Unsupervised Entailment Detection between Dependency Graph Fragments

Entailment detection systems are generally designed to work either on single words, relations or full sentences. We propose a new task – detecting entailment between dependency graph fragments of any type – which relaxes these restrictions and leads to much wider entailment discovery. An unsupervised framework is described that uses intrinsic similarity, multi-level extrinsic similarity and the detection of negation and hedged language to assign a confidence score to entailment relations between two fragments. The final system achieves 84.1% average precision on a data set of entailment examples from the biomedical domain.