Transition-based dependency DAG parsing using dynamic oracles

Abstract

In most of the dependency parsing studies, dependency relations within a sentence are often presented as a tree structure. Whilst the tree structure is sufficient to represent the surface relations, deep dependencies which may result to multi-headed relations require more general dependency structures, namely Directed Acyclic Graphs (DAGs). This study proposes a new dependency DAG parsing approach which uses a dynamic oracle within a shift-reduce transitionbased parsing framework. Although there is still room for improvement on performance with more feature engineering, we already obtain competitive performances compared to static oracles as a result of our initial experiments conducted on the ITU-METU-Sabanci Turkish Treebank (IMST).