Research map / Label-efficient and robust learning / Recognition and long-tail
All-but-the-Top: Simple and Effective Postprocessing for Word Representations
Jiaqi Mu, Pramod Viswanath
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Center embeddings and remove the top principal components | WORD2VEC and GLOVE · Word similarity, seven datasets | Spearman correlation: average improvement of 2.3% | Improved |
| Center embeddings and remove the top principal components | WORD2VEC and GLOVE · Concept categorization, three datasets | purity: average improvement of 2.5% | Improved |
| Center embeddings and remove the top principal components | WORD2VEC · Word analogy, all | accuracy: 72.93 -> 73.44 | Improved |
| Center embeddings and remove the top principal components | GLOVE · Semantic textual similarity, all | Pearson correlation: 49.19 -> 56.76 | Improved |
| Postprocess embeddings before classification | Four neural-network architectures with WORD2VEC or GLOVE · Text classification, five datasets | classification performance: 34 out of 40; 2.32% on average | Mixed |
| Center vectors and remove top principal components | TSCCA · Word similarity, German and Spanish datasets | Spearman correlation: 60.33 -> 60.37; 61.75 -> 64.39; 44.91 -> 46.59; 30.37 -> 32.92 | Improved |
| Set D according to the d/100 rule | Skip-gram embeddings across dimensions 300 to 1000 · Word similarity | Spearman correlation: 2.27% | Improved |
| Set D according to the d/100 rule | Skip-gram embeddings across dimensions 300 to 1000 · Word analogy | accuracy: 0.01 | No clear effect |
| Set D according to the d/100 rule | Skip-gram embeddings across dimensions 300 to 1000 · Concept categorization | purity: 3.37% | Mixed |
| Set D according to the d/100 rule | Skip-gram embeddings across dimensions 300 to 1000 · Semantic textual similarity | Pearson correlation: 1.92% | Improved |
| Remove the top principal components | RAND-WALK and TSCCA · Word similarity | Spearman correlation: 2.23% average improvement for RAND-WALK and TSCCA | Mixed |
| Remove the top principal components | RAND-WALK and TSCCA · Semantic textual similarity | Pearson correlation: 0.61% average improvement for RAND-WALK and TSCCA | Mixed |
Closest papers by meaning
- Learning Semantic Representations for Novel Words: Leveraging Both Form and Context (AAAI 2019 · similarity 0.88)
- Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models (NeurIPS 2025 · similarity 0.86)
- Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model (AAAI 2024 · similarity 0.86)
- On Affine Homotopy between Language Encoders (NeurIPS 2024 · similarity 0.86)
- The Foundations of Tokenization: Statistical and Computational Concerns (ICLR 2025 · similarity 0.86)
- On the Origins of Linear Representations in Large Language Models (ICML 2024 · similarity 0.85)
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.
