Method for Extracting User Needs With Chat-GPT for Dialogue Recommendation

Table of Links

Abstract & Introduction

Related works

Method

Experiment

Conclusion & References

3 Method

Our framework consists of three modules, "Controller" "Assistant" and "User Simulator" and is outlined in1. This chapter describes the structure of the framework, the role of each module, and the relationship between them.

3.1 Controller and Assistant

This subsection describes two of the three elements: the controller and the assistant. First, the assistant is responsible for interacting directly with the user, generating questions for the user that elicit user characteristics and needs. The controller does not interact with the user, but is responsible for monitoring the assistant’s speech and controlling the interaction between the user and the assistant from a higher level. The flow of the system is as follows As shown in 1, the query generated by the assistant is passed to the controller, who determines whether the content of the query is appropriate in terms of the dialogue scenario and communicates it to the assistant.

In Chat-GPT, the dialogue schema is predetermined, and the flow of dialogue can be configured according to the schema by inputting it into the prompts. However, this alone will fix the content of the dialogue and lead to a decrease in user satisfaction in terms of personalization of the dialogue. Therefore, instead of a hard constraint such as specifying the schema in the prompt, we can change it to a soft constraint such as instructions on question generation to the dialog module by the dialog management module, thereby allowing more freedom in the content of the dialog and realizing dynamic dialog. The following are prompts from the controller and assistant on the subject of apartment property recommendation.

3.2 User simulator

In making our system work, how we implement user utterances is very important. In the case of general interactive recommendation, user speech is often implemented using a publicly available dialogue recommendation dataset. Since the data contained in a dialogue recommendation dataset is a pair of a question by the user and a response by the system, and there is no pair data of a question by the system and a response by the user, it is difficult to experiment with a public dataset. Therefore, we attempt to solve this problem by simulating users using Chat-GPT. We have created a module that incorporates the user’s gender, age, and occupation into the prompts and outputs the user’s responses, which act as input to the assistant’s utterances. This allows us to validate a dynamic dialogue system between the assistant and the controller. Below are the prompts for the user simulator on the subject of apartment property recommendations as well as the controller and assistant.