SR'11 (English)

Overview

The SR Task data consists of two types of input representations—one shallow, one deep. In both, sentences are represented as sets of unordered labeled dependencies (with the exception of named entities (see Section 2.2.4 below), which are ordered). The shallow input representation is a more ‘surfacey’, syntactic represention of the sentence. The deep(er) input type is closer to a semantic, more abstract, representation of the meaning of the sentence.

The data has been constructed by post-processing the CoNLL 2008 Shared Task data (Surdeanu et al., 2008). For the preparation of the CoNLL-08 Shared task data, the Penn WSJ Treebank was converted to syntactic dependencies via the LTH Constituent-to-Dependency Conversion Tool for Penn-style Treebanks (Pennconverter) (Johansson and Nugues, 2007). The resulting dependency bank was then merged with the Nombank (Meyers et al., 2004) and Propbank corpora (Palmer et al., 2005). Named entity information from the BBC Entity Type corpus was also integrated into the CoNLL-08 data. Our shallow level representation is based on the Pennconverter dependencies. The semantic level representation is derived from the merged Nombank, Propbank and syntactic dependencies in a process similar to the graph completion algorithm outlined in (Bohnet et al., 2010) (see Section 2.2, in particular Subsection 2.2.2, for differences).

Documentation and Instructions for Participants:

SR-Task-2011-Doc.pdf

Shared Task Results:

pdf bib The First Surface Realisation Shared Task: Overview and Evaluation Results