Detecting Entailment Using an Extended Implementation of the Basic Elements Overlap Metric

In this paper we evaluate the utility of the recently proposed Basic Elements (BE) summarisation evaluation metric as a means of detecting entailment in a text/hypothesis sentence pair. Basic Elements are tuples representing important syntactic chunks in the Minipar dependency parse of a sentence. We extend the BE matching function with additional semantic knowledge, and show that both the basic and extended metrics outperform a baseline cosine similarity metric. Our results also show that a simple linear combination of BE–style sentence features can achieve an accuracy of 57.0% on the RTE 2006 test set.