Architecture for Debating AI

Abstract

This paper discusses architecture for artificial intelligence (AI) that can debate problems with humans. The proposed architecture is composed of (1) offline system which creates text annotations over large amount of text data in advance, and (2) online system which runs various algorithms asynchronously in real time. We built a prototype of debating AI system based on the architecture and evaluated it over 50 discussion topics extracted from a popular debate website. The AI system outputs three opinions per a discussion topic; each opinion is composed of 7 sentences. Thus, it outputs 150 opinions for 50 topics in total. We found that in 61 opinions out of 150 the argumentations are understandable. We plan to develop a feature of combinatorial optimization of algorithms onto the architecture as a future work.