Research map / Architectures and efficiency / RNNs and normalization
Dynamic Compositionality in Recursive Neural Networks with Structure-Aware Tag Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add structure-aware tag representations from a tag-level tree-LSTM to control word-level composition gates (ablation) | SATA Tree-LSTM · SST-2 | accuracy | Helped |
| Add naive tag embeddings directly at each tree node instead of structure-aware tag representations (ablation) | SATA Tree-LSTM · SST-2 | accuracy | No clear effect |
| Add a leaf-LSTM to contextualize word inputs before tree composition instead of a fully connected tanh layer (ablation) | SATA Tree-LSTM · SST-2 | accuracy | Helped |
| Use a unidirectional rather than bidirectional leaf-LSTM (ablation) | SATA Tree-LSTM · SST-2 | accuracy | Mixed |
| Use SATA Tree-LSTM rather than the strongest reported tree-structured model | sentence classification models · SST-2 | accuracy: 90.7 -> 91.3 | Helped |
| Use SATA Tree-LSTM rather than the strongest reported tree-structured model | sentence classification models · SST-5 | accuracy: 53.7 -> 54.4 | Helped |
| Use SATA Tree-LSTM rather than the strongest reported tree-structured model | sentence classification models · MR | accuracy: 83.6 -> 83.8 | Helped |
| Use SATA Tree-LSTM rather than the strongest reported tree-structured model | sentence classification models · SUBJ | accuracy: 95.9 -> 95.4 | Hurt |
| Use SATA Tree-LSTM rather than the strongest reported tree-structured model | sentence classification models · TREC | accuracy: 96.1 -> 96.2 | Helped |
| Use SATA Tree-LSTM rather than Gumbel Tree-LSTM | sentence encoders · SNLI | accuracy: 85.6 -> 85.9 | Helped |
Closest papers by meaning
- Modeling Hierarchical Structures with Continuous Recursive Neural Networks (ICML 2021 · similarity 0.89)
- Learning to Compose Task-Specific Tree Structures (AAAI 2018 · similarity 0.88)
- Characterizing intrinsic compositionality in transformers with Tree Projections (ICLR 2023 · similarity 0.85)
- RNNs implicitly implement tensor-product representations (ICLR 2019 · similarity 0.85)
- Augmenting Transformers with Recursively Composed Multi-grained Representations (ICLR 2024 · similarity 0.84)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.83)
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.