Open Text Classification Based on Dynamic Boundary Balance

Abstract Traditional text classification tasks require the test set to contain the same classes (known/seen classes) as the training set. However, when the new classes (unknown/unseen classes) appear during testing, the traditional classifier cannot recognize them effectively. As a result, a qualified open text classifier should be able to classify known classes as well as identify unknown classes, e.i., open classfication. But current open text classifiers perform poorly on the task. To address the issue, we propose a two-stage learning model based on dynamic balance of decision boundary to complete the open text classification task. We call it open text classification with dynamic boundary balance (OCD2B). First, we construct a traditional classifier based on BERT(Bidirectional Encoder Representation from Transformers) model to classify the known classes. Second, the prior knowledge of known classes is used to dynamically determine the decision boundaries between known classes and unknown classes. We propose a boundary loss function as boundary balance strategy to simultaneously reduce open space risk and empirical risk. Experiments on two standard datasets show that our method achieves better performance than existing other methods. In particular, OCD2B has even greater advantages over the others especially when the proportion of unknown classes is large.