Research map / Architectures and efficiency / RNNs and normalization
Hash Embeddings for Efficient Word Representations
Dan Tito Svenstrup, Jonas Hansen, Ole Winther
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 |
|---|---|---|---|
| Replace standard embeddings with hash embeddings using 5 times fewer parameters | Shallow text classifier without a dictionary · AG’s news, DBPedia, Yelp Review Polarity, Yelp Review Full, Amazon Review Full, Amazon Review Polarity, Yahoo! Answers | test accuracy (%): AG 92.0 -> 92.4; Amazon full 58.3 -> 60.0; DBPedia 98.6 -> 98.5; Yahoo 72.3 -> 72.3; Yelp full 62.6 -> 63.8; Amazon pol | Mixed |
| Use hash embeddings instead of standard embeddings | Deep text classifier using a dictionary · AG’s news, DBPedia, Yelp Review Polarity, Yelp Review Full, Amazon Review Full, Amazon Review Polarity, Yahoo! Answers | test accuracy (%): AG 91.7 -> 91.5; Amazon full 58.5 -> 59.4; DBPedia 98.6 -> 98.7; Yahoo 65.8 -> 71.3; Yelp full 61.4 -> 62.6; Amazon pol | Mixed |
| Add a 10-model soft-voting ensemble of dictionary hash-embedding models | Dictionary hash-embedding classifier · AG’s news, DBPedia, Yelp Review Polarity, Yelp Review Full, Amazon Review Full, Amazon Review Polarity, Yahoo! Answers | test accuracy (%): AG 91.5 -> 92.0; Amazon full 59.4 -> 60.5; DBPedia 98.7 -> 98.8; Yahoo 71.3 -> 72.9; Yelp full 62.6 -> 62.9; Amazon pol | Mixed |
| Use a different hash function for importance-parameter indices rather than reusing the component-vector token-to-id function | Hash embeddings · Text classification (initial findings; dataset unspecified) | performance | Improved |
Closest papers by meaning
- Learning Semantic Representations for Novel Words: Leveraging Both Form and Context (AAAI 2019 · similarity 0.87)
- Probabilistic Hash Embeddings for Online Learning of Categorical Features (AAAI 2026 · similarity 0.87)
- All Word Embeddings from One Embedding (NeurIPS 2020 · similarity 0.87)
- On the Dimensionality of Word Embedding (NeurIPS 2018 · similarity 0.86)
- Local Density Estimation in High Dimensions (ICML 2018 · similarity 0.86)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (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.
