Research map / Probabilistic and scientific modeling / Variational and generative sampling
Hierarchical Density Order Embeddings
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 |
|---|---|---|---|
| Use density-order training with thresholded KL divergence and graph-based negatives | Gaussian density embeddings · WordNet hypernym test set | accuracy: 90.6 -> 92.3 | Helped |
| Use asymmetric KL rather than symmetric ELK for training and scoring (ablation) | DOE · HyperLex nouns | Spearman's correlation: 0.455 -> 0.590 | Helped |
| Train with ELK but calculate scores using KL (ablation) | DOE · HyperLex nouns | Spearman's correlation: 0.455 -> 0.532 | Helped |
| Add S2 direction-reversal negative samples to S1 (ablation) | DOE with KL · HyperLex nouns | Spearman's correlation: 0.527 -> 0.567 | Helped |
| Add S2 and S4 graph-based negatives to S1 (ablation) | DOE with KL · HyperLex nouns | Spearman's correlation: 0.527 -> 0.590 | Helped |
| Use S3 rather than S4 with S1 and S2 (ablation) | DOE with KL · HyperLex nouns | Spearman's correlation: 0.590 -> 0.570 | Hurt |
| Replace the proposed loss with the original Gaussian-embedding loss (ablation) | DOE with S1 + S2 + S4 · WordNet hypernym test set | accuracy: 0.911 -> 0.796 | Hurt |
| Replace the proposed loss with the original Gaussian-embedding loss (ablation) | DOE with S1 + S2 + S4 · HyperLex nouns | Spearman's correlation: 0.590 -> 0.489 | Hurt |
| Reverse the KL divergence direction (ablation) | DOE · WordNet hypernym test set | accuracy: 92.3 -> 83.2 | Hurt |
| Reduce embedding dimension from 50 to 5 (ablation) | DOE with S1 + S2 + S4 · HyperLex nouns | Spearman's correlation: 0.483 -> 0.590 | Hurt |
| Use 2 dimensions instead of the default 50 (ablation) | DOE with KL · WordNet hypernym test set | accuracy: 92.3 -> 89.2 | Hurt |
| Set the divergence threshold to zero rather than using a tuned non-zero threshold (ablation) | DOE with KL · HyperLex nouns | Spearman's correlation | Hurt |
Closest papers by meaning
- Smoothing the Geometry of Probabilistic Box Embeddings (ICLR 2019 · similarity 0.88)
- Exponential Family Embeddings (NeurIPS 2016 · similarity 0.87)
- A Generative Model of Words and Relationships from Multiple Sources (AAAI 2016 · similarity 0.86)
- Encoding word order in complex embeddings (ICLR 2020 · similarity 0.86)
- By Tying Embeddings You Are Assuming the Distributional Hypothesis (ICML 2024 · similarity 0.85)
- Language Models as Hierarchy Encoders (NeurIPS 2024 · similarity 0.84)
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.