Code and datasets for our AAAI-2020 paper "Associating Natural Language Comment and Source Code Entities"

If you find this work useful, please consider citing our paper:

@inproceedings{panthaplackel2020associating,
  author={Sheena Panthaplackel and Milos Gligoric and Raymond J. Mooney and Junyi Jessy Li},
  title={Associating Natural Language Comment and Source Code Entities},
  booktitle={AAAI},
  year={2020},
}

Data is available in model_data/. It can be parsed using the load_data method in models/model_utils.py.

Download embeddings.tar.gz continaining pretrained embeddings from here. Unzip the file in the root directory:

tar zxvf embeddings.tar.gz

A directory with the name embeddings should appear, in the root directory, with 3 json files.

You will need to create a checkpoints directory under the root directory.

Run models from within the models directory. Commands to train models are structured as below:

python run_model.py -model [MODEL_TYPE] -dropout [DROPOUT_KEEP_PROBABILITY] -lr [LEARNING_RATE] -decay [DECAY_RATE] -decay_steps [NUM_DECAY_STEPS] -num_layers [NUM_LAYERS] -layer_units [LAYER_DIMENSIONS] -model_name [MOEL_NAME] -delete_size [NUM_EXAMPLES_FROM_DELETIONS]

Insert one of the following model types in place of MODEL_TYPE:

• feedforward
• more_data_feedforward
• crf
• more_data_crf
• subtoken_matching_baseline
• return_line_baseline
• random_baseline
• majority_class_random_baseline

Matching model types with those in the paper:

Learned models:

• feedforward = binary classifier
• more_data_feedforward = binary classifier w/ data from deletions set
• crf = CRF for joint classification
• more_data_crf = CRF for joint classication w/ data from deletions set

Baselines:

• subtoken_matching_baseline = subtoken matching
• return_line_baseline = presence in return line
• random_baseline = random
• majority_class_random_baseline = weighted random

Sample commands can be found in models/run.sh.

Please email [email protected] for any questions.