Abstract: We present ParlAI (pronounced ""parley""), an open-source python framework for researching, building, training, evaluating, and deploying open-domain chatbots. ParlAI provides everything you need to be successful in dialogue research: common model implementations, over 80 datasets, readily available state-of-the-art pretrained models, and standardized evaluation frameworks. We discuss the core abstractions of ParlAI, including Agents, Worlds, and Messages, and how these allow unified presentation of model training, model evaluation, and model deployment. We highlight the most important models and features, including support for general chit chat, image commenting, and seamless integration with Amazon Mechanical Turk. We show how you can start training a specialized using GPT-2 with just a single command.
We briefly highlight some of our recent advances in dialogue research with deep learning, including how we are teaching chatbots to be personable, knowledgeable, and empathetic. We discuss our methods for robustly evaluating the quality of research chatbots, how we teach models to respond safely. We show off some of our recently deployed research bots, which you can talk to today!
This tutorial is targeted at a technical audience without an existing background in dialogue research. Familiarity with deep learning is recommended, but not required, for the second half of the talk.
Bio: Stephen Roller is a research engineer at Facebook. He received his PhD in Natural Language Processing at the University of Texas at Austin in 2017. His research interests include natural language processing, deep learning, language generation, and conversational agents.
Stephen Roller, PhD
NLP/ML Researcher & Engineer | Facebook AI Research