What it does

This is the original code for Kim’s 2014 EMNLP paper, which showed that 1-D convolutions over word embeddings could classify sentences surprisingly well. It runs three model variants—random embeddings, static word2vec, and fine-tuned word2vec—on the Pang & Lee movie review dataset. The README is refreshingly honest about what was and wasn’t tested.

The interesting bit

The author openly admits the paper’s limitations: no GPU access during original experiments, missing ablation studies, and “premature” conclusions about regularization. That’s unusual candor in ML research code. The repo also serves as a historical marker—Theano 0.7 and Python 2.7 place it firmly in deep learning’s Jurassic period.

Key highlights

• Reproduces CNN-rand, CNN-static, and CNN-nonstatic from the paper
• Expect >81% CV accuracy with CNN-nonstatic (different folds than paper, though)
• GPU gives 10–20× speedup over CPU (219s → 16s per epoch)
• Links to TensorFlow and Torch reimplementations by others
• Author cites follow-up work (Ye Zhang 2015) that properly explored hyperparameters

Caveats

• Requires Python 2.7 and Theano 0.7—both long dead
• Fold assignments differ from the paper; direct comparison is slightly off
• No modern framework support; this is strictly for historical reference

Verdict

Worth a look if you’re tracing the evolution of NLP architectures or teaching the history of the field. Skip it if you need working, modern code—use one of the linked reimplementations instead.