu.edu.tw
Suggested Citation
Mahdin Rohmatillah and Jen-Tzung Chien (2023), "Advances and Challenges in Multi-Domain Task-Oriented Dialogue Policy Optimization", APSIPA Transactions on Signal and Information Processing: Vol. 12: No. 1, e37. http://dx.doi.org/10.1561/116.00000132

Publication Date: 05 Sep 2023
© 2023 M. Rohmatillah and J.-T. Chien
Multi-domain task-oriented dialogue systemdialogue policy optimizationreinforcement learningimitation learningdialogue act predictionword-level policy learning


Open Access

This is published under the terms of CC BY-NC.

Downloaded: 699 times

In this article:
Multi-Domain Task-Oriented Dialogue System 
Policy Optimization in Dialogue Act Level 
Policy Optimization in Word Level 
Challenges and Difficulties 


Developing a successful dialogue policy for a multi-domain task-oriented dialogue (MDTD) system is a challenging task. Basically, a desirable dialogue policy acts as the decision-making agent who understands the user’s intention to provide suitable responses within a short conversation. Furthermore, offering the precise answers to satisfy the user requirements makes the task even more challenging. This paper surveys recent advances in multi-domain task-oriented dialogue policy optimization and summarizes a number of solutions to policy learning. In particular, the case study on the task of travel assistance using the MDTD dataset based on MultiWOZ containing seven different domains is investigated. The dialogue policy optimization methods, categorized into dialogue act level and word level, are systematically presented. Moreover, this paper addresses a number of challenges and difficulties including the user simulator design and the dialogue policy evaluation which need to be resolved to further enhance the robustness and effectiveness in multi-domain dialogue policy representation.