Developing Interactive Tourism Planning: A Dialogue Robot System Powered by a Large Language Model*

In recent years, large language models (LLMs) have rapidly proliferated and have been utilized in various tasks, including research in dialogue systems. We aimed to construct a system that not only leverages the flexible conversational abilities of LLMs but also their advanced planning capabilities to reduce the speaking load on human interlocutors and efficiently plan trips. Furthermore, we propose a method that divides the complex task of a travel agency into multiple subtasks, managing each as a separate phase to effectively accomplish the task. Our proposed system confirmed a certain level of success by achieving fourth place in the DRC2023[1] preliminaries rounds. We report on the challenges identified through the competition.

Our system advances the conversation based on a pre-determined scenario-based approach. Table I shows the scenarios used, along with methods for controlling the robot’s movements and the display. The scenario consists of five phases as shown in the table, each aimed at accomplishing a subtask in tourist guidance. A large-scale language model is employed for response generation, and different types of prompts are used in each phase. For the large-scale language model, the system uses the API of gpt-4-32k-0613 (referred to as GPT) provided by OpenAI. We used the Streaming API and input the text into the Text-to-Speech (TTS) system every time a punctuation mark appeared, in order to speed up the speech delivery.

We implemented two types of transitions between phases. One is to instruct the system to output a termination sign [END] when the objective of the phase is achieved. When [END] is output, a transition is made to the next phase. This allows, for example, an automatic transition to the next phase when the system has heard everything it needs to hear in the inquiry phase. However, the model often does not output [END], so the first transition condition may not be satisfied. To address this issue, we implemented a second transition that sets a maximum number of turns for each phase and forcibly terminates the phase if the number of turns is exceeded.

For system evaluation, we used android robots set up at the storefronts of travel agencies in Fukuoka and Nagoya, and had people who visited the travel agencies evaluate our system. The evaluation of the system was divided into two parts: the evaluation of the plan and the satisfaction evaluation. Our system achieved a score of 0.85 in the plan evaluation, and an average of 4.41 across nine metrics in the satisfaction evaluation. For the satisfaction evaluation, there was a large variance of 3.44, indicating a significant variation in perceptions among the participants.

In the 2022 Dialogue Robot Competition (DRC2022) and before [2], the task was to recommend one tourist destination from two options within a five-minute dialogue. However, this year the competition has become more challenging due to the increased dialogue duration and a greater number of tourist destinations to choose from. Previous works participated in DRC2022 [3, 4] implemented a scenario-based dialogue system and utilized an LLM for response generation. This paper focuses on leveraging the performance improvements and enhanced controllability of large-scale language models, which have benefited from recent advancements like instruction tuning and scaling up. To increase reliance on the LLM and introduce more flexibility in response generation, we have expanded the granularity of the scenarios. Additionally, in response to the increased complexity of the task, we have also expanded our external knowledge sources, including the use of route searching APIs and model course datasets.

In this paper, we have outlined the proposed system that was entered in DRC2023 and discussed the issues discovered and insights gained through the competition.