Ming-Hung Wang
Author directory2026
DR-CUP: A Dataset on Real-time Commentary in U.S. Presidential Debates
Yu-Yu Chang | Huan-Wen Ho | Chung-Chi Chen | Ming-Hung Wang
Proceedings of the Fifteenth Language Resources and Evaluation Conference
Yu-Yu Chang | Huan-Wen Ho | Chung-Chi Chen | Ming-Hung Wang
Proceedings of the Fifteenth Language Resources and Evaluation Conference
Presidential debates are critical platforms for political discourse, yet existing research lacks datasets tailored for analyzing real-time professional commentary. To address this, we introduce the Dataset on Real-time Commentary in U.S. Presidential debates (DR-CUP), which aligns U.S. presidential debate transcripts (2016–2024) with professional commentary and annotations. DR-CUP supports research on commentary understanding, planning, and generation, offering insights into expert analysis and its role in contextualizing complex political discourse. In pilot studies, we evaluated state-of-the-art large language models (LLMs), revealing notable performance differences in understanding expert commentary and planning for generating professional commentary. DR-CUP is the first dataset to incorporate real-time cross-document alignment for debate data, providing a comprehensive resource for advancing research in political communication and computational social science.
Linguistic Cues in Political Debates: Lexical Choices and Evaluative Framing in GPT-Generated Commentary
Yu-Yu Chang | Chung-Chi Chen | Ming-Hung Wang
Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
Yu-Yu Chang | Chung-Chi Chen | Ming-Hung Wang
Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
Political debates shape public opinion, but many readers increasingly encounter debate takeaways through AI-generated summaries and commentary rather than full broadcasts. This paper studies whether fine-grained lexical choices in debate transcripts can influence the evaluative framing of GPT-generated political commentary. We introduce controlled word-level interventions by replacing embedding-identified key lexical items with near-synonymous alternatives, allowing us to examine how minimally invasive lexical changes affect model-generated interpretation. Across U.S. presidential debate segments, we find that single-word substitutions can shift the stance of GPT-generated commentary, but these effects are probabilistic, context-dependent, and sensitive to the speaker’s communicative goal. Substitutions are more likely to produce favorable stance shifts in critical contexts than in promotional ones, suggesting that GPT-generated commentary may amplify adversarial framing more readily than self-promotional framing. We further analyze which linguistic cue types are associated with substitution success, including sentiment direction, part of speech, and contextual suitability, and show that sentiment alone is insufficient to explain model responses. Finally, a small exploratory human evaluation suggests partial alignment between model-based stance shifts and reader preferences in promotional cases, while also revealing substantial subjectivity in human judgments. These findings highlight how candidates’ word choices may be transformed when political discourse is filtered through GPT-generated commentary, and caution against treating model-internal stance metrics as direct substitutes for large-scale human audience evaluation.
2025
Search