A simple example test implementation of the princeton graphMERT paper.
https://arxiv.org/abs/2510.09580
All rights w/ authors: GraphMERT: Efficient and Scalable Distillation of Reliable Knowledge Graphs from Unstructured Data Margarita Belova, Jiaxin Xiao, Shikhar Tuli and Niraj K. Jha from Department of Electrical and Computer Engineering, Princeton University ArXiv link arXiv:2510.09580 from 10 October 2025
Based on the notebooks use of Romeo + Juliet text for the graphMERT model training.
GraphMERT Node Embeddings (t-SNE View)
import matplotlib.pyplot as plt from sklearn.manifold import TSNE
tsne = TSNE(n_components=2, random_state=42) emb_2d = tsne.fit_transform(embeddings)
plt.figure(figsize=(8,8)) plt.scatter(emb_2d[:,0], emb_2d[:,1], c=labels, cmap='Spectral', alpha=0.7) for i, text in enumerate(corpus[:50]): # optional show partial labels plt.text(emb_2d[i,0]+0.01, emb_2d[i,1]+0.01, str(labels[i]), fontsize=6) plt.title("GraphMERT Node Embeddings (t-SNE View)") plt.show()
GraphMERT Semantic Graph Visualization
import networkx as nx
G = nx.Graph() for i, emb in enumerate(embeddings): G.add_node(i, text=corpus[i], cluster=int(labels[i]))
from scipy.spatial.distance import pdist, squareform dist = squareform(pdist(embeddings)) threshold = np.percentile(dist, 5) for i in range(len(embeddings)): for j in range(i+1, len(embeddings)): if dist[i,j] < threshold: G.add_edge(i, j)
plt.figure(figsize=(10,8)) pos = nx.spring_layout(G) nx.draw(G, pos, node_color=labels, cmap='Spectral', with_labels=False, node_size=40) plt.title("GraphMERT Semantic Graph Visualization") plt.show()
Query search on the graphs results
- This is what we want bacause the search of the graph is linear based on chained knowledge with nodes having data of them.
