Corpus Statistics Empowered Document Classification

Abstract

In natural language processing (NLP), document classification is an important task that relies on the proper thematic representation of the documents. Gaussian mixture-based clustering is widespread for capturing rich thematic semantics but ignores emphasizing potential terms in the corpus. Moreover, the soft clustering approach causes long-tail noise by putting every word into every cluster, which affects the natural thematic representation of documents and their proper classification. It is more challenging to capture semantic insights when dealing with short-length documents where word co-occurrence information is limited. In this context, for long texts, we proposed Weighted Sparse Document Vector (WSDV), which performs clustering on the weighted data that emphasizes vital terms and moderates the soft clustering by removing outliers from the converged clusters. Besides the removal of outliers, WSDV utilizes corpus statistics in different steps for the vectorial representation of the document. For short texts, we proposed Weighted Compact Document Vector (WCDV), which captures better semantic insights in building document vectors by emphasizing potential terms and capturing uncertainty information while measuring the affinity between distributions of words. Using available corpus statistics, WCDV sufficiently handles the data sparsity of short texts without depending on external knowledge sources. To evaluate the proposed models, we performed a multiclass document classification using standard performance measures (precision, recall, f1-score, and accuracy) on three long- and two short-text benchmark datasets that outperform some state-of-the-art models. The experimental results demonstrate that in the long-text classification, WSDV reached 97.83% accuracy on the AgNews dataset, 86.05% accuracy on the 20Newsgroup dataset, and 98.67% accuracy on the R8 dataset. In the short-text classification, WCDV reached 72.7% accuracy on the SearchSnippets dataset and 89.4% accuracy on the Twitter dataset.