Analysing Personal Attacks in U.S. Presidential Debates
Ruban Goyal, Rohitash Chandra, Sonit Singh
TL;DR
The paper tackles the detection of personal attacks in US presidential debates by creating a manually annotated transcript dataset spanning 2016, 2020, and 2024. It evaluates both task-specific fine-tuned BERT models and general-purpose large language models (LLMs) such as ChatGPT, Claude, Gemini, Grok, and DeepSeek, plus domain-adapted LLaMA using LoRA. Key findings show that fine-tuned BERT achieves high accuracy and robustness, while LLMs offer complementary strengths in precision and recall depending on the model. The study demonstrates that domain adaptation and careful evaluation across debate contexts can significantly improve automated analysis of political discourse, with practical implications for journalism, transparency, and public understanding of debates.
Abstract
Personal attacks have become a notable feature of U.S. presidential debates and play an important role in shaping public perception during elections. Detecting such attacks can improve transparency in political discourse and provide insights for journalists, analysts and the public. Advances in deep learning and transformer-based models, particularly BERT and large language models (LLMs) have created new opportunities for automated detection of harmful language. Motivated by these developments, we present a framework for analysing personal attacks in U.S. presidential debates. Our work involves manual annotation of debate transcripts across the 2016, 2020 and 2024 election cycles, followed by statistical and language-model based analysis. We investigate the potential of fine-tuned transformer models alongside general-purpose LLMs to detect personal attacks in formal political speech. This study demonstrates how task-specific adaptation of modern language models can contribute to a deeper understanding of political communication.