A novel similarity measure for dependency trees

We propose a new tree similarity measure based on the Connectivity-Integrality Principle (CI). CI is a concept from graph theory that makes the tree similarity to be measured by taking account of both the partial and the integral structures of trees. We have proved in theory that the new measure has better flexibility than the other two typical similarity measures, when given a variety of common substructures between two trees. We apply this new measure to a specific-domain QA system to undertake the task of sentence-level disambiguation. Experimental results show that, the new model can effectively enhance the rejection rate that aims at irrelevant documents.