UniRQR: A Unified Model for Retrieval Decision, Query, and Response Generation in internet-based knowledge dialogue systems

Knowledge-based dialogue systems with internet retrieval have recently attracted considerable attention from researchers. Compared to traditional knowledge dialogue systems, those with internet retrieval capabilities overcome a major limitation: the inability to ensure the timeliness of knowledge. This advancement provides greater practical application value. Internet-based dialogue systems can be typically segmented into three tasks: Retrieval Decision, Query Generation, and Response Generation. However, previous research faces two main issues. First, most of studies only focus on the latter two tasks, assuming that all conversations require external knowledge to continue, neglecting the critical step of Retrieval Decision. This assumption often leads to an over-dependence on external knowledge. Second, these studies often model the two tasks separately, employing distinct models for query generation and response generation respectively. UniRQR addresses these oversights by employing a single unified model facilitated by prompt and multi-task learning approaches. Integrating these functions, UniRQR leverages the full potential of pre-trained models and fully capitalizes on the synergistic relationship among the three tasks, significantly reducing complexity and costs in comparison to previous research that deployed multiple models. Consequently, UniRQR can decide whether retrieval is necessary, and generate retrieval queries and responses. The experimental outcomes reveal that our unified model not only outperforms state-of-the-art (SOTA) systems in the majority of cases but also achieves this while avoiding the deployment of separate, specialized models for each task, significantly reducing the resources required for deployment. What is more, we conducted extensive experiments to investigate the mutual enhancement among the three tasks in our model. •We propose UniRQR model for Internet-based dialogue systems.•UniRQR is a unified model that manages three tasks in Internet dialogue systems.•UniRQR smartly discerns when to seek external knowledge from search engine.•UniRQR achieves SOTA results with simple design and small model size.•UniRQR is versatile for real-world deployment.