Applying the Collins Parser to Czech

Group Participants

Michael Collins
Frederick Jelinek
Lance Ramshaw
Oren Schwartz
Christoph Tillmann

About the Collins Parser

Background about parsing lexicalized grammars (talk given 29th July 1998)
Tutorial about probabilistic parsing (given in pre-workshop week)
For workshop participants: Instructions for running the parser
Description of the baseline system (conversion from dependencies to trees)
First progress update (given August 5th 1998)

Final Presentations

POSTags Results (given August 19th 1998)

Workshop Research Subgroups

Mapping from dependencies to trees: handling of coordination
Mapping from morphology strings to POS tags
Dealing with crossing dependencies

Summary Of Results

Initial work involved testing and training on a preliminary version of the data. A baseline system (which built flat parse trees directly from the dependencies and used single letter POS tags) scored 71% in terms of correctly-predicted dependencies. During the second week, initial improvements in the tree construction and POS tagset led to a score on that data set of 77.5%.