A Convolutional Architecture for Short Text Expansion and Classification

Abstract

In this paper, we propose a convolutional framework for short texts expansion and classification. Particularly, by using additive composition over word embeddings from context with variable window width, the representations of multi-scale semantic units are computed first. Empirically, the semantically related words are usually close to each other in embedding spaces. Thus, the restricted nearest word embeddings of semantic units are chosen to constitute expanded matrices. Then, for a short text, the projected matrix and the expanded matrices are fed to a convolutional neural network. Experimental results on two open benchmarks validate the effectiveness of the proposed method.