Publication at Faculty of Mathematics and Physics |
2015
Abstract
We describe a transition-based, non-projective dependency parser which uses a neural network classifier for prediction and requires no feature engineering. We propose a new, search-based oracle, which improves parsing accuracy similarly to a dynamic oracle, but is applicable to any transition system, such as the fully non-projective swap system.
The parser has excellent parsing speed, compact models, and achieves high accuracy without requiring any additional resources such as raw corpora. We tested it on all 19 treebanks of the Universal Dependencies project.
The C++ implementation of the parser is being released as an open-source tool.