Resolving Knowledge Conflicts in Large Language Models

Yike Wang*, Shangbin Feng*, Heng Wang, Weijia Shi, Vidhisha Balachandran, Tianxing He, Yulia Tsvetkov

Dataset

The top-level keys in the json file correspond to primary fields, and each data point within a field is represented as a dictionary, with the following key-value pairs:

• main_entity(str): an entity from the generated entity list
• parametric_knowledge(str): extracted parametric knowledge about the main_entity
• named_entity_lst(lst): named entities with corresponding types returned by NER models
• conflict_generation_method(str): either "substitution" or "shuffling", representing in-domain named entity substitution and in-domain entity shuffling respectively
• entity_before(str): an entity originally presents in the parametric_knowledge before substitution or shuffling
• entity_after(str): the entity that replaces the entity_before in cases of substitution or shuffling
• conflicting_knowledge(str): the conflicting knowledge created by substitution or shuffling
• question_about_conflicting_segments(str): a question related to the conflicting segments of conflicting_knowledge
• question_about_nonconflicting_segments(str): a question related to the nonconflicting segments of conflicting_knowledge

Setup

Install dependencies:

pip install -r requirements.txt

Set your OpenAI API key:

export OPENAI_API_KEY="your_openai_api_key"

Experiments

The exact prompts used for all experiments are included in the prompts folder, with the corresponding samples provided in Appendix E of the paper. You can run the experiments using the following command:

# example
python run.py \
  --input_file dataset/gpt-3.5-turbo/data.json \
  --prompt_file prompts/task2/zero-shot.prompt \
  --output_file results/task2/zero-shot.json

Questions

If you have any questions or comments about our paper or data, feel free to reach out via email at [email protected]. We will do our best to respond within one business day.

Citing

If you found this work helpful, please consider starring this repository and citing our paper as shown below:

@article{wang2023resolving,
  title={Resolving knowledge conflicts in large language models},
  author={Wang, Yike and Feng, Shangbin and Wang, Heng and Shi, Weijia and Balachandran, Vidhisha and He, Tianxing and Tsvetkov, Yulia},
  journal={arXiv preprint arXiv:2310.00935},
  year={2023}
}