Abstract: We present ParlAI (pronounced "parley"), an all-in-one dialogue and chatbot research platform. ParlAI provides tools for all aspects of conversational research, including allowing one to develop novel models; train and test models on a wide variety of existing datasets; compare to standard baselines; collect rich new datasets and conduct human evaluations using Mechanical Turk, and deploy models to users over Facebook Messenger.
We present two research projects built using ParlAI. The first paper we present, "What Makes a Good Conversation?", compares two methods for controllable text generation in neural sequence models, and applies them to chit chat. We consider four controllable variables for text, and provide a detailed analysis of their effects on high-level human judgements of conversational aspects. We show that by controlling combinations of these variables, our models demonstrate clear improvements in human quality judgements.
The second paper we present, "The Wizard of Wikipedia", considers the expectation of users that chatbots should exhibit strong, open-domain factual world knowledge. We present a new dataset of conversations which contains explicit groundings in facts from Wikipedia. We propose novel models for integrating knowledge into conversations, and compare to standard "generate and hope" neural models. We show our models exhibit the ability to recall and incorporate factual knowledge, even when discussing previously unseen topics.
Bio: Stephen Roller is a Research Engineer at Facebook AI Research. His research focuses primarily on generation in neural dialogue agents. Before joining FAIR, Stephen completed his PhD at the University of Texas at Austin under the supervision of Katrin Erk, where his research focused primarily on hypernymy and lexical entailment. Stephen additionally completed part of studies as a visiting PhD student at the University of Stuttgart, where he researched multimodal lexical semantics under the supervision of Sabine Schulte im Walde.