Context for semantic similarity calculation in scenario template creation

Abstract

Scenario Template Creation (STC) is a Natural Language Processing (NLP) task to detect the commonalities among articles on similar events and summarize them into an abstract representation -- a scenario template (ST). For this task, the estimation of verb-centric text span similarity is the key. Various approaches have been proposed. Contextual information by intuition would enhance text span similarity estimation. But it has yet to be well exploited. In this thesis, I first devise an intrinsic similarity measure for predicate-argument tuples (PATs). I hypothesize that the semantic similarity between two PATs can also be reflected by their extrinsic similarity. I show experimentally that there is strong correlation between such an extrinsic similarity and the semantic similarity of PATs. To integrate intrinsic and extrinsic similarities for PAT clustering, I propose a graphical framework, using a novel core algorithm called Context Sensitive Clustering (CSC). Experiments show that the proposed algorithm outperforms baseline systems in all the scenarios tested.