PAK: A Survey-Grounded Synthetic Persona Dataset of the Korean Cultural Arts Workforce
This paper introduces PAK, the first synthetic persona dataset of the Korean cultural arts workforce comprising 30,000 statistically grounded personas reconstructed from the 2024 Survey of Korean Artists using iterative proportional fitting, alongside a comprehensive evaluation of its distributional accuracy and narrative quality.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, computers are learning to speak, write, and even pretend to be people. To teach them how to act like humans, researchers often use "personas"—digital profiles that describe a person's age, job, hobbies, and history. These profiles act as a script, telling the computer how to respond in a conversation. For years, these scripts have been made up or pulled from random internet posts, which means they often miss the mark when it comes to real-world accuracy. They might invent a job that doesn't exist or give a person a life story that contradicts their age. To build better, more trustworthy AI, scientists need a way to create these digital people based on hard facts, ensuring that every detail, from income to career length, matches what we know about real populations.
A researcher at the Korea Advanced Institute of Science and Technology has tackled this problem by creating a new resource called PAK, a collection of thirty thousand synthetic personas representing the Korean cultural arts workforce. Instead of guessing or scraping the web, this project builds every single digital person from the ground up using official government statistics. The team started with the 2024 Survey of Korean Artists, a massive national report that tracks the lives of over 330,000 registered artists. This survey provides a detailed map of the population, showing how many artists work in literature versus film, how old they are, where they live, and how much they earn. The challenge was that the survey only released summary tables, not the full list of individual answers. The researchers had to use a mathematical method to reconstruct the missing details, ensuring that the final group of thirty thousand digital people perfectly matched the official numbers for every category, from gender to career breaks.
Once the statistical skeleton was built, the researchers gave each digital person a voice. Using a large language model, they wrote seventeen different narrative sections for every persona, covering everything from their creative philosophy and family life to their daily hobbies and travel habits. To make sure these stories made sense, the team added a special check for people who started their careers later in life. In early tests, the computer often made a mistake: it would write a story about a sixty-year-old artist who had been painting since childhood, even though the data said they only started working in their fifties. The researchers fixed this by teaching the computer to look at the person's age and career length first. If the data showed a late start, the story was rewritten to describe a life of other jobs before the artistic career began, removing the contradiction and making the character feel more real.
The result is a dataset where every number can be traced back to a published government table, and every story is consistent with those numbers. The researchers tested the work in several ways. They checked the statistics and found that the digital group matched the real survey results almost perfectly, with differences of less than one percent in key areas like full-time employment and contract experience. They also asked human experts and other computer models to read the stories and judge their quality. The experts found that while the stories were generally coherent and grounded in the data, they were not always perfect; about half of the fluency scores were average, suggesting the writing followed a template rather than sounding like a polished human essay. However, the computer models were able to read the stories and correctly guess the hidden facts about the people, proving that the narratives contained real information rather than just random words.
This project does not claim to have captured the true, lived experiences of real artists, nor does it contain any private data from actual survey respondents. It is a controlled simulation, a tool designed to help researchers test how well AI systems understand specific cultural groups. By providing a set of digital people that are statistically accurate and internally consistent, PAK offers a new standard for evaluating language models in the Korean language. It shows that it is possible to move beyond making up facts and instead build synthetic worlds that respect the boundaries of real data, creating a resource that is both reusable and reproducible for future studies in artificial intelligence.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.