Text Classification with Enriched Word Features

Abstract

Text classification is a fundamental task in natural language processing. Most existing text classification models focus on constructing sophisticated high-level text features but ignore the importance of word features. Those models only use low-level word features obtained from a linear layer as input. To explore how the quality of word representations affects text classification, we propose a deep architecture which can extract high-level word features to perform text classification. Specifically, we use different temporal convolution filters, which vary in size, to capture different contextual features. Then a transition layer is used to coalesce the contextual features and form an enriched high-level word representations. We also find that word feature reuse is useful in our architecture to enrich word representations. Extensive experiments on six publically available datasets show that enriched word representations can significantly improve the performance of classification models.