Research map / Label-efficient and robust learning / Recognition and long-tail
InterpretARA: Enhancing Hybrid Automatic Readability Assessment with Linguistic Feature Interpreter and Contrastive Learning
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 InterpretARA, a hybrid readability assessment model that verbalizes linguistic features into natural-language templates, encodes them with BERT, and combines them with document- and segment-level BERT representations plus a supervised contrastive loss.
Problem
Hybrid automatic readability assessment models underuse linguistic features because they treat feature values as numbers and ignore their structure, and they do not model relative difficulty among texts.
Approach
A linguistic feature interpreter turns each feature name and value into a natural-language sentence that is encoded by BERT in the deep representation space, combined with segment-level BERT representations and a contrastive loss over document representations, with orthogonal projection filtering to remove redundancy.
Key results
- InterpretARA reaches 93.12% accuracy on Weebit and 90.00% on Cambridge, outperforming BERT-FP-LBL (92.70, 87.78) and BigBird on these corpora.
- On Newsela, InterpretARA (87.15% accuracy) is second to BigBird (87.34%); on the Chinese CMT corpus it reaches 44.87% accuracy versus 44.42% for DTRA.
- Ablation on Cambridge shows removing linguistic features drops F1 from 89.88 to 81.44, and removing segment-level representations drops it further to 71.93.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the linguistic feature interpreter (ablation) | InterpretARA · Weebit | accuracy: 92.80 -> 93.12 | Helped |
| Add the linguistic feature interpreter (ablation) | InterpretARA · Cambridge | accuracy: 82.22 -> 90.00 | Helped |
| Add the linguistic feature interpreter (ablation) | InterpretARA · CMT | accuracy: 43.04 -> 44.87 | Helped |
| Add contrastive learning to the model without linguistic features (ablation) | InterpretARA · Weebit | accuracy: 91.96 -> 92.80 | Helped |
| Add contrastive learning to the model without linguistic features (ablation) | InterpretARA · Cambridge | accuracy: 80.00 -> 82.22 | Helped |
| Add contrastive learning to the model without linguistic features (ablation) | InterpretARA · CMT | accuracy: 40.11 -> 43.04 | Helped |
| Add segment-level representations to the model without linguistic features and contrastive learning (ablation) | InterpretARA · Weebit | accuracy: 91.43 -> 91.96 | Helped |
| Add segment-level representations to the model without linguistic features and contrastive learning (ablation) | InterpretARA · Cambridge | accuracy: 73.33 -> 80.00 | Helped |
| Add segment-level representations to the model without linguistic features and contrastive learning (ablation) | InterpretARA · CMT | accuracy: 39.74 -> 40.11 | Helped |
| Use raw linguistic feature concatenation (ablation) | InterpretARA · Weebit, Cambridge, CMT | accuracy: 92.17; 85.56; 37.91 | Mixed |
| Use linear transformation of linguistic features (ablation) | InterpretARA · Weebit, Cambridge, CMT | accuracy: 91.75; 74.44; 37.36 | Mixed |
| Add layer normalization to the linearly transformed linguistic features (ablation) | InterpretARA · Weebit, Cambridge, CMT | accuracy: 92.70; 84.44; 43.59 | Mixed |
Surprising findings
- Direct concatenation of raw linguistic feature vectors (way 1) and linear-transformed features (way 2) make the hybrid model worse than the deep-only baseline on Weebit and Cambridge.
Limitations
- Baselines Lee2021 and BERT-FP-LBL results are taken from the literature because their code was unavailable, so comparisons are not fully controlled.
- Cambridge corpus is small (300 texts, 60 per class after downsampling), making results noisy.
- Fusion scheme between deep and linguistic representations is simple orthogonal projection; the authors note more effective fusion remains future work.
Experimental setup
Datasets: Weebit Cambridge Newsela CLEAR CMT
Benchmarks: Weebit Cambridge Newsela CLEAR CMT
Metrics: accuracy precision macro F1 QWK
Baselines: BERT HAN BigBird Lee2021 DTRA BERT-FP-LBL
Builds on: BERT supervised contrastive learning HAN DTRA BERT-FP-LBL
Closest papers by meaning
- Self-Supervised Collaborative Information Bottleneck for Text Readability Assessment (AAAI 2025 · similarity 0.89)
- Multi-Attribute Multi-Grained Adaptation of Pre-Trained Language Models for Text Understanding from Bayesian Perspective (AAAI 2025 · similarity 0.82)
- Analyze Feature Flow to Enhance Interpretation and Steering in Language Models (ICML 2025 · similarity 0.82)
- Language Model Meets Prototypes: Towards Interpretable Text Classification Models through Prototypical Networks (AAAI 2025 · similarity 0.82)
- A General Framework for Producing Interpretable Semantic Text Embeddings (ICLR 2025 · similarity 0.82)
- Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability (ICLR 2025 · 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.