Introduction

This is the repositoy for "Perils and Opportunities in Using Large Language Models in Psychological Research". All code and data necessary to replicate the analysis in the paper are presented here.

General Setup / First Step

1. In a terminal execute the command conda env create --file=textgen.yaml. This creates a conda environment and installs the necessary packages to execute all codebooks.
2. Activate the environment using conda activate textgen.
3. Install the required packages using pip install -r requirements.txt
4. If the cuda-toolkit could not be directly installed from the requirements.txt, run pip install nvidia-cudnn-cu11==8.6.0.163 to force an update.
5. You might have to link tensorRT using: export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/anaconda3/envs/textgen/lib/python3.10/site-packages/tensorrt-libs/ or where your python libs are saved. (tensorRT is optional but might increase inference speed)
6. Add your OpenAI API key to your environment before running the ChatGPT files: setx OPENAI_API_KEY “<yourkey>” (windows), echo "export OPENAI_API_KEY='yourkey'" >> ~/.zshrc and source ~/.zshrc (linux/mac)

Text Annotations

Preparation

Sample the test data from the MFRC:

• Open jupyter lab or any other editor that can handle pythonbooks: E.g., in a terminal, navigate to the project folder, activate the conda environment, and execute jupyter lab.
• Open annotations/train_bert_model/prepare_data.ipynb. This codebook processes the MFRC data set and creates a training sample to fine-tune BERT and a separate test sample to compare BERT and ChatGPT.
• Run all cells in the pythonbooks. This should create the following files:
  • data/preprocessed/mfrc_eval_full.csv
  • data/preprocessed/mfrc_meta_sample_full.csv
  • data/preprocessed/mfrc_sample_full.csv
  • data/preprocessed/mfrc_cleaned_full.csv
  • data/train_test/mfrc_train_full.pkl

BERT

1. To fine-tune BERT, open annoations/train_bert_model/train_BERT.ipynb and run all cells

   • You will need a GPU for this step! We used a v100 on a computing cluster but it should also run on less powerful GPUs (e.g., RTX2060)!
   • Make sure to have all GPU related packages (e.g., CUDA, CUDA Toolkit, etc) correctly installed.
   • This should be done automatically when creating the conda environment from the .yaml file.
   • Check this website if you encounter issues with using your GPU and tensorflow: https://www.tensorflow.org/install/pip

2. After running the file, the model will be saved in annotations/models/mfrc_normal_full.h5

3. Create the predictions on the test sample by running all cells in annoations/train_bert_model/predict_BERT.ipynb:

   • The predictions will be saved under results/predictions/mfrc_labels_normal_full.csv

4. We also created regular .py files for training and prediction via command line. If you are using a computing cluster with slurm, we also created exemplary .job files. Adjust as needed.

   • If you use the command line, the arguments for training are "mfrc", "full", "normal" (corpus, aggregate level for moral values, training type)
   • If you want to optimize the model paramters (e.g., add classification layers, change the bert model, etc), you can train using "eval" instead of "normal", which will return a cross-validated performance on the training data. In that case you also need to specify the threshold for classifying a text as a containing a moral sentiment (between 0-1).

LIWC

1. We provide the embedded texts for training in ../data/preprocessed/mfrc_train_full_liwc.csv and for testing in ../data/preprocessed/mfrc_sample_full_liwc.csv`
   • To create LIWC embeddings from text files, you will need a LIWC license and the respective toolkit from: https://www.liwc.app
2. Open and run all cells in annotations/train_bert_model/train_LIWC.ipynb to train the LIWC based model
3. Open and run all cells in annotations/train_bert_model/predict_LIWC.ipynb to annotate the test data
4. Open and run all cells in annotaitons/train_bert_model/LIWC_performance.ipynb to calculate the annotation performance (F1 score)

ChatGPT

1. Open annotations/codes/chatGPT_annotations.ipynb

   • THIS WILL CHARGE YOUR ACCOUNT!
   • Make sure that you know the prices before running (check https://openai.com/pricing)

2. Run all cells:

   • This will save the ChatGPT annotations under results/predictions/gpt_mfrc_labels_full.csv

