TACIT: A Target-Agnostic Feature Disentanglement Framework for Cross-Domain Text Classification

• DOI: 10.48550/arxiv.2312.17263
• Publication Date: 2023-01-01
• Authors: Rui Song; Fausto Giunchiglia; Yingji Li; Mingjie Tian; Hao Xu
• Keywords: Feature (linguistics)

Cross-domain text classification aims to transfer models from label-rich source domains label-poor target domains, giving it a wide range of practical applications. Many approaches promote cross-domain generalization by capturing domain-invariant features. However, these methods rely on unlabeled samples provided the which renders model ineffective when domain is agnostic. Furthermore, are easily disturbed shortcut learning in domain, also hinders improvement ability. To solve aforementioned issues, this paper proposes TACIT, agnostic feature disentanglement framework adaptively decouples robust and unrobust features Variational Auto-Encoders. Additionally, encourage separation features, we design distillation task that compels approximate output teacher. The teacher trained with few easy carry potential unknown shortcuts. Experimental results verify our achieves comparable state-of-the-art baselines while utilizing only data.