ekdinan

Indeed, consistency (with the lengthy dialogue histories, but also the game state as well as game plans) was a huge challenge for us. We spent a lot of time working on techniques for detecting and filtering these kinds of low quality messages. You can see the Supplementary Materials in the paper for full details, but TL;DR: we built a suite of classifiers for detecting common mistakes (training classifiers to discriminate between human messages and counterfactuals), and used these classifiers as an ensemble which acted as a filter on top of message generation.

One of the tasks that we fine-tuned the model on is ConvAI2 (or "Persona-chat") which specifically aims to improve the model's consistency by conditioning its responses on a given persona. See here: https://arxiv.org/abs/1801.07243. In this research we found that conditioning on a persona improves consistency, but that the models still aren't always perfectly consistent. In particular, the model can only see a few turns back in the dialogue history, and may contradict itself if the conversation is long enough that the model simply does not see the full conversation.

As far as facts go, we also fine-tuned the model on the Wizard of Wikipedia task (https://arxiv.org/abs/1811.01241), which helps improve its knowledgeability. Again, it still isn't perfect.

Hi, paper author here. The model was fine-tuned on (non-Reddit) data that was specifically designed to give it positive conversational skills, like empathy, knowledge, and personality (see: https://arxiv.org/abs/2004.13637).

No FB data was used to train these models, which is what allowed us to open source it.