Research map / Label-efficient and robust learning / Recognition and long-tail
Counterfactual-Enhanced Information Bottleneck for Aspect-Based Sentiment Analysis
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 CEIB, which uses T5-XXL multi-pattern cloze prompting to generate counterfactual ABSA samples and an information bottleneck objective that balances factual and counterfactual label information to reduce spurious correlations in BERT-based aspect sentiment classification.
Problem
Deep ABSA models learn spurious correlations between superficial context patterns and sentiment labels, which hurts robustness on out-of-distribution and adversarial data.
Approach
Combines LLM-generated counterfactual samples (same spurious context, opposite label) with a variational information bottleneck objective that splits mutual information into factual and counterfactual parts and balances their label predictiveness.
Key results
- CEIB reaches 87.77% accuracy and 82.08% macro-F1 on REST14 with BERT-base, outperforming SenticGCN-BERT and SSEGCN-BERT.
- On ARTS-REST14, CEIB drops only 7.40% accuracy versus 14.96% for RGAT-BERT and 26.30% for BERT-SPC under adversarial test sets.
- Ablations show removing the IB module causes a sharp drop (e.g., REST14 accuracy 87.77 to 85.54).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the information bottleneck module (ablation) | CEIB · REST14 | accuracy: 85.54 -> 87.77 | Helped |
| Add the information bottleneck module (ablation) | CEIB · LAP14 | F1: 75.96 -> 79.50 | Helped |
| Add the information bottleneck module (ablation) | CEIB · REST15 | F1: 69.44 -> 72.97 | Helped |
| Add the information bottleneck module (ablation) | CEIB · REST16 | F1: 76.96 -> 81.08 | Helped |
| Add the information bottleneck module (ablation) | CEIB · MAMS | accuracy: 83.28 -> 84.95 | Helped |
| Add counterfactual data augmentation (ablation) | CEIB · REST14 | F1: 80.00 -> 82.08 | Helped |
| Add counterfactual data augmentation (ablation) | CEIB · LAP14 | accuracy: 80.94 -> 82.92 | Helped |
| Add counterfactual data augmentation (ablation) | CEIB · REST15 | accuracy: 85.54 -> 86.16 | Helped |
| Add counterfactual data augmentation (ablation) | CEIB · REST16 | F1: 77.74 -> 81.08 | Helped |
| Add counterfactual data augmentation (ablation) | CEIB · MAMS | accuracy: 84.06 -> 84.95 | Helped |
| Use CEIB instead of the strongest baseline | ABSA models · REST14 | F1: 81.35 -> 82.08 | Helped |
| Use CEIB instead of the strongest baseline | ABSA models · LAP14 | accuracy: 82.12 -> 82.92 | Helped |
Surprising findings
- Naively combining original and counterfactual data without modeling their interaction degrades performance, per the authors' w/o IB ablation.
Limitations
- Relies on BERT-base as encoder and the generated counterfactuals' quality depends on T5-XXL and BERT-based selection.
- Gains on some datasets are modest (e.g., MAMS accuracy 84.95 vs 83.98 baseline).
Experimental setup
Datasets: REST14 LAP14 REST15 REST16 MAMS ARTS
Benchmarks: REST14 LAP14 REST15 REST16 MAMS REST14-ARTS LAP14-ARTS
Metrics: accuracy macro-F1
Baselines: BERT-SPC BERT-PT CapsNet-BERT DGEDT-BERT RGAT-BERT DualGCN-BERT TGCN-BERT SenticGCN-BERT SSEGCN-BERT
Builds on: Variational Information Bottleneck BERT T5 counterfactual data augmentation
Closest papers by meaning
- Aspect-Based Sentiment Analysis with Explicit Sentiment Augmentations (AAAI 2024 · similarity 0.84)
- A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.82)
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.82)
- DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (AAAI 2026 · similarity 0.82)
- Counterfactual Concept Bottleneck Models (ICLR 2025 · similarity 0.82)
- Correlation-Decoupled Knowledge Distillation for Multimodal Sentiment Analysis with Incomplete Modalities (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.