Backoff DOP: Parameter Estimation by Backoff

The Data Oriented Parsing (DOP) model currently achieves state-ofthe-art parsing on benchmark corpora. However, existing DOP parameter estimation methods are known to be biased, and ad hoc adjustments are needed in order to reduce the effects of these biases on performance. This paper presents a novel estimation procedure that exploits a unique property of DOP: different derivations can generate the same parse-tree. We show that the different derivations represent different “Markov orders” that the DOP model interpolates together. The idea behind the present method is to combine the different derivation orders by backoff instead of interpolation. This allows for a novel estimation procedure that employs Katz backoff for estimation. We report on experiments showing error reduction of up to 15% with respect to earlier methods.