Random Silicon Sampling: Simulating Human Sub-Population Opinion Using a Large Language Model Based on Group-Level Demographic Information.
This repository contains the code for the paper "Random Silicon Sampling: Simulating Human Sub-Population Opinion Using a Large Language Model Based on Group-Level Demographic Information." All code has been anonymized for confidentiality.
Within the data folder, you'll find the ANES data from the years 2012, 2016, and 2020. Additionally, the questionnaire file ANES_2020_multiple_questions_selected.xlsx, used for our multiple question experiment, is also included. Each dataset can be downloaded from American National Election Studies (ANES).
We have reproduced and extended the silicon sampling method described in Argyle et al., (2023). The code has been modified and augmented based on the code used by Argyle et al., (2023).
common.py is a script for using the OpenAI API. Insert your own OpenAI API key and select the desired model as the engine argument. In our experiment, gpt-3.5-turbo-0613 was adopted.
anes2012.py, anes2016.py, and anes2020.py are scripts for converting demographic information of respondents from each respective ANES dataset into first-person prompts.
main.py is a script for performing random silicon sampling on the U.S. presidential election candidate choice for each year. You can run
python main.py <year>
to conduct random silicon sampling on the ANES data for the specified year.
main_st.py is a script for stratified experiments. It allows for random silicon sampling from 23 subgroups extracted from the ANES 2020 data.
main_mq.py is a script for the multiple question experiment. It facilitates random silicon sampling using ANES 2020 data for 10 surveys selected in our study.
main_mq_st.py is a script for stratified experiments on multiple questions. It performs random silicon sampling on 23 subgroups extracted from ANES 2020 data for 10 surveys chosen in our study.