Using uncertainty for multi-domain text classification

Abstract

Multi-domain learning allows for joint feature detection to promote the performance on a learning task. The shared feature space, however, has limited capacity and should include only the most discriminative task-independent features that are useful for all the tasks. To this end, we proposed a global-local task uncertainty measure to monitor the usefulness of features for all tasks, increasing their effectiveness and generalizability while disentangling them from task-specific features that are not helpful for other tasks. Besides, this measure can utilize unlabeled domain data, tapping the vast reserves of unlabeled data to have even better features. An experiment on a multi-domain text classification shows that the proposed method consistently improves the baseline’s performance and improves the knowledge transfer of learned features to unseen data.