Active Learning for Anaphora Resolution

In this paper we present our experiments with active learning to improve the performance of our probabilistic anaphora resolution system. We have adopted entropy-based uncertainty measures to select new instances to be added to our training data. The actively selected instances, however, were not more successful in improving the performance of the system than the same amount of randomly selected instances. The uncertainty measures we used behave differently from each other when selecting new instances, but none of them achieved remarkable performance. Further studies on active sample selection for anaphora resolution are necessary.