Backoff Parameter Estimation for the DOP Model

Abstract

The Data Oriented Parsing (DOP) model currently achieves state-of-the-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. In contrast with earlier work, in this paper we show that the DOP parameters constitute a hierarchically structured space of correlated events (rather than a set of disjoint events). The correlations between the different parameters can be expressed by an asymmetric relation called “backoff”. Subsequently, we present a novel recursive estimation algorithm that exploits this hierarchical structure for parameter estimation through discounting and backoff. Finally, we report on experiments showing error reductions of up to 15% in comparison to earlier estimation methods.