Word Embedding of Dimensionality Reduction for Document Clustering

Word embedding models have been extensively used in document analysis. Even though many models have been created for embedding documents into vector spaces, their document clustering performance is not noticeably better than that of conventional bag-of-words representations. This paper proposes a document clustering called Word Embedding of Dimensionality Reduction (WERD) that can be used in conjunction with any word embedding method and can provide a semantic explanation of the clustering outcomes. Stopwords and a lexical reduction are first used to preprocess the documents. A pre-trained embedding model is used to embed documents. Then a dimension reduction is used to reduce the dimension of the embedded data to remove redundant features and create more compact document vectors used as document features for clustering. After clustering, the Non-Negative Matrix Factorization approach extracts the keywords from each cluster to produce semantic descriptions. Numerous experiments on two datasets show that WERD can produce superior clustering results.