lgalke/text-clf-baselines

Implements comparative benchmarks across Bag-of-Words MLPs, graph neural networks (TextGCN, HeteGCN), and Transformer models (BERT, DistilBERT) on standard text classification datasets. The wide MLP approach uses GloVe embeddings with a simple dense architecture, demonstrating competitive or superior performance to graph-based methods while maintaining significantly lower computational overhead—avoiding the O(N²) graph construction and O(L²) attention computations. Includes modular implementations of tokenization, data loading for five benchmark datasets (20ng, R8, R52, OHSUMED, MR), and reproducible experiment scripts from the ACL 2022 paper.

No commits in the last 6 months.