Comparative Analysis of DeepBank and the Penn Treebank

Abstract

I examined the differences between the DeepBank and Penn Treebank and the effect of hand-created and grammar-derived annotations on dependency representations. The dependencies comparison involved transforming the DeepBank trees into Penn Treebank format, training the Stanford parser on the resulting output, and testing the trained parser vs known dependencies data; this task yielded a null result. A detailed analysis of the remaining differences between the Penn Treebank and modified DeepBank was done after the transformation process, showing many differences including parse selection, clause and phrase attachment, labeling of modifiers, and the treatment of proper noun phrases like movie titles. This yielded useful information for future work in this area.