Comparing Learning Approaches to Coreference Resolution. There is More to it Than 'Bias'

On the basis of results on three coreference resolution data sets we show that when following current practice in comparing learning methods, we cannot reliably conclude much about their suitability for a given task. In an empirical study of the behavior of representatives of two machine learning paradigms, viz. lazy learning and rule induction on the task of coreference resolution we show that the initial differences between learning techniques are easily overruled when taking into account factors such as feature selection, algorithm parameter optimization, sample selection and their interaction. We propose genetic algorithms as an elegant method to overcome this costly optimization.