Dependency Parsing with Structure Knowledge

A new probability-based parsing model DPSK (Dependency Parsing with Structure Knowledge) is presented for dependency parsing. Similar to a bottom-up chart parsing algorithm, DPSK select the best dependency arc between two words in a sentence according to the probability. The arc probability depends on two kinds of information: (1) Features extracting from two words of the arc. (2) Features extracting according to arc child word and those children words of arc parent. The first kind feature information is relevant to the case grammar theory, while the second kind feature information is relevant to Context-free grammars. A Maximum Entropy model is used to train and calculate single arc probabilities. DPSK is evaluated experimentally using the dataset distributed in CoNLL 2008 share-task. An unlabelled arc score of 90.2 % is reported. This work will contribute to and stimulate other researches in the field of parsing.