Sparse Lexical and Expansion Model using BERT for learned sparse retrieval, combining the interpretability of lexical search with the semantic power of neural models for enhanced keyword search.
SPLADE (Sparse Lexical and Expansion Model) uses pretrained language models like BERT to identify connections between words/sub-words and uses that knowledge to enhance sparse vector embeddings.
Unlike dense embeddings, SPLADE outputs vectors aligned with a vocabulary (similar to bag-of-words), but with learned weights reflecting semantic and contextual information.
For the keyword "study":
| Feature | BM25 | SPLADE |
|---|---|---|
| Non-zero values | Single term only | Multiple related terms |
| Semantic understanding | None | Yes (via BERT) |
| Term expansion | No | Yes |
| Context awareness | No | Yes |
| Performance | Baseline | Superior recall |
Open-source model, free to use.
Loading more......