C3PO: A Lightweight Copying Mechanism for Translating Pseudocode to Code

Source Code for the C3PO paper published at AACL-IJCNLP 2022 (Student Research Workshop) [Paper]

Stages of the Pipline

1. Copy Phase
2. Generate Phase
3. Combine Phase

Getting Started

1. Install conda or miniconda
2. Create a conda environment

conda create -f c3po.yml 
conda activate c3po

3. Download the SPOC dataset as a zip and unzip into into the data/ folder

Data Preprocessing

Data preprocessing of the SPOC dataset to mask [CPY] tags using a decision tree is included under cpy_preprocess/.
The preprocessed datasets are released in the data/ folder.
There are 2 versions of both the train and eval datasets.

1. Normally masking tokens with [CPY] tokens. In data/CPY_dataset.pkl (for train set) and data/CPY_dataset_eval.pkl (for eval set)
2. Numbering the [CPY] tokens as [CPY_1] ... [CPY_n]. In data/CPY_dataset_numbered.pkl (for train set) and data/CPY_dataset_numbered_eval.pkl (for eval set)

Models

Experiments were carried out with the 2 following architectures, which is included in the models/ folder.

1. Vanilla Seq2Seq (models/vanilla_seq2seq.py)
2. Attention Seq2Seq (models/attention_seq2seq.py)

Training Seq2Seq models

Scripts in train_scripts/seq2seq_training.py. Can train both Usage:

python seq2seq_training.py
--attention # Boolean flag if attention seq2seq model to be trained (Default Vanilla Seq2Seq)
--non-copy # Boolean flag if non-copy version to be trained (no masking of tokens) (Default copy version)

The hyperparameters used have been set by default for each model

Inference and Results

Inference scripts are included in inference/scripts/. The predictions are released as .pkl files under inference/predictions/ for the following versions

1. Attention Seq2Seq with C3PO (CPY masking)
2. Attention Seq2Seq w/o C3PO (No CPY masking)
3. Vanilla Seq2Seq with C3PO (CPY masking)
4. Vanilla Seq2Seq w/o C3PO (No CPY masking)

The results (BLEU scores) are computed using these prediction files in inference/results.ipynb

Citation

If you used this work in your research, please cite:

@inproceedings{veerendranath2022c3po,
    title={C3PO: A Lightweight Copying Mechanism for Translating Pseudocode to Code},
    author={Veerendranath, Vishruth and Masti, Vibha and Anagani, Prajwal and Hr, Mamatha},
    booktitle={Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing: Student Research Workshop},
    pages={47--53},
    year={2022}
}