Abstract
As a subtask of qualitative media reputation research, human annotators manually encode the polarity of actors in media products. Seeking to automate this process, wehave implemented twobaseline classifiers that categorize actors in newspaper articles under six and four polarityclasses. Experiments have shown that our approach is not suitable for distinguishing between six finegrained classes, which has turned out to bedifficult for humans also. In contrast, we have obtained promising results for the fourclass model, through which we argue thatautomated sentiment analysis has a considerable potential in qualitative reputation research.