Linguistic Cues in Political Debates: Lexical Choices and Ideological Framing in LLM-Generated Commentary

ACL ARR 2026 January Submission1315 Authors

Keywords: Political Debates, Transcript, LLM generation, data analysis

Abstract:

Political debates are critical forums where candidates' language can shape public opinion. Yet, an increasingly common real-world mediation layer is often overlooked: many people now consume debate takeaways through LLM-generated summaries and commentary rather than by watching full debates. This shift changes the persuasion target from only ``the live audience'' to an additional audience of LLM users, whose impressions are filtered through model-generated commentary. In this paper, we study whether candidates’ linguistic cues, fine-grained lexical and rhetorical choices in debate transcripts, can systematically influence how an LLM comments on their performance, and how these altered commentaries in turn shape downstream audience perception.
In our experiment, we mirror contemporary information consumption practices, and introduce controlled word-level interventions by replacing embedding-identified key lexical items with near-synonymous alternatives.
The experimental results show that even minimally invasive lexical substitutions can systematically shift the tone of LLM-generated commentary and, consequently, alter perceived audience sentiment. We further examine these effects across multiple linguistic dimensions to identify which types of linguistic cues are most likely to be amplified or attenuated in LLM-mediated political discourse. These findings highlight how candidates’ language choices may indirectly shape public perception in settings where LLM-generated commentary serves as a primary interpretive layer.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: human behavior analysis, stance detection, human-computer interaction

Contribution Types: NLP engineering experiment, Reproduction study, Data analysis

Languages Studied: English

Submission Number: 1315