Argumentation Versus Optimization for Supervised Acceptability Learning

This paper deals with the question of how one should predict agent's psychological opinions regarding acceptability statuses of arguments. We give a formalization of argumentation-based acceptability learning (ABAL) by introducing argument-based reasoning into supervised learning. A baseline classifier is defined based on an optimization method of graph-based semi-supervised learning with dissimilarity network where neighbor nodes represent arguments attacking each other, and therefore, the optimization method adjusts them to have different acceptability statuses. A detailed comparison between ABAL instantiated with a decision tree and naive Bayes, and the optimization method is made using each of 29 examinees' psychological opinions regarding acceptability statuses of 22 arguments extracted from an online discussion forum. We demonstrate that ABAL with the leave-one-out cross-validation method shows better learning performance than the optimization method in most criteria under the restricted conditions that the number of training examples is small and a test set is used to select the best models of both methods.