DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation
This work advances open-domain dialogue systems by providing a publicly available pre-trained model to improve response relevance and consistency, though it is incremental as it builds on existing transformer architectures.
The authors tackled conversational response generation by training DialoGPT, a large-scale generative model on Reddit data, achieving performance close to human levels in single-turn dialogues with strong automatic and human evaluations.
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.