Knowledge graph construction

Pick a set of documents (or upload your own), say which kinds of things and relationships you care about, and build. MaskLM reads each document, finds them, and merges duplicates across documents. Drag nodes, scroll to zoom, and click anything to see the sentence it came from.

1 · Documents: try a demo set or upload your own

Drop .txt, .md or .pdf files here, or · up to 20 files · text is extracted in your browser

2 · Node types and their attributes

Attributes are filled by the model from what each document states (unknown otherwise). "name" becomes the node's label and is used to match nodes across documents. Write "attr: description" to say exactly what you mean. Experimental with this model: it was trained on organisation fields only, so values are checked (confidence, must sit next to the node, a name must be one of its mentions) and many are dropped.

3 · Relationships (source → relation → target)

For each relationship the model is asked for only the entities that take part in it, typed by role ("person that works for an organisation", "organisation that a person works for"). Those are matched by mention overlap to the nodes from step 2, and joined when they share a sentence.

4 · Build

Ctrl+Enter