notesum.ai
Published at November 26Isotropy Matters: Soft-ZCA Whitening of Embeddings for Semantic Code Search
cs.CL
Released Date: November 26, 2024
Authors: Andor Diera1, Lukas Galke2, Ansgar Scherp1
Aff.: 1Ulm University - Data Science and Big Data Analytics; 2University of Southern Denmark - Dept of Mathematics and Computer Science

| Embedding Size 256 | Embedding Size 768 | |||
|---|---|---|---|---|
| MRR | IsoScores | MRR | IsoScores | |
| Ruby | 0.705 | 0.350 / 0.296 | 0.463 | 0.102 / 0.102 |
| Javascript | 0.638 | 0.365 / 0.335 | 0.360 | 0.110 / 0.118 |
| Go | 0.757 | 0.234 / 0.196 | 0.413 | 0.042 / 0.044 |
| Java | 0.595 | 0.388 / 0.313 | 0.323 | 0.120 / 0.081 |
| Python | 0.721 | 0.394 / 0.356 | 0.387 | 0.122 / 0.139 |
| PHP | 0.537 | 0.400 / 0.262 | 0.266 | 0.107 / 0.065 |
| R | 0.045 | 0.139 / 0.118 | 0.026 | 0.026 / 0.012 |