Benchmarking for syntax-based sentential inference

We propose a methodology for investigating how well NLP systems handle meaning preserving syntactic variations. We start by presenting a method for the semi automated creation of a benchmark where entailment is mediated solely by meaning preserving syntactic variations. We then use this benchmark to compare a semantic role labeller and two grammar based RTE systems. We argue that the proposed methodology (i) supports a modular evaluation of the ability of NLP systems to handle the syntax/semantic interface and (ii) permits focused error mining and error analysis.