Introduction
Public speaking is a high-stakes social interaction where applauding serves as a "risky" coordination game. For an audience member, clapping alone carries a social cost, while collective applause represents a shared moment of triumph. This paper introduces a computational framework to predict audience applause in political campaign speeches by modeling the "verbal design" used by speakers to signal when a clap is licensed.
Modeling Tension and Release
The core innovation of this work is the operationalization of "tension and release" through Rhetorical Structure Theory (RST). The researchers hypothesize that speakers build tension through specific rhetorical structures and provide a satisfying release that triggers applause.
By using an RST parser, the model maps logical relationships—such as ANTITHESIS (contrast) or PURPOSE—between Elementary Discourse Units (EDUs). A "rhetorical closure" occurs when a phrase ties off multiple rhetorical branches, acting as a structural anchor and a release valve for built-up tension, significantly increasing the probability of applause.
Experimental Setup and Dataset
The study utilizes a new C-SPAN Campaign Corpus consisting of 310 speeches from 16 candidates during the 2016 U.S. election cycle. This dataset is notably dense, containing over 80,000 utterances and 19,000 instances of applause (averaging 62 instances per speech).
The modeling approach compares interpretable Logistic Regression with LSTMs designed to capture long-term contextual dependencies. Features include:
- Acoustic Modeling: MFCC features and energy statistics (pitch and RMS) to capture prosody.
- Textual Features: N-grams, repetition patterns, and RST-derived structural features.
Key Findings and Feature Analysis
The models demonstrate that while speaker-specific "stock phrases" (e.g., "Make America...") are strong predictors within a single speaker's data, rhetorical structure and audio energy are crucial for generalizing to new speakers. In inter-speaker tests, the best LSTM model achieved an accuracy of 0.629 and an F1 score of 0.621.
Analysis of the model's coefficients reveals which linguistic and acoustic features most strongly influence audience response.
Positive predictors include expressions of gratitude, future-oriented language ("will," "going to"), and high mean audio energy. Conversely, past-tense verbs and "tentative" language tend to discourage applause. Interestingly, the model correctly predicted a high probability of applause for the line preceding Jeb Bush’s infamous "Please clap" request, suggesting his rhetorical structure was sound even if the delivery or sentiment failed.
Conclusion and Future Directions
While the current framework is robust to the "messy" real-world data of uncorrected captions, several avenues for improvement remain. The exclusion of visual cues—such as gaze and gestures—leaves out a significant portion of "applause invitations." Additionally, distinguishing between organic audience responses and manufactured applause from "claques" (planted supporters) remains a challenge for future computational rhetoric research.
Summary updated