Dialogue system based on pre-trained model

Abstract

This study builds a task-oriented dialogue system using the pre-trained model DialoGPT.It intends to examine the adaptability of fine-tuning pre-trained model in task-oriented dialogue systems. Therefore, a hotel reservation dataset was selected as the training corpus to fine-tune the DialoGPT-medium version in a local environment. The constructed dialogue system adopts a frontend and backend separated architecture. The frontend is designed for displaying the user interface and interacting with the user. The backend is designed for managing the dialogue context and calling the model to generate responses.

The BERTScore F1 and Perplexity metrics are applied to assess the performance of the fine-tuned DialoGPT. They compare the generation quality of a model from two perspectives, semantic similarity and language fluency. The results provide evidence that the fine-tuned DialoGPT outperforms the original model in both semantic understanding and language generation fluency.

This study demonstrates that pre-trained models can achieve better performance in task-oriented dialogue systems through fine-tuning. However, due to the limitations of the hardware resources and dataset size, the generalization ability of the model still could be improved. Nevertheless, this study provides a valuable reference for the design of future task-oriented dialogue systems.