Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Collaborative Information Bottleneck for Text Readability Assessment
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 SCIB-ARA, a hybrid readability assessment model that fuses document-, sentence-, word-level BigBird deep features and linguistic features via a self-supervised collaborative information bottleneck module capturing common and specific information.
Problem
Hybrid automatic readability assessment models fuse deep and linguistic features with orthogonal projection, which ignores specific-intrinsic information of each feature level and is sensitive to the choice of base feature.
Approach
Replace orthogonal projection fusion with a self-supervised information bottleneck: a common encoder-decoder over all four representations plus per-level specific encoder-decoder pairs with reconstruction losses, whose refined outputs are concatenated for classification.
Key results
- SCIB-ARA reaches 95.56% accuracy on Cambridge and 94.07% on Weebit, versus 90.00% and 93.12% for InterpretARA.
- On CMER, SCIB-ARA achieves 31.79% accuracy vs 25.39% for InterpretARA, the best of the baselines.
- Removing the self-supervised scheme lowers accuracy by 2.22 to 6.84 points across five corpora; removing both common and specific sub-modules lowers accuracy by about 4.58 points on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SCIB-ARA instead of PromptARA | Readability assessment · WeeBit | accuracy: 93.12 -> 94.07 | Helped |
| Use SCIB-ARA instead of PromptARA | Readability assessment · Cambridge | accuracy: 91.11 -> 95.56 | Helped |
| Use SCIB-ARA instead of InterpretARA | Readability assessment · CLEAR | accuracy: 83.30 -> 83.72 | Helped |
| Use SCIB-ARA instead of PromptARA | Readability assessment · Newsela | accuracy: 88.40 -> 87.57 | Hurt |
| Use SCIB-ARA instead of InterpretARA | Readability assessment · CMT | accuracy: 44.87 -> 45.79 | Helped |
| Use SCIB-ARA instead of DTRA | Readability assessment · CMER | accuracy: 26.50 -> 31.79 | Helped |
| Add the common representation (ablation) | SCIB-ARA · Cambridge | accuracy: 94.44 -> 95.56 | Helped |
| Add the specific representations (ablation) | SCIB-ARA · WeeBit | accuracy: 93.44 -> 94.07 | Helped |
| Add both common and specific representations (ablation) | SCIB-ARA · CMER | accuracy: 24.28 -> 31.79 | Helped |
| Add the self-supervised scheme (ablation) | SCIB-ARA · Cambridge | accuracy: 88.89 -> 95.56 | Helped |
| Add the linguistic representation (ablation) | SCIB-ARA · WeeBit | accuracy: 94.01 -> 94.07 | No clear effect |
Surprising findings
- Chinese corpora remain hard: accuracy is only 31.79% on CMER and 45.79% on CMT, which the authors attribute to semantic ambiguity.
- The ablation shows the specific sub-module gains slightly more than the common sub-module on average.
Limitations
- Lee-2021 and BERT-FP-LBL results are copied from the literature because their code was not available, so comparisons are not fully controlled.
- Accuracy on the Chinese corpora is low, indicating the approach struggles with Chinese readability.
Experimental setup
Datasets: WeeBit Cambridge Newsela CLEAR CMT CMER
Benchmarks: WeeBit Cambridge Newsela CLEAR CMT CMER
Metrics: Accuracy Precision macro F1 quadratic weighted kappa
Baselines: BERT BigBird HAN DTRA Lee-2021 BERT-FP-LBL PromptARA InterpretARA
Builds on: information bottleneck orthogonal projection fusion HAN BigBird PromptARA InterpretARA
Closest papers by meaning
- InterpretARA: Enhancing Hybrid Automatic Readability Assessment with Linguistic Feature Interpreter and Contrastive Learning (AAAI 2024 · similarity 0.89)
- Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability (ICLR 2025 · similarity 0.84)
- RMIB: Representation Matching Information Bottleneck for Matching Text Representations (ICML 2024 · similarity 0.82)
- NovelQA: Benchmarking Question Answering on Documents Exceeding 200K Tokens (ICLR 2025 · similarity 0.82)
- Concept Bottleneck Large Language Models (ICLR 2025 · similarity 0.82)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · 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.