UWB Participation in the Multiling’s OnForumS Task

This paper presents a system used for the Online Forum Summarization task of Multiling 2015. We drafted an approach to all 3 subtasks: linking comment sentences to relevant content of the article, detecting sentiment polarity of the comment and agreement between the linked texts. For the comment linking we use vector space model and latent dirichlet allocation. The sentiment and argument structure labeling is based on a maximum entropy classifier. The preliminary results indicate a good precision for English but worse for Italian.