View Proposal

---

Proposer

Gavin Abercrombie

Title

Safety and Bias in Dialogue Systems and Large Language Models

Goal

Evaluate and develop datasets, methods, and systems for the detection and mitigation of safety issues and biases in Dialogue Systems and LLMs.

Description

Automated dialogue systems are becoming ubiquitous in our homes, on our smart devices, and on the internet, and, with recent advances in Large Language Models, the quality of chatbots and voice assistants is rapidly improving — to the extent that they can sometimes be mistaken for humans [1]. But as the quality of end-to-end dialogue systems improves, so does their capacity to learn unsafe behaviours from data on which they are trained [2]. They also run the risk of responding inappropriately to unsafe or toxic user input [3]. Potentially undesirable behaviours include offensive outputs (abuse, hate speech etc.), as well as the failure to detect and mitigate such language in user inputs, and generation of inappropriate responses in safety-critical situations, such as offering medical or legal advice, or responding to/generating sensitive content. In this project, we will examine one or more of the following aspects of safety for conversational AI: - Evaluation and/or detection of unsafe user input and/or system outputs (abusive language, safety critical topics, sensitive content etc.). - Evaluation of societal biases in the outputs of conversational systems and Large Language Models (LLMs) - Mitigation of unsafe content e.g. evaluation of system response strategies, generation of appropriate responses.

Resources

[1] Abercrombie, G., Cercas Curry, A., Pandya, M., & Rieser, V. (2021). Alexa, Google, Siri: What are Your Pronouns? Gender and Anthropomorphism in the Design and Perception of Conversational Assistants. Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing. [2] Dinan, E., Gavin Abercrombie, A. Bergman, Shannon Spruit, Dirk Hovy, Y-Lan Boureau, and Verena Rieser. 2022. Safety Kit: First Aid for Measuring Safety in Open-Domain Conversational Systems. ACL.

Background

Url

Difficulty Level

Moderate

Ethical Approval

Full

Number Of Students

0

Supervisor

Gavin Abercrombie

Keywords

safety, bias, nlp, dialogue systems, large language models

Degrees

Master of Science in Artificial Intelligence
Master of Science in Artificial Intelligence with SMI
Master of Science in Data Science