Hanjia Lyu¹, Jiebo Luo¹, Jian Kang¹, Allison Koenecke²

¹ University of Rochester

² Cornell University

Accepted for publication in FAccT 2025

Will also be presented at IC2S2 2025 as a Parallel Talk

• Introduction
• Example Usage
• Prompts of Our Benchmark Dataset
• Datasets We Used to Create Our Prompts
• Requirements
• Reproducibility
  • Figure 2
  • Figure 3
  • Figure 4
  • Figure 5
  • Figure 6
  • Figure 7
  • Figure 8
  • Figure 9
  • Figure 10
  • Figure 11
  • Tables 2 & 21
  • Table 11
  • Table 13
  • Tables 14-17
  • Table 18
  • Table 19
  • Table 20
  • Table 22
  • Table 23
  • Table 24
  • Table 25
  • Table 26
  • Table 27
  • Table 28
  • Table 29
  • Table 30
  • Table 31
  • Table 32
  • Table 33
  • Table 34
  • Table 35
  • Table 36
  • Table 37
• Citation

While the capabilities of Large Language Models (LLMs) have been studied in both Simplified and Traditional Chinese, it is yet unclear whether LLMs exhibit differential performance when prompted in these two variants of written Chinese. This understanding is critical, as disparities in the quality of LLM responses can perpetuate representational harms by ignoring the different cultural contexts underlying Simplified versus Traditional Chinese, and can exacerbate downstream harms in LLM-facilitated decision-making in domains such as education or hiring.

To investigate potential LLM performance disparities, we design two benchmark tasks that reflect real-world scenarios: regional term choice (prompting the LLM to name a described item which is referred to differently in Mainland China and Taiwan), and regional name choice (prompting the LLM to choose who to hire from a list of names in both Simplified and Traditional Chinese).

For both tasks, we audit the performance of 11 leading commercial LLM services and open-sourced models---spanning those primarily trained on English, Simplified Chinese, or Traditional Chinese. Our analyses indicate that biases in LLM responses are dependent on both the task and prompting language: while most LLMs disproportionately favored Simplified Chinese responses in the regional term choice task, they surprisingly favored Traditional Chinese names in the regional name choice task. We find that these disparities may arise from differences in training data representation, written character preferences, and tokenization of Simplified and Traditional Chinese.

• Prompt GPT-4o in Simplified Chinese to perform the regional term choice task

python infer.py --llm gpt4o --task term --lang simplified --prompt_id 1

• Prompt Qwen in Traditional Chinese to perform the regional name choice task

python infer.py --llm qwen --task name --lang traditional --prompt_id 1

• Prompt Llama3-70b in English to perform the regional name choice task

python infer.py --llm llama3-70b --task name --lang english --prompt_id 1

• Use GPT-4o-mini to annotate the result of Qwen on the regional term task when prompted in English and the prompt_id is 2

python gpt_eval.py --llm qwen --lang english --task term --prompt_id 2

• Use GPT-4o-mini to annotate the result of Breeze on the regional name task when prompted in Traditional Chinese and the prompt_id is 1

python gpt_eval.py --llm breeze --lang traditional --task name --prompt_id 1

• prompt/regional_term/{language}_{prompt_id}.csv

These datasets contain the prompts of the regional term choice task. prompt_id represents the prompt version.

• prompt/regional_name/{language}_{prompt_id}.csv

These datasets contain the prompts of the regional name choice task. prompt_id represents the research questions in Section 4. prompt_id_0: Section 4.1, prompt_id_1: Section 4.3.1, prompt_id_2: Section 4.6, prompt_id_3: Section 4.4, prompt_id_4: Section 4.5.

• source_data/regional_term_and_definition.csv

This dataset includes all 110 regional terms, along with their definitions and their usage in the contexts of Mainland China and Taiwan.

• source_data/regional_name_and_characteristics.csv

This dataset includes all 352 regional names, along with their population-based popularity decile assignments and their gender labels---predicted for Mainland Chinese names and reported for Taiwanese names.

• Clone this repository

git clone https://github.com/brucelyu17/SC-TC-Bench.git

• Create a conda virtual environment and activate it

conda create -n sc-tc-bench python=3.10
source activate sc-tc-bench

• Install packages

pip install -r requirements.txt

• To run inference with ChatGLM2, install transformers==4.40.0

python -m reproducibility.fig_2 --prompt_id 1

python -m reproducibility.fig_3 --prompt_id 0

python -m reproducibility.fig_3 --prompt_id 2 --arrow

python -m reproducibility.fig_2 --prompt_id 1 --no_gpt

python -m reproducibility.fig_6

python -m reproducibility.fig_2 --prompt_id 2

python -m reproducibility.fig_2 --prompt_id 3

python -m reproducibility.fig_9

python -m reproducibility.fig_3 --prompt_id 1

python -m reproducibility.fig_3 --prompt_id 3 --arrow

python -m reproducibility.tab_2

python -m reproducibility.tab_11 --once

python -m reproducibility.tab_11

python -m reproducibility.fig_2 --prompt_id 1

python -m reproducibility.tab_18

python -m reproducibility.tab_19

python -m reproducibility.tab_20

python -m reproducibility.tab_22

python -m reproducibility.tab_23

python -m reproducibility.tab_24 --lang simplified --overall

python -m reproducibility.tab_24 --lang traditional --overall

python -m reproducibility.tab_24 --lang english --overall

python -m reproducibility.tab_24 --lang simplified --gender

python -m reproducibility.tab_24 --lang traditional --gender

python -m reproducibility.tab_24 --lang english --gender

python -m reproducibility.tab_30

python -m reproducibility.tab_30 --llm baichuan2

python -m reproducibility.tab_30 --llm qwen

python -m reproducibility.tab_31 --llm baichuan2 --name_example
python -m reproducibility.tab_31 --llm qwen --name_example

python -m reproducibility.tab_34

python -m reproducibility.tab_35

python -m reproducibility.tab_36

python -m reproducibility.tab_37

@inproceedings{sctcbench-facct25,
    title={Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinese},
    author={Lyu, Hanjia and Luo, Jiebo and Kang, Jian and Koenecke, Allison},
    year={2025},
    isbn = {9798400714825},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3715275.3732182},
    doi = {10.1145/3715275.3732182},
    booktitle = {Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency},
    location = {Athens, Greece},
    series = {FAccT '25}
}