Repository files navigation

CodeQA is a free-form question answering dataset for the purpose of source code comprehension: given a code snippet and a question, a textual answer is required to be generated.

To obtain natural and faithful questions and answers, we implement syntactic rules and semantic analysis to transform code comments into question-answer pairs.

We hope this new dataset can serve as a useful research benchmark for source code comprehension.

You can find more details, analyses, and baseline results in our Findings of EMNLP 2021 paper "CodeQA: A Question Answering Dataset for Source Code Comprehension".

The dataset (ver. 1.0) can be downloaded fromGoogle Drive.

It contains a Java dataset with 119,778 question-answer pairs and a Python dataset with 70,085 question-answer pairs.

A few examples of CodeQA data format are shown indata_sample. Each one contains a question, an answer and a code snippet. For the code snippet, we provide an original version (.code.original) as well as a processed version (.code). Details of the processing are available in the following code summarization datasets.

We follow the same evaluation method of automatic metrics (BLEU, ROUGE-L, METEOR) as inAhmad et al. (2020).

Source code can be foundhere.

cd codeBERT

• pip install torch
• pip install transformers

You can download data fromGoogle Drive. Unzip it and move it to./data.

We fine-tune the model on 3*1080Ti GPUs.

Please run the following scripts:

bash java_script.sh [gpu-id] [model-name]

bash python_script.sh [gpu-id] [model-name]

After the training process, several best checkpoints are stored in a folder named after your model name, for example,./output/[model-name]/checkpoint-best-bleu/pytorch_model.bin. You can run the following scripts to get the results on test set:

bash java_script_test.sh [gpu-id] [model-name]

bash python_script_test.sh [gpu-id] [model-name]

Java and Python pre-trained models (20 epochs) are availablehere.

Our CodeQA dataset is based on two code summarization datasets,code-docstring-corpus andTL-CodeSum.

We are thankful to the authors for making dataset and code available.

If you use our dataset, please cite us!

@inproceedings{liu2021codeqa,  title={CodeQA: A Question Answering Dataset for Source Code Comprehension},  author={Liu, Chenxiao and Wan, Xiaojun},  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2021},  pages={2618--2632},  year={2021}}