Construction and Analysis of a Persuasive Dialogue Corpus

Abstract

Persuasive dialogue systems, systems which are not passive actors, but actually try to change the thoughts or actions of dialogue participants, have gained some interest in recent dialogue literature. In order to construct more effective persuasive dialogue systems, it is important to understand how the system’s human counterparts perform persuasion. In this paper, we describe the construction of a corpus of persuasive dialogues between real humans, and an analysis of the factors that contribute to the persuasiveness of the speaker. Specifically, we collect dialogue between 3 professional salespeople and 19 subjects, where the salesperson is trying to convince a customer to buy a particular product. We annotate dialogue acts of the collected corpus, and based on this annotated corpus, perform an analysis of factors that influence persuasion. The results of the analysis indicate that most common dialog acts are information exchange, and about 30 % of the persuader’s utterances are argumentation with framing aiming at making listener select a particular alternative. Finally, we perform a regression analysis of factors contributing to the satisfaction of the customer and persuasive power of the salesperson. We find that factors derived from dialogue acts are particularly effective predictor of satisfaction, and factors regarding framing are particularly effective predictors of persuasive power.