Paper Repository: Evaluating Prompt Engineering Strategies for Sentiment Control in AI-Generated Texts

Overview

This repository contains the code and supplementary materials for the paper "Evaluating Prompt Engineering Strategies for Sentiment Control in AI-Generated Texts" by Kerstin Sahler and Sophie Jentzsch published at HHAI 2025: The 4th International Conference Series on Hybrid Human-Artificial Intelligence.

Contents

The prompts for each optimisation step are organised and saved based on their respective prompting techniques, accompanied by util.py which contains essential utility functions supporting various techniques. For additional reference, other_prompts.ipynb contains a comprehensive collection of supplementary prompts, specifically designed for generating Few-Shot examples and Chain-of-Thought reasoning texts. However, the Chain-of-Thought reasoning texts used in the original experiments for the Automatic as well as the Manual approach are also provided in the cot_examples/ folder together with the examples from the Human approach of Few-Shot prompting.

prompt-sentiment-control/
├── experiments/
│   ├── cot_examples/
│       ├── automatic_factual_reasoning
│       ├── automatic_subjective_reasoning
│       ├── factual_examples
│       ├── manual_factual_reasoning
│       ├── manual_subjective_reasoning
│       └── subjective_examples
│   ├── queries/
│       ├── factual-queries.txt
│       └── subjective-queries.txt
│   ├── cot.ipynb
│   ├── few_shot.ipynb
│   ├── fine_tuning.ipynb
│   ├── other_prompts.ipynb
│   ├── util.py
│   ├── vanilla_baseline.ipynb
│   ├── zero_shot_cot.ipynb
│   └── zero_shot.ipynb
├── Readme.md
└── Supplementary_Material.pdf

Setup

git clone https://github.com/DLR-SC/prompt-sentiment-control.git
cd prompt-sentiment-control
# setup environment
pip install -r requirements.txt

Usage

All experiments were implemented in different Jupyter Notebooks, allowing for modular execution. This implementation allows to run each optimisation step independently, facilitating targeted analysis and individual refinement of specific components without requiring a complete rerun.

Citation

tbd

License

MIT License