Unified Approach for Scalable Task-Oriented Dialogue System

Task-oriented dialogue (TOD) systems are currently the subject of extensive research owing to their immense significance in the fields of human-computer interaction and natural language processing. These systems assist users to accomplish certain tasks efficiently. However, most commercial TOD systems rely on handcrafted rules and offer functionalities in a single domain. These systems perform well but are not scalable to adapt multiple domains without manual efforts. Pretrained language models (PLMs) have been popularly applied to enhance these systems via fine-tuning. Recently, large language models (LLMs) have made significant advancements in this field but lack the ability to converse proactively in multiple turns, which is an essential parameter for designing TOD systems. To address these challenges, this paper initially studies the impact of language understanding on the overall performance of a TOD system in a multi-domain environment. Furthermore, to design an efficient TOD system, we propose a unified approach by leveraging LLM with reinforcement learning (RL) based dialogue policy. The experimental results demonstrate that a unified approach using LLM is more promising for scaling the capabilities of TOD systems with prompt adaptive instructions with more user friendly and human-like response generation.