M2P2: Multimodal Persuasion Prediction Using Adaptive Fusion

Identifying persuasive speakers in an adversarial environment is a critical task. In national election, politicians would like to have campaign on their behalf. When company faces adverse publicity, they engage advocates for position the presence of adversaries who are them. Debates represent common platform these forms persuasion. This paper solves two problems: Debate Outcome Prediction (DOP) problem predicts wins debate while Intensity Persuasion (IPP) change number votes before and after speaker speaks. Though DOP has been previously studied, we first study IPP. Past studies fail leverage important aspects multimodal data: 1) multiple modalities often semantically aligned, 2) different may provide diverse information prediction. Our $\mathsf{M2P2}$ (Multimodal Prediction) framework use (acoustic, visual, language) data solve IPP problem. To alignment maintaining diversity cues provide, devises novel adaptive fusion learning which fuses embeddings obtained from modules – alignment module that extracts shared between xmlns:xlink="http://www.w3.org/1999/xlink">heterogeneity learns weights with guidance three separately trained unimodal reference models. We test popular IQ2US dataset designed DOP. also introduce new called QPS (from Qipashuo, Chinese TV show) significantly outperforms 4 recent baselines both datasets.