Diversity-Based Generalization for Unsupervised Text Classification Under Domain Shift

Domain adaptation approaches seek to learn from a source domain and generalize it an unseen target domain. At present, the state-of-the-art unsupervised for subjective text classification problems leverage unlabeled data along with labeled data. In this paper, we propose novel method of single-task based on simple but effective idea diversity-based generalization that does not require still matches in performance. Diversity plays role promoting model better be indiscriminate towards shift by forcing rely same features prediction. We apply concept most explainable component neural networks, attention layer. To generate sufficient diversity, create multi-head infuse diversity constraint between heads such each head will differently. further expand upon our tri-training designing procedure additional tri-trained classifiers. Extensive evaluation using standard benchmark dataset Amazon reviews newly constructed Crisis events shows fully competing baselines uses Our results demonstrate machine learning architectures ensure can better; encouraging future research design ubiquitously usable models without