Automatically selecting complementary vector representations for semantic textual similarity

Abstract

The goal of the Semantic Textual Similarity task is to automatically quantify the semantic similarity of two text snippets. Since 2012, the task has been organized on a yearly basis as a part of the SemEval evaluation campaign. This paper presents a method that aims to combine different sentence-based vector representations in order to improve the computation of semantic similarity values. Our hypothesis is that such a combination of different representations allows us to pinpoint different semantic aspects, which improves the accuracy of similarity computations. The method's main difficulty lies in the selection of the most complementary representations, for which we present an optimization method. Our final system is based on the winning system of the 2015 evaluation campaign, augmented with the complementary vector representations selected by our optimization method. We also present evaluation results on the dataset of the 2016 campaign, which confirms the benefit of our method.