Multimodal Modelling of Persuasion Strategies in Social Deduction Games
Author(s)
Pariani, Aryan J.
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Collecting reliable conversational data from realistic social interactions for the study of persuasion is a difficult task, due to the implicit nature of persuasive cues and the privacy concerns in such interactions. This difficulty makes it challenging to design automated systems capable of recognizing the use of persuasion strategies in human social interactions and inferring the belief state of the target audience. Therefore in this thesis, we introduce a large, rich corpus of videos and conversation transcriptions from recorded playthroughs of the social deduction board game One Night Ultimate Werewolf, along with annotations of identified persuasion strategies and final voting decisions of every player. The dataset is presented with a taxonomy of predefined six persuasion strategies commonly used in social deduction games. Two primary prediction tasks are presented to study persuasion in social deduction games. The primary task is aimed at predicting possible persuasion strategies exhibited in an utterance by a player. Baseline models and results for this task are presented to show how prior conversational context and corresponding visual features can benefit persuasion strategy prediction based on the text of the utterance. The second task is that of predicting game outcomes in the form of the final voting decisions of every player, based on players’ dialogue, their game-based roles, and the persuasion strategy distribution for all players. The multimodal nature of social interactions and the indirectly observable belief states of participants make this dataset an important contribution to the study of persuasion strategies.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis