Tutorials¶
Five notebooks that take you from "what is a text graph?" to training on your own corpus. Each one answers the question the previous one leaves open, so they read best in order.
flowchart LR
A["1 · Text as a graph<br/><small>what the graph is</small>"] --> B["2 · Your first classifier<br/><small>raw text → accuracy</small>"]
B --> C["3 · Choosing a graph builder<br/><small>which edges?</small>"]
C --> D["4 · Refining the graph<br/><small>fewer edges, same accuracy?</small>"]
D --> E["5 · Bring your own corpus<br/><small>your data, any language</small>"]| # | Tutorial | You will learn | Runtime |
|---|---|---|---|
| 1 | Text as a graph | What nodes and edges Text4GCN builds, and why an unlabelled document can be classified from its place in the graph | seconds, CPU |
| 2 | Your first classifier | The full pipeline on the 20 Newsgroups benchmark: download, TextPipeline, Frequency.tfidf(), GNN.fit() |
~2.5 min GPU |
| 3 | Choosing a graph builder | TF-IDF + PMI vs. Word2Vec RWF vs. cosine similarity: edges, build time and accuracy side by side | ~4.5 min GPU |
| 4 | Refining the graph | Measuring and pruning a built graph with RefinementConfig, and what it does to accuracy |
~2.5 min GPU |
| 5 | Bring your own corpus | From a DataFrame to the input files, non-English text, encodings and memory sizing | seconds, CPU |
The story in one paragraph¶
TextGCN turns a corpus into one graph of documents and words and lets a GCN pass labels along its edges (1). On a real benchmark that graph classifies well, but it has millions of edges (2). The Related Word Filter of Pereira & de Castro (DCAI 2022) connects only related words and keeps comparable accuracy with an order of magnitude fewer edges, because the document–word edges carry most of the signal (3). Refinement prunes an existing graph further, without rebuilding it (4). Everything carries over to your own data and language (5).
Running the notebooks¶
Every notebook has an Open in Colab badge and installs text4gcn in its
first cell. Use a GPU runtime for tutorials 2–4 (Runtime → Change runtime
type → T4 GPU). To run them locally: