Abstract : Many argumentation tools have been proposed nowadays to support the users in on-line social discussions. However, the main drawback of these tools is that they do not cope with the automatic generation of the arguments from the natural language discussions of the users. In this paper, we propose to use a technique from computational linguistics, namely textual entailment, to generate in an automatic way the abstract arguments from the dialogues. The abstract arguments as well as their relationships are then structured in an argumentation graph to evaluate the dialogue as a whole. The success criteria of the proposed approach is that it is able to represent the dynamics of the dialogues among users allowing to find the use of argumentation natural enough to be really adopted.