We built a system that detects linguistic anomalies in political speeches by comparing them against LLMs fine-tuned on a speaker's past rhetoric. Applied to Trump's second inaugural address for some interesting results.
Links below include reproducible methodology, code, and annotated examples.
Blogpost: https://alignedforesight.substack.com/p/is-trump-normal Colab: https://colab.research.google.com/drive/1uiTgQApEHQrJZB6tpY0... Data: https://drive.google.com/drive/u/0/folders/1_ifNoJxvgUe9oiY8... Data processing code: https://github.com/Aligned-Foresight/speech-finetuning/