Shortest Path and Word Vector Based Relation Representation and Clustering

Abstract

Relation representation plays an important role in text understanding. In this paper, different from previously published supervised methods or semi-supervised methods, an new method of relation representation and clustering based on shortest path and word vector was proposed. By accumulating the word vector along the shortest path within dependency tree, we can not only obtain the essential representation of the relation, but also can map the relation into semantic space simultaneously. Therefore, reliable distance between any two relations could be measured. Moreover, further applications such as relations clustering can be performed conveniently by direct analysis on the collection of vectors.