Centroid-Means-Embedding: An Approach to Infusing Word Embeddings into Features for Text Classification

This paper presents word embedding-based approach to text classification. In this study, we introduce a new vector space model called Semantically-Augmented Statistical Vector Space Model (SAS-VSM) that is a statistical VSM with a semantic VSM for information access systems, especially for automatic text classification. In the SAS-VSM, we first implement a primary approach to concatenate continuous-valued semantic features with an existing statistical VSM. We, then, introduce the Centroid-Means-Embedding (CME) method that updates existing statistical feature vectors with semantic knowledge. Experimental results show that the proposed CME-based SAS-VSM approaches are promising over the different weighting approaches on the 20 Newsgroups and RCV1-v2/LYRL2004 datasets using Support Vector Machine (SVM) classifiers to enhance the classification tasks. Our approach outperformed other approaches in both micro-F\(_1\) and categorical performance.