Differences

This shows you the differences between two versions of the page.

--- courses:rg:2011-report-parser [2012/09/26 13:56]
ufal
+++ courses:rg:2011-report-parser [2012/09/27 11:22] (current)
ufal
@@ Line 1: / Line 1: @@
-====== Using a Wikipedia-based Semantic Relatedness Measure for Document Clustering ======
+====== A Fast, Accurate, Non-Projective, Semantically-Enriched Parser ======
 written by Stephen Tratz and Eduard Hovy (Information Sciences Institute, University of Southern Carolina)
-spoken by Martin Popel
+presented by Martin Popel
 reported by Michal Novák
@@ Line 9: / Line 9: @@
 ===== 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's one. In addition, they extended the non-directional easy-first first algorithm of Goldberg and Elhadad to support non-projective trees by adding "move" actions inspired by Nivre's swap-based reordering for shift-reduce parsing.
+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's one. In addition, they extend the non-directional easy-first first algorithm of Goldberg and Elhadad to support non-projective trees by adding "move" actions inspired by Nivre's swap-based reordering for shift-reduce parsing. Their parser is capable of producing shallow semantic annotations for prepositions, possesives and noun compounds.
 ===== 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's dependency labels
+==== 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 "dep/DEP" relation
+              * Stanford tags are hierarchical and "dep/DEP" is the top-most one
+          * correcting of POS using the syntactic info + additional rules for specific word forms
+          * 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): {{:courses:rg:dependency-conversion.png|}}
+      - additional changes and the final conversion from the intermediate output to a dependency structure
+==== Parser ====
+  * we illustrated a step of the parser:
+{{:courses:rg:ndef_parsing.png|}}
+  * we compared time complexity of this system with other commonly used ones
+| MST parser | <latex>\mathop O(n^2)</latex> | |
+| MALT parser | <latex>\mathop O(n)</latex> | in fact slower |
+| this parser | <latex>\mathop O(n\log(n))</latex> | <latex>\mathop O(n^2)</latex> - naive implementation |
+| this parser - non-projective | <latex>\mathop O(n^2\log(n))</latex> | <latex>\mathop O(n^3)</latex> - naive implementation |
+    * implemented by a heap, it can reach <latex>\mathop O(n\log(n))</latex>; however, they don't use it because it's fast enough
+    * Algorithm1
+        * we weren't sure what the variable "index" is (best index of parent or any index of queue of unprocessed words)
+        * it again confirms that pseudocode is usually more confusing than a normal code or verbal explanation
+    * "move" operation:
+{{:courses:rg:move_operation.png|}}
+==== Features ====
+  * Brown et al. clusters - they are surprisingly used rarely
+==== Features ====
+  * not much discussed in the paper
+==== Evaluation ====
+  * they make the same transformation as they did in Section 2
+==== Shallow semantic annotation ====
+  * 4 optional modules
+      * preposition sense disambiguation
+      * noun compound interpretation
+      * possesives interpretation
+      * PropBank semantic role labelling
+          * (Hajič et al., 2009) is not in the list of references
+===== Conclusion =====
+  * non-directional easy-first parsing
+  * new features - Brown et al. clusters
+  * fast and accurate
+  * modified Penn converter
+      * changes in 9.500 POS tags
+      * labels copula, coordination
+  * semantic info

[ Back to the navigation ] [ Back to the content ]

Institute of Formal and Applied Linguistics Wiki

Differences