Training Millions of Personalized Dialogue Agents

TL;DR

This work extends end-to-end dialogue modeling by introducing a massive Reddit-derived dataset of 5 million personas and 700 million persona-based dialogues, enabling training of persona-conditioned dialogue agents at scale. It shows that conditioning on personas improves response selection across encoder architectures and that larger, well-chosen personas further boost performance. The study also demonstrates strong transfer learning gains, achieving state-of-the-art results on the Persona-chat benchmark by pretraining on Reddit and fine-tuning on Persona-chat, while highlighting challenges in cross-dataset transfer. Overall, the dataset and findings suggest substantial practical impact for building more engaging, personalized conversational agents and indicate directions for refining persona selection and adaptation strategies.

Abstract

Current dialogue systems are not very engaging for users, especially when trained end-to-end without relying on proactive reengaging scripted strategies. Zhang et al. (2018) showed that the engagement level of end-to-end dialogue models increases when conditioning them on text personas providing some personalized back-story to the model. However, the dataset used in Zhang et al. (2018) is synthetic and of limited size as it contains around 1k different personas. In this paper we introduce a new dataset providing 5 million personas and 700 million persona-based dialogues. Our experiments show that, at this scale, training using personas still improves the performance of end-to-end systems. In addition, we show that other tasks benefit from the wide coverage of our dataset by fine-tuning our model on the data from Zhang et al. (2018) and achieving state-of-the-art results.

Figures (1)

• Figure 1: Persona-based network architecture.