Ensembles of diverse clustering-based discriminative dependency parsers

摘要: Syntactic parsing and dependency in particular are a core component of many Natural Language Processing (NLP) tasks applications. Improvements can help improve machine translation information extraction applications among others. In this thesis, we extend the framework (Koo, Carreras, Collins, 2008) for which uses single clustering method semi-supervised learning. We make use multiple diverse methods to build discriminative models Maximum Spanning Tree (MST) (McDonald, Crammer, Pereira, 2005). All these clustering-based parsers then combined together using novel ensemble model, performs exact inference on shared hypothesis space all parser models. show that significantly improves unlabeled accuracy from 90.82% 92.46% Section 23 Penn Treebank. also significant improvements domain adaptation Switchboard Brown corpora.