A PyTorch implementation of a Tree2Tree neural network that translates natural language descriptions into source code, built for my MS thesis "Generation of code from text description with syntactic parsing and Tree2Tree model".
The idea: parse the input sentence into a syntax tree (CCG, PCFG, or dependency), encode it with a recursive tree-LSTM, and decode the output as a sequence of grammar rules that assemble a target-language AST — rather than generating code token-by-token. It's heavily inspired by Yin & Neubig's "A Syntactic Neural Model for General-Purpose Code Generation" and follows a similar structure to their reference implementation, NL2code.
Trained and evaluated on two NL-to-code datasets:
- HearthStone (hs) — card text → Python card implementations
- Django — Django docstrings/comments → Python code
| CCG | Dependency | PCFG |
|---|---|---|
 |
 |
 |
sh fetch_and_preprocess.sh # downloads + preprocesses the datasets
python experiment_hs.py # train/evaluate on HearthStone
python experiment_django.py # train/evaluate on DjangoModel and training options (encoder type, syntax formalism, embedding sizes, copy mechanism, etc.) live in config.py.
This was a research/thesis project, not production code — expect rough edges. Kept around as a record of a fun grad-school rabbit hole into grammar-based code generation.