Sgardine writesup Bunescu 2005 Shortest Path

This is a review of Bunescu_2005_a_shortest_path_dependency_kernel_for_relation_extraction by user:Sgardine

Summary

The authors propose a new kernel which allows margin classifiers to achieve better performance on relation extraction tasks; the authors believe that the kernel succinctly represents the information required for the task. The kernel's feature space is the set of all word-classes of the words in the shortest path in the dependency graph of the sentence (each feature being binary). This can be efficiently computed without enumerating the prohibitively large space. The kernel was evaluated against previous kernels, using both CCG and the Collins parser for construction of the underlying dependency graph. The best results come from their kernel calculated with the Collins parser.

Commentary

Maybe I'm just grouchy, but I don't like their example sentence: "protesters seized several stations" I'm not completely sure the protesters are located in the stations, as in "Castro seized the Hilton" possibly from the comfort of his own desk. Which is ultimately fine, there are always judgment calls, but I got distracted seeing it as an example again and again.

Would be interesting to compare to more restricted sets of word classes to see which ones are helping most.