Recipes for Safety in Open-domain Chatbots

Abstract

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.

References in corpus (20)

• A Deep Reinforced Model for Abstractive Summarization
• CTRL: A Conditional Transformer Language Model for Controllable Generation
• Wizard of Wikipedia: Knowledge-Powered Conversational agents
• How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation
• Towards a Human-like Open-Domain Chatbot
• Conversational AI: The Science Behind the Alexa Prize
• fairseq: A Fast, Extensible Toolkit for Sequence Modeling
• Plug and Play Language Models: A Simple Approach to Controlled Text Generation
• DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation
• ACUTE-EVAL: Improved Dialogue Evaluation with Optimized Questions and Multi-turn Comparisons
• Adversarial NLI: A New Benchmark for Natural Language Understanding
• Advancing the State of the Art in Open Domain Dialog Systems through the Alexa Prize
• SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval)
• Controlling Style in Generated Dialogue
• Neural Generation Meets Real People: Towards Emotionally Engaging Mixed-Initiative Conversations
• Chat as Expected: Learning to Manipulate Black-box Neural Dialogue Models
• Zero-Shot Fine-Grained Style Transfer: Leveraging Distributed Continuous Style Representations to Transfer To Unseen Styles
• Detecting the Hate Code on Social Media
• Detecting and Classifying Malevolent Dialogue Responses: Taxonomy, Data and Methodology
• Demoting Racial Bias in Hate Speech Detection

Cited by in corpus (18)

• A Literature Survey of Recent Advances in Chatbots
• Identifying and Mitigating the Security Risks of Generative AI
• Prompted LLMs as Chatbot Modules for Long Open-domain Conversation
• Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling
• Automatic Evaluation and Moderation of Open-domain Dialogue Systems
• Hi, my name is Martha: Using names to measure and mitigate bias in generative dialogue models
• Detoxifying Language Models Risks Marginalizing Minority Voices
• AutoConv: Automatically Generating Information-seeking Conversations with Large Language Models
• A Unified Pre-training Framework for Conversational AI
• GeDi: Generative Discriminator Guided Sequence Generation
• Proto: A Neural Cocktail for Generating Appealing Conversations
• Assessing Political Prudence of Open-domain Chatbots
• Did Chatbots Miss Their 'Apollo Moment'? A Survey of the Potential, Gaps and Lessons from Using Collaboration Assistants During COVID-19
• Just Say No: Analyzing the Stance of Neural Dialogue Generation in Offensive Contexts
• Viola: A Topic Agnostic Generate-and-Rank Dialogue System
• Distilling the Knowledge of Large-scale Generative Models into Retrieval Models for Efficient Open-domain Conversation
• Automatically Exposing Problems with Neural Dialog Models
• Investigating Robustness of Dialog Models to Popular Figurative Language Constructs