Knowledge Graphs
Turning text into connected facts.
Who owns the company that published this book? No single document says so. One says who founded Bright Books. Another says it bought a publisher. To answer, you have to connect the facts. That's what a knowledge graph is for.
- No single document has the answer
- You have to connect the facts
- A map of how things relate
Extracting facts
We use the dependency parsing that we learned earlier. For each sentence, find the verb, its subject, and its object. Each one becomes an edge in a graph, between two entities. Run it. Plain sentences have become a set of connected facts. For messier text, you'd use a language model with a schema to extract these, which works far better on real documents.
import spacy
from collections import defaultdict
nlp = spacy.load("en_core_web_sm")
text = ("Bright Books acquired Northern Press. Northern Press published Dune. "
"Sam Lee founded Bright Books. Anna Chen leads Northern Press.")
graph = defaultdict(list)
for sent in nlp(text).sents:
root = sent.root
subj = [t for t in root.children if t.dep_ == "nsubj"]
obj = [t for t in root.children if t.dep_ == "dobj"]
if subj and obj:
a = " ".join(w.text for w in subj[0].subtree)
b = " ".join(w.text for w in obj[0].subtree)
graph[a].append((root.lemma_, b))
for head, facts in graph.items():
print(head, "->", facts)Bright Books -> [('acquire', 'Northern Press')]
Northern Press -> [('publish', 'Dune')]
Sam Lee -> [('found', 'Bright Books')]
Anna Chen -> [('lead', 'Northern Press')]Answering by hopping
Now we can answer questions that need several steps. How is Sam Lee connected to Dune? We walk through the graph, from one fact to the next. Run it. Three hops, and a clear answer. Sam founded Bright Books, which bought the press, which published Dune. Ordinary RAG would struggle with this, since no single chunk contains the link. This idea, called graph RAG, is used when questions connect many documents.
graph = {
"Sam Lee": [("found", "Bright Books")],
"Bright Books": [("acquire", "Northern Press")],
"Northern Press": [("publish", "Dune")],
}
def path(start, goal, seen=()):
for rel, next_node in graph.get(start, []):
if next_node == goal:
return [(start, rel, next_node)]
if next_node not in seen:
rest = path(next_node, goal, seen + (start,))
if rest:
return [(start, rel, next_node)] + rest
return []
for step in path("Sam Lee", "Dune"):
print(step)('Sam Lee', 'found', 'Bright Books')
('Bright Books', 'acquire', 'Northern Press')
('Northern Press', 'publish', 'Dune')Summary
Knowledge graphs connect facts across documents, and they answer questions that search alone can't. Next, what about trends? What are customers saying this month, compared to last?
Key takeaways- Subject, verb, object become edges
- Hop across facts
- Graph RAG links documents