Learning Argument Acceptability from Abstract Argumentation Frameworks

This paper introduces argument-based decision-tree for learning acceptability of arguments. We specifically examine an attack relation existing between arguments, without referring to any contents, either sentences or words, existing in individual arguments. This idea is formalized using decision trees in which their attributes are instantiated by complete, preferred, stable and grounded extensions, respectively, defined by acceptability semantics. This study extracted 38 arguments and 4 utterers from an argument about euthanasia that actually took place on a social media site. Also, 21 training data were collected by asking them to express their attitudes either for or against the individual 38 arguments. By stratifying audiences in accordance with consistency with utterers, leave-two-out cross validation yielded results with a 0.73 AUC value, on average. This fact demonstrates that our argument-based decision-tree learning is expected to be fairly useful for agents who have a definite position on an issue of argument.