DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation

TL;DR

This paper introduces DialoGPT, a large-scale pre-trained transformer model for conversational response generation trained on Reddit data, achieving near-human performance and improving response relevance and consistency in dialogue systems.

Contribution

The paper presents DialoGPT, a novel large-scale pre-trained model specifically designed for open-domain dialogue, with publicly available training pipeline and model.

Findings

DialoGPT achieves near-human performance in single-turn dialogue.

Responses generated are more relevant and contextually consistent.

The model outperforms strong baseline systems in automatic and human evaluations.

Abstract

We present a large, tunable neural conversational response generation model, DialoGPT (dialogue generative pre-trained transformer). Trained on 147M conversation-like exchanges extracted from Reddit comment chains over a period spanning from 2005 through 2017, DialoGPT extends the Hugging Face PyTorch transformer to attain a performance close to human both in terms of automatic and human evaluation in single-turn dialogue settings. We show that conversational systems that leverage DialoGPT generate more relevant, contentful and context-consistent responses than strong baseline systems. The pre-trained model and training pipeline are publicly released to facilitate research into neural response generation and the development of more intelligent open-domain dialogue systems.