A Probabilistic Model for Dependency Parsing Considering Ascending Dependencies

In this study, we propose a new probability model for disambiguation in dependency parsing. In order to enhance accuracy, early dependency models used two parameters: distance and the part of speech tag of the left sister derived from the word order within a sentence. However, these parameters are not appropriate in free order or elliptical languages such as German, Korean, and Russian. In the proposed model, instead of adopting parameters based on word order, ascending dependency was used to improve accuracy. Ascending dependency is defined as a relationship between a word and any ascendant word in the dependency hierarchy. This approach is useful for free order and the frequently elliptical languages, because in the dependency grammar the relationship between words is critical. The performance of the three models, including the proposed models of Collins and Eisner, was compared, when they were applied to Korean. These models were trained by 9,000 sentences and tested by 300 sentences sampled from the KIBS tree bank. The result showed that the proposed model works better in accuracy than Collins' model by 2.4% and Eisner's model by 2.1%.