Using In-Context Learning to Improve Dialogue Safety

By Nicholas Meade, Spandana Gella, Devamanyu Hazarika, Prakhar Gupta, Di Jin, Siva Reddy, Yang Liu, Dilek Z. Hakkani-Tür. 2023

60 citations recorded by the source corpus.

While large neural-based conversational models have become increasingly proficient dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, which often perpetuates social biases or stereotypes. We investigate a retrieval-based method for reducing bias and toxicity in responses from chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe responses to similar dialogue contexts. We find our method performs competitively with strong baselines without requiring training. For instance, using automatic evaluation, we find our best fine-tuned baseline only generates safe responses to unsafe dialogue contexts from DiaSafety 4.04% more than our approach. Finally, we also propose a re-ranking procedure which can further improve response safeness.

Subfields: Artificial Intelligence, Safety, Robustness & Interpretability, Dialogue & Conversational AI

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Damai Dai
2. Xu Sun
3. Lei Li
4. Tong Yu
5. Zhiyong Wu
6. Ryan A. Rossi
7. Ruiyi Zhang
8. Qingxiu Dong

In Safety, Robustness & Interpretability

1. Xu Sun
2. Lei Li
3. Franck Dernoncourt

Citations in the corpus

• Bias and Fairness in Large Language Models: A Survey (2023, methodology): “Similar to model reporting practices established by Mitchell et al. (2019), we suggest that the evaluation of bias and fairness issues become standard additions to model documentation.”
• A Survey on In-context Learning (2022, background): “, 2023), (Meade et al., 2023) use ICL for retrieved demonstrations to steer a model towards safer generations, reduc-”
• High-Dimension Human Value Representation in Large Language Models (2024, unknown): “We prioritize LLMs that have undergone human value and preference tuning such as safety tuning [109, 110, 111], reinforcement learning with human feedback (RLHF) [34, 5, 105], direct preference optimization (DPO) [112], etc.”