Improving Domain-Generalized Few-Shot Text Classification with Multi-Level Distributional Signatures

Domain-generalized few-shot text classification (DG-FSTC) is a new setting for (FSTC). In DG-FSTC, the model meta-trained on multi-domain dataset, and meta-tested unseen datasets with different domains. However, previous methods mostly construct semantic representations by learning from words directly, which limited in domain adaptability. this study, we enhance adaptability of utilizing distributional signatures texts that indicate domain-related features specific We propose Multi-level Distributional Signatures based model, namely MultiDS. Firstly, inspired pretrained language models, compute an extra large news corpus, denote these as domain-agnostic features. Then calculate same class, respectively. These two kinds information are regarded domain-specific class-specific features, After that, fuse translate three into word-level attention values, enables to capture informative changes. addition, utilize calibration feature The vectors produced word embeddings help adapt various Extensive experiments performed four benchmarks. results demonstrate our proposed method beats state-of-the-art average improvement 1.41% datasets. Compared five competitive baselines, achieves best performance. ablation studies prove effectiveness each module.