Abstract

Abstract:

Social values are a key factor in human decision-making. Some people, for example, oppose the death penalty while others support it, and there is no single objective truth. Large language models are pre-trained on texts authored by many different people with different social values. But when prompted to answer an ethical question or complete a subjective task, model responses will necessarily align with some social values, and not others. This leads to two questions that I want to answer in my research: 1) What social values are reflected in model behaviour? 2) How can we control these values, and by extension model behaviour? In my talk, I will introduce role-playing as a framework for exploring these questions, differentiating between generic roles that models play by default, and specific roles that we ask them to play, for example based on sociodemographic attributes. I will discuss requirements for successful role-playing, including role stability, internal and external alignment, as well as the limitations of role-playing. Lastly, I will present initial role-playing experiments for hate speech detection, as a highly subjective task.

Bio:

Paul Röttger is a final-year DPhil student at the University of Oxford, working on natural language processing. In his thesis, he focused on evaluating and improving hate speech detection models, adapting language models to language change, and managing subjectivity in data annotation. His main research interest now is in exploring and controlling the behaviour of large language models in relation to social values, as part of a larger goal to make models more helpful and less harmful.