christian-bick/edugraph-embedding

EduGraph Embed

This model generates embeddings for labels from the EduGraph Ontology.

When combined with an EduGraph Classification Model, we can determine similarity between any type of learning content covered by the EduGraph ontology. For example, in tandem, the two models can determine whether some content of a math learning app trains the exact same set of skills tested in a paper quiz, by providing nothing else than a screenshot and a photo.

How it works

The model determines similarity based on the structure of the EduGraph Ontology. It respects various types of entity relationships to determine similarity, most importantly, parent-child and sibling relationships within the graph in addition to the semantic similarity of their definitions.

For example, the model will reliably place labels like IntegerAddition and FractionAddition closer together than, say, ShapeIdentification.

To accomplish this, the model generates knowledge graph embeddings that map the ontology structure into a high-dimensional vector space using a Relational Graph Convolutional Network (R-GCN).

Limitations

This model is centered around the EduGraph ontology. The embedding model was trained on the entities and relationships in this ontology. Consequently, it can only embed labels that are defined as entities within this ontology.

Risks

Important: Currently this model is in a research status and has not been evaluated under real-world conditions.

• ONLY use this model for research, experimentation and evaluation
• Do NOT use in a classroom environment
• Do NOT use for automations that might impact children

Using the Model

Preparation

1. Download the following files:

• embed_entities_biased.onnx
• embed_entities.pt

2. Install the following dependencies:

• torch
• numpy
• onnxruntime

Reference Example

See entity_embeddings_infer.py for reference usage.

License

This project is licensed under the GNU Affero General Public License. See the LICENSE file for details.

If these license terms are not working for you, then get in touch, and we can discuss your options.