# PRODIGy: a PROfile-based DIalogue Generation dataset

**Authors:** Daniela Occhipinti, Serra Sinem Tekiroglu, Marco Guerini  
**arXiv:** [2311.05195](https://arxiv.org/abs/2311.05195)  
**Date:** 2023-11-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.05195v1/

## Key points

- Authors explore benefits of providing dialogue agents with profile representations to enhance consistency and coherence in conversations
- Existing profile-based dialogue datasets often contain simplistic or challenging-to-collect profiles
- Proposed unified framework integrates standard and sophisticated profile representations aligned with speaker attributes like communication style, biographies, and personality
- Profile-based models show superior generalization capabilities compared to models trained solely on dialogues in both in-domain and cross-domain scenarios
- Human evaluation confirms preference for responses consistent with speaker's profile and conversation context

## Summary

, , , , 

In their work titled "PRODIGy: a PROfile-based DIalogue Generation dataset," authors Daniela Occhipinti, Serra Sinem Tekiroglu, and Marco Guerini explore the benefits of providing dialogue agents with profile representations to enhance consistency and coherence in conversations. They note that existing profile-based dialogue datasets often contain either simplistic and dialogue-specific explicit profiles or challenging-to-collect implicit representations. To address this limitation, the authors propose a unified framework that integrates both standard and sophisticated profile representations by aligning each dialogue with various speaker attributes such as communication style, biographies, and personality. Through their new resource, the authors evaluate several baseline models constructed using generative language models with different profile configurations. Their automatic evaluation reveals that profile-based models exhibit superior generalization capabilities compared to models trained solely on dialogues, across both in-domain and cross-domain scenarios. These findings hold true for fine-tuned models and instruction-based Language Models (LLMs). Human evaluation further confirms a preference for responses that are consistent with both the speaker's profile and the context of the conversation. To address privacy concerns, the experiments are conducted under two configurations: inter-character and intra-character. In inter-character setups, the language model stores character information within its internal representation, while in intra-character settings, personal data is not retained but only utilized during inference. This study highlights the importance of incorporating comprehensive profile representations in training dialogue agents for improved conversational outcomes.