Learning refined features for open-world text classification with class description and commonsense knowledge

Open-world classification requires a classifier not only to classify samples of the observed classes but also to detect samples which are not suitable to be classified as the known classes. State-of-the-art methods train a feature extractor to extract features for separating known classes with limited training data. Then some strategies, such as outlier detector, are used to reject samples from unknown classes based on the feature space. However, they are prone to extract the discriminative features among known classes and cannot model comprehensive features of known classes, which causes the classification errors when detecting the samples from the unknown classes in an open world scenario. Motivated by the theory of psychology and cognitive science, we utilize both class descriptions and commonsense knowledge summarized by human to refine the discriminant features and propose a regularization strategy. The regularization is incorporated into the feature extractor, which is enabled to further improve the performance of our model in an open-world environment. Extensive experiments and visualization analysis are conducted to evaluate the effectiveness of our proposed model.