Paper: https://arxiv.org/abs/2305.14928
Kellin Pelrine, Anne Imouza, Camille Thibault, Meilina Reksoprodjo, Caleb Gupta, Joël Christoph, Jean-François Godbout, Reihaneh Rabbany
Greatly updated version of our misinformation mitigation paper, forthcoming at EMNLP 2023, is out!
We propose 3 key elements for building a reliable misinfo mitigation system: recent LLMs, graceful failure, and a focus on generalization.
In addition to results on each of those, we conducted experiments on temperature, prompting, comparing LLMs, versioning, explainability, web retrieval, and more. We also published the LIAR-New dataset, which has every example in both English and French, plus novel Possibility labels which mark whether an input has sufficient context or is too ambiguous for veracity evaluation.
Please check it out! And I'd be happy to hear your thoughts.