SHACKLE: SHape-based pAtterns for Constraining KnowLedge graph Embeddings
About
Neural network-based AI methods have become pervasive in a large range of domains, including some with major societal impacts such as health or justice where trust is key. Those methods are proficient but still error-prone (e.g. hallucinations) and not explainable, hindering their operationalization. On the other hand, knowledge graphs (KGs) are recognized as interpretable and trustworthy symbolic representations of knowledge. They are often equipped with deductive (i.e. ontologies) and validation (i.e. shapes) schemata to guarantee soundness and completeness, and can also be used in neural methods with embedding models. As a stepping stone toward a neuro-symbolic integration to reduce errors, increase explainability, and thus enforce trust, SHACKLE will target the combination of sub-symbolic methods, especially Knowledge Graph Embedding Models (KGEMs), and validation schemata, a currently unexplored area.Scientific tasks
T1: Datasets involving KGs and shapes
We propose benchmark datasets at the intersection of KG link prediction and SHACL shapes. These datasets are available below:
T2: Naive combination between KGEMs and shapes
We propose naive approaches to inject shapes in KGEMs to improve the quality of their inferences.
T3: Develop a measure for shape validation from models
We propose a metric to measure the compliance of model inferences with constraints expressed in shapes.
T4: Advanced combination approaches between KGEMs and shapes
We extend results from T2 and T3 with advanced combination approaches.
T5: Extended collaboration
SHACKLE served as a stepping stone to an extended European collaboration. Stay tuned for more info!
Team
Pierre Monnin
Wimmics team
Université Côte d'Azur, Inria, CNRS, I3S, Sophia Antipolis, France
Visiting researcher
Mehwish Alam (host)
DIG team
Télécom Paris, Institut Polytechnique de Paris, LTCI, France
Host researcher
Stays at Télécom Paris
- February 17-20, 2025
- March 17-21, 2025
- March 24-26, 2025
- April 14-18, 2025
- July 9-11, 2025
Dissemination
- Pierre Monnin presentation at Telecom Paris on March 18, 2025: “Neurosymbolic approaches for the knowledge graph lifecycle”
Funding
The "SHACKLE: SHape-based pAtterns for Constraining KnowLedge graph Embeddings" project has received funding from the European Union, via the oc2-2024-TES-02 issued and implemented by the ENFIELD project, under the grant agreement No 101120657.
