Reliability of Psychological Scales on LLMs:

Execution Process

Create Utils File

An example utils.py:

api_key = "<API key>"
temperature = <model temperature>
delay_time = <the seconds between each request>
model = "<the name of the model>"

Specify Test Cases

In main.py, specify the server parameters:

1. template: a list of prompt templates.
2. version: a list of question versions.
3. language: a list of language versions.
4. label: a list of level option labels.
5. order: a list of level orders.
6. questionnaire_name: the selected questionnaire.
7. name_exp: name of the save file.

Start a Server class, all pre-testing cases are created and stored in save/<name_exp>.json

test = Server(questionnaire_name, template, version, language, label, order, name_exp)

Load the saved file as a new save, a protection mechanism for test interruption

test = load("<save_path>", "<new_save_name>")

Run for all pre-testing cases

test.run()

An Example Run

from server import *

template = ['t1','t2','t3','t4','t5']
version = ['v1','v2','v3','v4','v5']
language = ['En', 'Zh', 'Ko', 'Es', 'Fr', 'De', 'It', 'Ar', 'Ru', 'Ja']
label = ['n', 'al', 'au', 'rl', 'ru']
order = ['r', 'f']
questionnaire_name = 'BFI'
name_exp = 'bfi-save'

bfi_test = Server(questionnaire_name, template, version, language, label, order, name_exp)
bfi_test.run()

Rephrase the Statements

In main.py, execute:

rephrase("<questionnaire_name>", "<specified_language>")

References

For more details, please refer to this paper. Please remember to cite us if you find our work helpful in your work!

@inproceedings{huang2024reliability,
  author    = {Jen{-}tse Huang and
               Wenxiang Jiao and
               Man Ho Lam and
               Eric John Li and
               Wenxuan Wang and
               Michael R. Lyu},
  title     = {On the Reliability of Psychological Scales on Large Language Models},
  booktitle = {The 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP Main)},
  year      = {2024}
}