3. Open annotations/codes/chatGPT_annotations_Fewshot.ipynb (repeat the analysis using fewshot learning instead of zeroshot)

   • THIS WILL CHARGE YOUR ACCOUNT!
   • Make sure that you know the prices before running (check https://openai.com/pricing)

4. Run all cells:

   • This will save the ChatGPT annotations under results/predictions/gpt_fewshot_mfrc_labels_full.csv

5. Open annotations/codes/chatGPT_annotations_FT.ipynb (finetune ChatGPT3.5 and repeat the analysis)

   • THIS WILL CHARGE YOUR ACCOUNT!
   • Make sure that you know the prices before running (check https://openai.com/pricing)

6. Run all cells:

   • This will save the ChatGPT annotations under results/predictions/gpt_FT_mfrc_labels_full.csv

Statistical Anlysis

1. Open annotations/codes/chatGPT_performance.ipynb and run all cells

   • This will calculate the correct/false classifications for the ChatGPT annotations, add the annotator demographic information and save it under ../results/evals/gpt_mfrc_success_full.csv.

2. Open annotations/train_bert_model/BERT_performance.ipynb and run all cells

   • This will calculate the correct/false classifications for the BERT model, add the annotator demographic information and save it under ../results/evals/mfrc_success_normal_full.csv.

3. Open annotations/statistical_analyses/annotations_analyses.Rmd and run all cells

   • The output of ## Evaluate will show the logistic regression outputs for each set of annotator variables (e.g., demographics, moral values, etc). Under each regression output are the coefficients converted to percentage differences in odds. These results are presented in Table X of our work and express how each annotator characteristic is linked to the models predictions (i.e., how biased the classifier is towards said annotator characteristic).
   • The output of ## Fit Model (moral foundation ~ predictor) will show the logistic regression of predicting each set of moral sentiment as a function of Classifier (BERT, ChatGPT, compared to humans). The results show how much more or less likely a Classifier predicts a class compared to trained human annotators (i.e., how much it over or underpredicts each moral sentiment) and is shown in Table X of our paper.
   • The output of ## Extract Coefficients converts the coefficients above into percentage differences in odds (i.e., how much more in percent does a classifier predict a moral sentiment compared to trained humans).

Survey Responses

1. Open survey_predictions/code/prepare_data_gpt.ipynb and run all cells. This will create a data/processed/SURVEY_cleaned.csv file for each survey in the data/surveys folder. In our data, some information was not collected for all participants so we filter for those participants who responded to the items of interests. If you apply this pipeline on your own data this step will likely not be necessary or you will have to specify different items of interest in the COLS_META variable.

   • The code will also generate the prompts under data/prompts/SURVEY.pkl for each survey. The prompts is generated from the PROMPT_TEXT variables and the item texts. If you use different surveys, make sure to adjust PROMPT_TEXT to the respective response scales.

2. Open survey_predictions/code/run_prompts_gpt.ipynb

   • Specify, which surveys to run in d_list (list the names of all surveys from data/surveys that you want to collect responses from). The default are the surveys we ran in our study.

3. Run all cells. This will generate the ChatGPT responses and save them under results/SURVEY.csv for each SURVEY

Statistical Analyses

1. Open statistical_analyses/survey_analysis.Rmd and run all cells.

   • This will calculate all group diffferences between humans and ChatGPT's survey responses, output the results as tables and save figures under results/plots/
   • The output of ### Demographic Group Differences shows the differences of ChatGPT's survey responses and various demographic groups using Dunnett's Test. The test compares for each demographic variable the different levels with ChatGPT (e.g., for political orientation it compares Liberals, Moderates, Conservatives against ChatGPT). The results of this analysis are shown in Table X and Figure Y of our paper.
   • The output of ### Results shows the regression of various human demographic variables on similarity to ChatGPT's survey responses. This expresses how much more similar ChatGPT is to a certain demographic group when responding to surveys. The results of this analysis are shown in Table X and of our paper.

2. Repeat this for any survey you are investigating (in our paper: bigfive, closure, cognition, rwa, systems_feelings; change variable d = to these values).

