Current open-domain conversational models can easily be made to talk in inadequate ways. Online learning from conversational feedback given by the conversation partner is a promising avenue for a model to improve and adapt, so as to generate fewer of these safety failures. However, current state-of-the-art models tend to react to feedback with defensive or oblivious responses. This makes for an unpleasant experience and may discourage conversation partners from giving feedback in the future. This work proposes SaFeRDialogues, a task and dataset of graceful responses to conversational feedback about safety failures. We collect a dataset of 10k dialogues demonstrating safety failures, feedback signaling them, and a response acknowledging the feedback. We show how fine-tuning on this dataset results in conversations that human raters deem considerably more likely to lead to a civil conversation, without sacrificing engagingness or general conversational ability.

Despite recent improvements in open-domain dialogue models, state of the art models are trained and evaluated on short conversations with little context. In contrast, the long-term conversation setting has hardly been studied. In this work we collect and release a human-human dataset consisting of multiple chat sessions whereby the speaking partners learn about each other's interests and discuss the things they have learnt from past sessions. We show how existing models trained on existing datasets perform poorly in this long-term conversation setting in both automatic and human evaluations, and we study long-context models that can perform much better. In particular, we find retrieval-augmented methods and methods with an ability to summarize and recall previous conversations outperform the standard encoder-decoder architectures currently considered state of the art.

Models trained on large unlabeled corpora of human interactions will learn patterns and mimic behaviors therein, which include offensive or otherwise toxic behavior and unwanted biases. We investigate a variety of methods to mitigate these issues in the context of open-domain generative dialogue models. We introduce a new human-and-model-in-the-loop framework for both training safer models and for evaluating them, as well as a novel method to distill safety considerations inside generative models without the use of an external classifier at deployment time. We conduct experiments comparing these methods and find our new techniques are (i) safer than existing models as measured by automatic and human evaluations while (ii) maintaining usability metrics such as engagingness relative to the state of the art. We then discuss the limitations of this work by analyzing failure cases of our models.