# XPersona: Evaluating Multilingual Personalized Chatbot

**Authors:** Zhaojiang Lin, Zihan Liu, Genta Indra Winata, Samuel Cahyawijaya, Andrea Madotto, Yejin Bang, Etsuko Ishii, Pascale Fung  
**arXiv:** [2003.07568](https://arxiv.org/abs/2003.07568)  
**Date:** 2020-03-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.07568v1/

## Key points

- Authors address the need for personalized dialogue systems in multiple languages
- Existing dialogue agents are mostly designed for English conversations, limiting their usage in other languages
- Authors propose XPersona, a multi-lingual extension of Persona-Chat
- XPersona dataset includes persona conversations in six different languages apart from English
- Authors build and evaluate multilingual personalized agents using this dataset
- Multilingual trained models outperform translation-pipeline models and perform on par with monolingual models
- Cross-lingual trained models achieve inferior performance compared to other models, highlighting challenges of cross-lingual conversation modeling
- Authors hope to accelerate research in multilingual dialogue systems by releasing their dataset and baselines
- Work contributes to improving human-machine interaction by enabling personalized chatbots in various languages

## Summary

In their paper titled "XPersona: Evaluating Multilingual Personalized Chatbot," authors Zhaojiang Lin, Zihan Liu, Genta Indra Winata, Samuel Cahyawijaya, Andrea Madotto, Yejin Bang, Etsuko Ishii and Pascale Fung address the need for personalized dialogue systems in multiple languages. They highlight that existing dialogue agents are mostly designed for English conversations which limits their usage in other languages. To overcome this limitation, the authors propose XPersona; a multi-lingual extension of Persona-Chat. The XPersona dataset includes persona conversations in six different languages apart from English. The authors use this dataset to build and evaluate multilingual personalized agents. They experiment with both multilingual and cross-lingual trained baselines and compare them against monolingual models and translation-pipeline models using automatic and human evaluation methods. The experimental results demonstrate that the multilingual trained models outperform the translation-pipeline models and perform on par with the monolingual models. This indicates that having a single model across multiple languages provides an advantage. However, state-of-the-art cross-lingual trained models achieve inferior performance compared to other models highlighting the challenges of cross-lingual conversation modeling. The authors hope that by releasing their dataset and baselines they can accelerate research in multilingual dialogue systems. Overall their work contributes to improving human-machine interaction by enabling personalized chatbots in various languages.