Argumentation Mining from Acceptability of Arguments

15^th January 2019, 13:00 Ashton Lecture Theater
Dr. Hiroyuki Kido
Institute of Logic and Cognition
Sun Yat-Sen University

Abstract

Argumentation mining aims to identify an argumentative structure from unstructured data consisting mainly of natural language texts. The most challenging task of argumentation mining is considered to be a prediction of an attack relation between arguments. Its dominant approach is currently natural language processing (NLP) with machine learning. However, it is too optimistic to expect that NLP solves it perfectly. In my talk, the task is tackled by a combination of Bayesian statistics and computational argumentation, rather than NLP. I present a generative model to solve an inverse problem of argument-based reasoning. Given sentiments regarding acceptability of arguments, the inverse problem finds an attack relation justifying the sentiments in terms of acceptability semantics. It thus statistically constructs a justification explaining the sentiments. The results of an experiment support our hypothesis that sentiments collected in an online forum can be reasonably predicted through an attack relation estimated with our generative model.