text_gcn

Why we included this project

Researchers and practitioners who want to compare graph-based text classification against classic bag-of-words models will find the official Text GCN implementation from the AAAI 2019 paper here. Instead of treating each document as a bag of words, it builds a single heterogeneous graph where documents and words are nodes, edges capture word-document co-occurrence and word-word co-occurrence, and graph convolution runs over that structure. The repo includes scripts to clean a corpus, build the graph, and train the model, with the paper's datasets (20ng, R8, R52, ohsumed, mr) already wired in, so you can check your numbers against the published ones. A companion inductive version is referenced for cases where test documents must stay out of the training graph, though the code targets TensorFlow 1.x, so expect some adaptation for modern pipelines.