A Computational Approach for Generating Toulmin Model Argumentation

Proceedings of the 2nd Workshop on Argumentation Mining, pages 45–55, Denver, Colorado, June 4, 2015. c⃝2015 Association for Computational Linguistics A Computational Approach for Generating Toulmin Model Argumentation Paul Reisert Naoya Inoue Naoaki Okazaki Kentaro Inui Tohoku University Graduate School of Information Sciences 6-6 Aramaki Aza Aoba, Aobaku, Sendai, Miyagi 980-8579, Japan {preisert,naoya-i,okazai,inui}@ecei.tohoku.ac.jp Abstract Automatic generation of arguments is an im- portant task that can be useful for many ap- plications. For instance, the ability to gener- ate coherent arguments during a debate can be useful when determining strengths of support- ing evidence. However, with limited technolo- gies that automatically generate arguments, the development of computational models for debates, as well as other areas, is becoming increasingly important. For this task, we fo- cused on a promising argumentation model: the Toulmin model. The Toulmin model is both well-structured and general, and has been shown to be useful for policy debates. In this preliminary work we attempted to generate, with a given topic motion keyword or phrase, Toulmin model arguments by developing a computational model for detecting arguments spanned across multiple documents. This pa- per discusses our subjective results, observa- tions, and future work. 1 Introduction Given an input motion, or claim, the task of auto- matic argumentation generation is to generate co- herent and logically structured argumentation in various scenarios. In this paper, we examined two extreme types of scenarios: (i) an input claim should be supported, and (ii) a counterclaim should be sup- ported. For example, with the House should ban alcohol in schools as an input claim, our goal was to automatically generate supportive output, such as “The input claim is valid because alcohol causes brain damage. Brain damage loses concentration for study.”; and with the House should not ban alco- hol in schools as our counterclaim, our goal, like be- fore, was to generate supportive output, such as “The counterclaim is valid because alcohol makes people fun. Sociality can be developed by pleasure.”. The automatic generation of arguments is a challenging problem that is not only useful for identifying evi- dence of certain claims but also for why the evidence of certain claims is significant. As a basis for generating logically structured out- put, we required the utilization of a structured frame- work ideal for debate arguments. A promising op- tion for accomplishing this goal includes the inte- gration of the Toulmin model [18], which consists of three main components (claim, data, and warrant), where a claim is something an individual believes, data is support or evidence to the claim, and a war- rant is the hypothetical link between the claim and data. When considering this structure for debate topic motions, such as alcohol should be banned, then data such as alcohol causes liver disease and a warrant such as if alcohol causes liver disease, then it should be banned can be supportive for the claim, as the data’s relevance to the claim is provided by the warrant. Although many possibilities exist for constructing a Toulmin model, …

Content truncated to 500 words