Knowledge graph construction from unstructured data

text2kgraph is a small Python package that sketches a pipeline from unstructured text to a knowledge graph. It finds entities in text with spaCy, builds a directed graph with NetworkX, and exports the graph in several standard formats. He wrote it in early 2025 to accompany written work on knowledge-graph construction, just before his ontology-engineering research at JAIST.

Pipeline

1. Preprocessing cleans the input text.
2. Entity extraction runs spaCy's English model (en_core_web_sm) and keeps each named entity with its type, such as PERSON, DATE or WORK_OF_ART.
3. Relation extraction is the step that turns entities into facts. In this version it is a placeholder: it returns one fixed triple per sentence instead of real relations.
4. Graph building adds every entity as a node and every relation as a labelled, directed edge in a NetworkX DiGraph.
5. Export writes the graph as:
   • node-link JSON;
   • a tab-separated edge list (source → target → relation);
   • an adjacency matrix, with a separate file of node names;
   • GraphML, for tools such as Gephi and yEd.

Example

text = "Albert Einstein developed the theory of relativity in 1915."
entities = extract_entities(clean_text(text))
relationships = extract_relationships(text)
graph = build_graph(entities, relationships)
save_as_graphml(graph, "graph.graphml")

Status

The repository is an early scaffold rather than a finished extractor. The entity extraction, graph building and every exporter work, but relation extraction still needs to be implemented, for example with dependency parsing or a trained relation model. The later work went in that direction:

• At JAIST, his semantic extraction pipeline extracted structured facts and checked them with SPARQL.
• His ontology-augmented health recommendation project used a hand-built OWL ontology instead of extracting one from text.

See also

• Ontology and knowledge graphs
• Machine learning and data
• Natural language processing