Differences
This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision Next revision Both sides next revision | ||
courses:rg:2011-report-parser [2012/09/26 13:56] ufal |
courses:rg:2011-report-parser [2012/09/27 10:39] ufal |
||
---|---|---|---|
Line 3: | Line 3: | ||
written by Stephen Tratz and Eduard Hovy (Information Sciences Institute, University of Southern Carolina) | written by Stephen Tratz and Eduard Hovy (Information Sciences Institute, University of Southern Carolina) | ||
- | spoken | + | presented |
reported by Michal Novák | reported by Michal Novák | ||
Line 9: | Line 9: | ||
===== Introduction ===== | ===== Introduction ===== | ||
- | The paper describes a high-quality conversion of Penn Treebank to dependency trees. The authors introduce an improved labeled dependency scheme based on the Stanford' | + | The paper describes a high-quality conversion of Penn Treebank to dependency trees. The authors introduce an improved labeled dependency scheme based on the Stanford' |
===== Notes ===== | ===== Notes ===== | ||
+ | ==== Dependency conversion structure ==== | ||
+ | |||
+ | * in general, there are (at least) 3 possible types of dependency labels: | ||
+ | * unlabeled - is it really a set of labels? | ||
+ | * coarse labels of the CoNLL tasks | ||
+ | * 10-20 labels | ||
+ | * for example NMOD is always under a noun - it's an easy task and the result is not quite useful | ||
+ | * their scheme is based on the Stanford' | ||
+ | |||
+ | ==== Conversion process ==== | ||
+ | |||
+ | * converting phrase trees of Penn Treebank to dependency ones | ||
+ | * it consists of 3 steps: | ||
+ | - add structure to flat NPs | ||
+ | - constituent-to-dependency converter with some head-finding rule modifications | ||
+ | * a list of rules in Figure 2 is hardly understandable without reading a paper their conversion method is related to | ||
+ | * they reduced the number of generic " | ||
+ | * Stanford tags are hierarchical and " | ||
+ | * 1.3% of arcs are non-projective (out of 8.1% of all non-projective arcs) because of the following conversion (agreement can be a motivation for this, i.e. in Czech): | ||
+ | {{: | ||
+ | | ||
+ | ==== Parser ==== | ||
+ | |||
+ | * we illustrated a step of the parser: | ||
+ | {{: | ||
+ | * we compared time complexity of this system with other commonly used ones |