AI for Low-frequency Data

Nakamura

In the simulation of social surveys using Large Language Models (LLMs), appropriately extracting the diverse values held by the models is a significant challenge. Explicit attribute specification (persona assignment) via conventional prompts tends to converge on neutral responses, revealing limitations in adequately reproducing the complex differences in human values. In this study, as a new approach to complement this limitation, we focused on output instructions using “role language” (tone reminiscent of specific attributes) specific to Japanese, and investigated its impact on the distribution of responses. The experimental results demonstrated that the change in responses due to tone specification was consistently smaller than that of explicit attribute specification, indicating its practical value is currently limited. However, the direction of the response shift was consistent with explicit persona specification, suggesting that role language may access underlying values. We plan to further evaluate the effectiveness of role language through improved experimental design in the future.