Research map / Label-efficient and robust learning / Domain adaptation and generalization

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

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain text and graph transfer · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TACIT, a target-agnostic cross-domain text classifier that uses a VAE to split robust (mean) and unrobust (variance) features, distilling the unrobust branch from an easy-sample teacher, and trains only on source-domain data.

Problem

Cross-domain text classification when the target domain is completely agnostic (no unlabeled target data), where models also suffer from shortcut learning in the source domain.

Approach

Disentangle robust and unrobust features in a VAE latent space (mean vs. variance branches), and distill the unrobust branch toward a teacher trained only on high-confidence "easy" source samples, which carry potential shortcuts, so the robust branch alone is used for inference.

Key results

• Single-source with RoBERTa backbone: TACIT averages 93.57%, the best among compared methods, 0.25 points above UDALM (93.32%).
• With BERT backbone, TACIT averages 91.32%, slightly below UDALM (91.74%) but above CFd (90.63%) and COBE (90.39%) while using only source data.
• Multi-source (RoBERTa): TACIT averages 94.65% versus AdSPT 93.75% and Vanilla 93.12%.

Measured effects reported in the paper

Surprising findings

• Vanilla fine-tuned RoBERTa (91.84%) outperforms UDALM with BERT (91.74%), suggesting backbone choice matters more than some domain-adaptation methods.
• TACIT does not beat UDALM on the BERT backbone, and the paper notes that multi-source configurations are not always beneficial (e.g., BDE->K worse than E->K).

Limitations

• Lambda weights are set by loss-magnitude heuristics rather than tuned per task, which may yield non-optimal results.
• Does not outperform UDALM on the BERT backbone and only roughly matches the best methods in the multi-source setting.
• Easy-sample selection is heuristic (top 35% by confidence of an underfitting DistilBERT), and the authors note more judicious selection is needed.

Experimental setup

Datasets: Amazon reviews (Books, DVDs, Electronics, Kitchen) Ling-spam SMS spam collection Spam email dataset SST-5 Twitter sentiment

Benchmarks: Amazon reviews cross-domain (single and multi-source) Spam detection (Ling, SMS, Emails) SST-5/Twitter multi-class sentiment DeBERTa and OPT-1.3b backbone comparisons

Metrics: accuracy

Baselines: DAAT R-PERL CFd UDALM COBE AdSPT Vanilla BERT Vanilla RoBERTa

Builds on: DistilBERT Variational Auto-Encoder UDALM CFd AdSPT

Closest papers by meaning

• Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.85)
• Transferable Semantic Augmentation for Domain Adaptation (CVPR 2021 · similarity 0.85)
• Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.84)
• Towards Generalizing to Unseen Domains with Few Labels (CVPR 2024 · similarity 0.84)
• Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.83)
• Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing (AAAI 2024 · similarity 0.83)