IR-776: (2010) Wu, X. and Smith, D., "Right-branching tree transformation for eager dependency parsing," CIIR Technical Report. [View bibtex]
This paper presents a reversible transformation on dependency trees
that produces purely right-branching structures. On their own, these
structures are easier for stack-based shift-reduce parsers to
learn. We describe experiments training parsers on the transformed
trees, parsing raw text, and then reversing the transformation on
the results to evaluate. While these conditions do not significantly
change parsing accuracy, a parser tailored to exploit the greater
predictability of the transformed trees is able to run ten times
faster than the widely-used MaltParser system, while using the same
modeling and learning infrastructure. We conclude with experiments
that explore further search strategies enabled by the transformed
structures.