Prompt Sensitivity

Text Annotations

1. Open annotations/codes/chatGPT_annotations_ALT.ipynb

   • THIS WILL CHARGE YOUR ACCOUNT!
   • Make sure that you know the prices before running (check https://openai.com/pricing)

2. Run all cells:

   • This will save the ChatGPT annotations under results/predictions/gpt_mfrc_labels_full_ALT.csv, one for each altered prompt (numbered; ALT1, ALT2, ...).

Statistical Analyses

1. Open annotations/codes/chatGPT_performance_ALT.ipynb and run all cells

   • This will calculate the correct/false classifications and add the annotator demographic information and save it under ../results/evals/gpt_mfrc_success_full_ALT.csv, one for each altered prompt.

2. Open annotations/statistical_analyses/annotations_analyses_prompting.Rmd and run all cells

   • The output of ## Evaluate will show the logistic regression outputs for each set of annotator variables (e.g., demographics, moral values, etc). Under each regression output are the coefficients converted to percentage differences in odds. These results are presented in Table X of our work and express how each annotator characteristic is linked to the models predictions (i.e., how biased the classifier is towards said annotator characteristic).
   • The output of ## Fit Model (moral foundation ~ predictor) will show the logistic regression of predicting each set of moral sentiment as a function of Classifier (BERT, ChatGPT, compared to humans). The results show how much more or less likely a Classifier predicts a class compared to trained human annotators (i.e., how much it over or underpredicts each moral sentiment) and is shown in Table X of our paper.
   • The output of ## Extract Coefficients converts the coefficients above into percentage differences in odds (i.e., how much more in percent does a classifier predict a moral sentiment compared to trained humans).

Survey Responses

1. Open survey_predictions/code/prepare_data_gpt.ipynb and run all cells. This will create a data/processed/SURVEY_cleaned.csv file for each survey in the data/surveys folder. In our data, some information was not collected for all participants so we filter for those participants who responded to the items of interests. If you apply this pipeline on your own data this step will likely not be necessary or you will have to specify different items of interest in the COLS_META variable.

   • The code will also generate the prompts under data/prompts/SURVEY.pkl for each survey. The prompts are generated from the PROMPT_TEXT variables and the item texts. If you use different surveys, make sure to adjust PROMPT_TEXT to the respective response scales.

2. Open survey_predictions/code/run_prompts_gpt.ipynb

   • Specify, which surveys to run in d_list (list the names of all surveys from data/surveys that you want to collect responses from). The default are the surveys we ran in our study.

3. Run all cells. This will generate the ChatGPT responses and save them under results/SURVEY.csv for each SURVEY

Statistical Analyses

1. Open statistical_analyses/survey_analysis_prompting.Rmd and run all cells.
   • This will calculate all group diffferences between humans and ChatGPT's survey responses, output the results as tables and save figures under results/plots/
   • The output of ### Demographic Group Differences shows the differences of ChatGPT's survey responses and various demographic groups using Dunnett's Test. The test compares for each demographic variable the different levels with ChatGPT (e.g., for political orientation it compares Liberals, Moderates, Conservatives against ChatGPT). The results of this analysis are shown in Table X and Figure Y of our paper.
   • The output of ### Results shows the regression of various human demographic variables on similarity to ChatGPT's survey responses. This expresses how much more similar ChatGPT is to a certain demographic group when responding to surveys. The results of this analysis are shown in Table X and of our paper.

Open-Source Pipeline

Preparation

• Follow https://github.com/oobabooga/text-generation-webui to install the interface for LLaMa (either use the "one-click-installer" or manually install).
• Start the interface via terminal (activate the conda environment, enter the textgen directory, run python server.py --api).
• In the interface, click on the "Model" tab. On the right pane, under "Download custom model or LoRA", enter "TheBloke/Luna-AI-Llama2-Uncensored-GPTQ:gptq-4bit-32g-actorder_True" and press download (this will download the model used in our studies.
• After loading is completed, on the left pane under Model choose the model on the drop down menu.
• Under "model loader" choose "ExLLama" and click on load. This will load the model so that our python script can process the prompts

Text Annotations

1. Open annotations/codes/llama_annotations.ipynb

   • THIS WILL CHARGE YOUR ACCOUNT!
   • Make sure that you know the prices before running (check https://openai.com/pricing)

2. Run all cells:

   • This will save the ChatGPT annotations under results/predictions/llama2_mfrc_labels_full.csv

Statistical Anlysis

1. Open annotations/codes/llama_performance.ipynb and run all cells

   • This will calculate the correct/false classifications and add the annotator demographic information and save it under ../results/evals/llama2_mfrc_success_full.csv.

