From Sentiment to Reputation

We report on our participation in the profiling task of the first edition of the CLEF RepLab evaluation initiative. We assume that a statement—such as a tweet—that caries negative sentiment can have a positive impact on the reputation of the entity it talks about (and vice versa). Our model directly captures this impact by observing the reactions—such as replies—the statement solicits. We present the assumptions behind our model and the model itself. We find that given the current setting, results on the test set are strongly entity-dependent and that the test data is very different from the trial data. We conclude with a proposal on how to create a task that avoids such dataset dependent problems.