---
language:
- en
widget:
- text: "The Italian Space Agency’s Light Italian CubeSat for Imaging of Asteroids, or LICIACube, will fly by Dimorphos to capture images and video of the impact plume as it sprays up off the asteroid and maybe even spy the crater it could leave behind."
tags:
- seq2seq
- relation-extraction
- triple-generation
- entity-linking
- entity-type-linking
- relation-linking
model-index:
- name: knowgl
  results:
  - task:
      name: Relation Extraction
      type: Relation-Extraction
    dataset:
      name: "Babelscape/rebel-dataset"
      type: REBEL
    metrics:
       - name: RE+ Macro F1 
         type: re+ macro f1
         value: 70.74
license: cc-by-nc-sa-4.0
---

# KnowGL: Knowledge Generation and Linking from Text

The `knowgl-large` model is trained by combining Wikidata with an extended version of the training data [REBEL](https://huggingface.co/datasets/Babelscape/rebel-dataset) dataset. Given a sentence, it generates triple(s) in the following format -
```
[(subject mentions # subject label # subject type) | relation label | (object mentions # object label # object type)]
```

It achieves state-of-the-art results for relation extraction on the test dataset of REBEL. See results comparison with [GenIE (Josifoski et al (NAACL 2022))](https://aclanthology.org/2022.naacl-main.342.pdf) in [Mihindukulasooriya et al (ISWC 2022)](https://arxiv.org/pdf/2207.05188.pdf).

The generated labels (for subject, relation and object) and types (subject and object) can be directly mapped to Wikidata IDs associated with them.