2. Open annotations/statistical_analyses/annotations_analyses_llama.Rmd and run all cells

   • The output of ## Evaluate will show the logistic regression outputs for each set of annotator variables (e.g., demographics, moral values, etc). Under each regression output are the coefficients converted to percentage differences in odds. These results are presented in Table X of our work and express how each annotator characteristic is linked to the models predictions (i.e., how biased the classifier is towards said annotator characteristic).
   • The output of ## Fit Model (moral foundation ~ predictor) will show the logistic regression of predicting each set of moral sentiment as a function of Classifier (BERT, ChatGPT, compared to humans). The results show how much more or less likely a Classifier predicts a class compared to trained human annotators (i.e., how much it over or underpredicts each moral sentiment) and is shown in Table X of our paper.
   • The output of ## Extract Coefficients converts the coefficients above into percentage differences in odds (i.e., how much more in percent does a classifier predict a moral sentiment compared to trained humans).

Survey Responses

1. Open survey_predictions/code/prepare_data_llama.ipynb and run all cells. This will create a data/processed/SURVEY_cleaned_llama2.csv file for each survey in the data/surveys folder. In our data, some information was not collected for all participants so we filter for those participants who responded to the items of interests. If you apply this pipeline on your own data this step will likely not be necessary or you will have to specify different items of interest in the COLS_META variable.

   • The code will also generate the prompts under data/prompts/SURVEY_llama2.pkl for each survey. The prompts are dynamically generated using scale_meaning_dict and the item texts. If you use different surveys, make sure to adjust scale_meaning_dict to the respective response scales.

2. Open survey_predictions/code/run_prompts_llama2.ipynb. Make sure that the textgen interface is running in the background.

   • Specify, which surveys to run in d_list (list the names of all surveys from data/surveys that you want to collect responses from). The default are the surveys we ran in our study.

3. Run all cells. This will generate the LLaMa2 responses and save them under results/SURVEY_llama2.csv for each SURVEY.

Statistical Analyses

1. Open statistical_analyses/survey_analysis_llama2.Rmd and run all cells.

   • This will calculate all group diffferences between humans and ChatGPT's survey responses, output the results as tables and save figures under results/plots/
   • The output of ### Demographic Group Differences shows the differences of ChatGPT's survey responses and various demographic groups using Dunnett's Test. The test compares for each demographic variable the different levels with ChatGPT (e.g., for political orientation it compares Liberals, Moderates, Conservatives against ChatGPT). The results of this analysis are shown in Table X and Figure Y of our paper.
   • The output of ### Results shows the regression of various human demographic variables on similarity to ChatGPT's survey responses. This expresses how much more similar ChatGPT is to a certain demographic group when responding to surveys. The results of this analysis are shown in Table X and of our paper.

2. Repeat this for any survey you are investigating (in our paper: bigfive, closure, cognition, rwa, systems_feelings; change variable d = to these values).

Comparison with top-down methods

1. Open ccr/code/chatGPT_predictions.ipynb
2. Run all cells:
   • This will save the ChatGPT predictions under results/predictions/gpt_topdown.csv

Statistical Analyses

1. Open ccr/statistical_analyses/topdown_analysis.Rmd and run all cells.
   • This will output the statistical analysis and relevant plots and save them under results/plots/.
   • The output of ### Dunnett's Test (gpt & gpt_ccr vs CCR) tests whether ChatGPT's predictions (on the item-level or construct-level) differ significantly from CCR (our topdown method).
   • The output of ### Correlation of model performances shows the correlation between the CCR performance and ChatGPT (on the item and construct level)
   • These analyses can be repeated with different topdown methods or ChatGPT prompting styles - Simply run the alternative topdown method and save the resulting performance in a file analogue to the current behavior_surve.csv or values_survey.csv files. For different GPT approaches, change the prompts in the chatGPT_predictions.ipynb according to your respective considerations.