Kastor: Fine-tuned Small Language Models for Shape-based Active Relation Extraction
Célian Ringwald, Fabien Gandon, Catherine Faron, Franck Michel, Hanna Abi Akl · 2025
Read Kastor: Fine-tuned Small Language Models for Shape-based Active Relation Extraction by Célian Ringwald, Fabien Gandon, Catherine Faron, Franck Michel, Hanna Abi Akl (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-03466.txt.
Description
Celian Ringwald and co-authors present Kastor, a framework that trains small language models to extract facts actively under shape rules. Shape-guided extraction focuses a model on one graph pattern at a time, working well with little data. Kastor widens the check from a single shape to every property combination the shape allows.
Picking the best combination per training example lifts generalization and accuracy. An iterative loop then cleans noisy knowledge bases, refining stored facts round by round and growing fresh ones. Specialist fields with thin data, such as heritage or niche science, need exactly that kind of upkeep. Kastor aims to make small, cheap models trustworthy curators rather than one-shot extractors.
Catalog record
| CREATOR | Célian Ringwald, Fabien Gandon, Catherine Faron, Franck Michel, Hanna Abi Akl |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03466