Open Text Classification Based on Dynamic Boundary Balance

Ganlin Xu, Jianzhou Feng, Qikai Wei

Abstract: Open classification is the problem where there exist some unseen/unknown classes in the test set, i.e., these unknown/unseen classes don’t appear when the model is trained. Existing work often maps samples to high-dimensional space to make decisions, which leads to unobservable and inexplicable results. To address the issue, we shift perspectives to two-dimensional space and put forward a two-stage learning method built on the dynamic decision boundaries balance. We refer it to open classification with dynamic boundary balance (OCD2B). First, we construct a vanilla classifier via known classes with BERT model. Then, we use the prior knowledge of known classes to dynamically determine the decision boundaries between known classes and unknown classes in low-dimensional space. We propose a novel boundary loss function as a boundary balance strategy to reduce open space risk and empirical risk. Experimental results on two standard datasets show that our method achieves performance gain over existing methods, providing easily observable results. In particular, the larger the ratio of unseen classes is, the more obvious the performance advantage the model achieves.