Abstract: "Opinion mining" has been a flourishing application of Computational Linguistics for several years now, one reason being its commercial relevance in tracking opinions about products on web sites. For these purposes, current systems sometimes yield respectable results. But when moving to opinions whose target is less clear-cut (as for instance in political debate), the task becomes much more difficult. In my talk, I will present work on the SO-CAL opinion recognition system, where I have collaborated with colleagues at
Simon-Fraser University (Taboada et al. 2011), and discuss the various problems surfacing when applying this lexicon-based approach to different kinds of editorial text. One central problem is the role of discourse structure, which needs to be taken into account for appropriately recognizing author's stances. In our current work at Potsdam, we combine layers of discourse structure (most importantly coreference and connectives with their scopes) with the sentence-based opinion recognition as in SO-CAL; I present some first results of these efforts.