Joint Representations of Texts and Labels with Compositional Loss for Short Text Classification

Short text classification is an important foundation for natural language processing (NLP) tasks. Though, the text classification based on deep language models (DLMs) has made a significant headway, in practical applications however, some texts are ambiguous and hard to classify in multi-class classification especially, for short texts whose context length is limited. The mainstream method improves the distinction of ambiguous text by adding context information. However, these methods rely only the text representation, and ignore that the categories overlap and are not completely independent of each other. In this paper, we establish a new general method to solve the problem of ambiguous text classification by introducing label embedding to represent each category, which makes measurable difference between the categories. Further, a new compositional loss function is proposed to train the model, which makes the text representation closer to the ground-truth label and farther away from others. Finally, a constraint is obtained by calculating the similarity between the text representation and label embedding. Errors caused by ambiguous text can be corrected by adding constraints to the output layer of the model. We apply the method to three classical models and conduct experiments on six public datasets. Experiments show that our method can effectively improve the classification accuracy of the ambiguous texts. In addition, combining our method with BERT, we obtain the state-of-the-art results on the CNT dataset.