Incorporating Task-Oriented Representation in Text Classification

Xue Lei, Yi Cai, Jingyun Xu, Da Ren, Qing Li, Ho-fung Leung

Abstract: Text classification (TC) is an important task in natural language processing. Recently neural network has been applied to text classification and achieves significant improvement in performance. Since some documents are short and ambiguous, recent research enriches document representation with concepts of words extracted from an external knowledge base. However, this approach might incorporate task-irrelevant concepts or coarse granularity concepts that could not discriminate classes in a TC task. This might add noise to document representation and degrade TC performance. To tackle this problem, we propose a task-oriented representation that captures word-class relevance as task-relevant information. We integrate task-oriented representation in a CNN classification model to perform TC. Experimental results on widely used datasets show our approach outperforms comparison models.