# Training Millions of Personalized Dialogue Agents

**Authors:** Pierre-Emmanuel Mazaré, Samuel Humeau, Martin Raison, Antoine Bordes  
**arXiv:** [1809.01984](https://arxiv.org/abs/1809.01984)  
**Date:** 2018-09-06  
**Summary page:** https://www.summarizepaper.com/en/paper/1809.01984v1/

## Key points

- Current dialogue systems lack engagement for users when trained end-to-end without scripted strategies
- Conditioning end-to-end dialogue models on text personas enhances user engagement
- Zhang et al.'s study used a synthetic dataset with limited size (1k personas)
- Mazaré et al. present a new dataset with 5 million personas and 700 million persona-based dialogues
- Training dialogue systems with personas at such a large scale improves performance for end-to-end models
- The dataset also benefits other tasks beyond engagement levels
- Fine-tuning the model using data from Zhang et al.'s study achieves state-of-the-art results across various tasks
- Mazaré et al.'s research introduces a vast and diverse dataset of personas and persona-based dialogues to address limitations of existing dialogue systems
- Training models with personalized backstories has broader applicability in dialogue systems research.

## Summary

In their paper titled "Training Millions of Personalized Dialogue Agents," Pierre-Emmanuel Mazaré, Samuel Humeau, Martin Raison, and Antoine Bordes address the issue of current dialogue systems lacking engagement for users. They highlight that these systems are particularly ineffective when trained end-to-end without relying on proactive reengaging scripted strategies. To tackle this problem, the authors draw inspiration from Zhang et al. 's (2018) work which demonstrated that conditioning end-to-end dialogue models on text personas can significantly enhance user engagement. However, the dataset used in Zhang et al. 's study is synthetic and limited in size consisting of only around 1k different personas. In response to this limitation, Mazaré et al. present a new dataset that offers a substantial improvement in both scale and diversity. Their dataset includes 5 million personas and an impressive 700 million persona-based dialogues. The authors conduct experiments using this extensive dataset and demonstrate that training dialogue systems with personas at such a large scale still leads to improved performance for end-to-end models. Moreover, they go beyond evaluating the impact on engagement levels and show that other tasks also benefit from the wide coverage provided by their dataset. To validate this claim further, Mazaré et al. fine-tune their model using data from Zhang et al. 's study and achieve state-of-the-art results across various tasks. This finding highlights the potential of leveraging their dataset not only for enhancing engagement but also for improving overall system performance. Overall, Mazaré et al. 's research contributes significantly to addressing the limitations of existing dialogue systems by introducing a vast and diverse dataset of personas and persona-based dialogues. Their findings underscore the effectiveness of training models with personalized backstories while showcasing its broader applicability across multiple tasks within dialogue systems research.