Baseline We compare IPL with other entity annotation methods. Our first group of baselines includes entity linking systems in domains of general text, Wikiminer (Milne and Witten, 2008), and short text, Tagme (Ferragina and Scaiella,2012). For each method, we use the default parameter settings, apply them for the individual tweets,and take the average of the annotation confidence scores as the prominence ranking function. The second group of baselines includes systems specifically designed for microblogs. For the contentbased methods, we compare against Meij et al.(2012), which uses a supervised method to rank entities with respect to tweets. We train the model using the same training data as in the original paper.For the graph-based method, we compare against KAURI (Shen et al., 2013), a method which uses user interest propagation to optimize the entity linking scores. To tune the parameters, we pickup four hashtags from different clusters, randomly sample 50 tweets for each, and manually annotate the tweets. For all baselines, we obtained the implementation from the authors. The exception is Meij method, where we implemented ourselves,but we clarified with the authors via emails on several settings. In addition, we also compare three variants of our method, using only local functions for entity ranking (referred to as M , C, and T for mention, context, and time, respectively)
đang được dịch, vui lòng đợi..
