Downstream use of syntactic analysis: does representation matter?

Research in syntactic parsing is largely driven by progress in intrinsic evaluation and there have been impressive developments in recent years in terms of evaluation measures, such as F-score or labeled accuracy. At the same time, a range of different syntactic representations have been put to use in treebank annotation projects and there have been studies measuring various aspects of the ”learnability” of these representations and their suitability for automatic parsing, mostly also evaluated in terms of intrinsic measures. In this talk I will provide a different perspective on these developments and give an overview of research that examines the usefulness of syntactic analysis in downstream applications. The talk will discuss both constituency-based and dependency-based representations, with a focus on various flavours of dependency-based representations, ranging from purely syntactic representations to more semantically oriented representations. The recently completed shared task on Extrinsic Parser Evaluation was aimed at assessing the utility of different types of dependency representations for downstream applications and I will discuss some of our findings based on the results from this task as well as follow-up experiments and analysis.