A new knowledge based probabilistic dependencyparsing (KPDP) is presented to overcome the localoptimization problem of native probabilistic models. KPDPis composed of two stages: (1) selecting a set of constituentparse trees with an extensive bottom-up chart parsingalgorithm which employs Maximum Entropy Models tocalculate single arc probabilities; (2) finding the bestparsing tree with the help of word knowledge. Differentfrom previous studies, we incorporate word knowledge intoparsing procedure. Based on case grammar theory, theword knowledge is represented as some patterns whichgroup those arcs with the same head. Thus, the KPDPcontains both single arc information and the relationshipinformation between relevant arcs. KPDP is evaluatedexperimentally using the dataset distributed in CoNLL 2008share-task. An unlabelled arc score of 87 % is reported,which is 3.39% higher than the native model without wordknowledge. This work will contribute to and stimulate otherresearches in the field of parsing.
展开▼
