feat: add code for keyword method, GPT-4, logistic regression experiments

[Maosef]

NLP for the German Federal Constitutional Court: Identifying the Justice Referee

In the judicial system, transparency about which judges decide a case is important for
providing accountability, and understanding political ideologies.
The German Federal Constitutional Court doesn't publicly disclose judge votes;
however, cases are often heavily influenced by a single "judge referee", and the structure of judge assignments to cases by domain provides clues.
We introduce and investigate the task of predicting referee judges from court rulings.
We filter candidate referee judge(s) based on domain keyword matching, and evaluate classification models.

• GPT-4
• logistic regression
• author-topic model

Summary of code

• 2024 Referee: R scripts for data pre and postprocessing for keyword method, GPT4
• python: Python scripts to run GPT-4
• notebooks: logistic regression experiments, calculate accuracy scores for GPT-4

Instructions to run GPT-4 experiments

• Download the data: case and domain data
  • 2024 Referee/bverfg240129.RData
  • 2024 Referee/domain240215.RData
• Generate prompts: 2024 Referee/prepare_GPT4_prompts.R
• Run GPT-4: python/run_GPT.py
• Extract domains from responses: 2024 Referee/process_GPT_responses.R
• Compute accuracy scores for domain and judge predictions: notebooks/judge_prediction_gpt4.ipynb