A Grammar-Based Structural CNN Decoder for Code Generation

Authors: Zeyu Sun, Qihao Zhu, Lili Mou, Yingfei Xiong, Ge Li, Lu Zhang7055-7062

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on the Hearth Stone benchmark dataset show that our CNN code generator significantly outperforms the previous state-of-the-art method by 5 percentage points; additional experiments on several semantic parsing tasks demonstrate the robustness of our model. We also conduct in-depth ablation test to better understand each component of our model.
Researcher Affiliation Collaboration Key Laboratory of High Confidence Software Technologies (Peking University), Mo E; Software Institute, Peking University, 100871, P. R. China {szy , zhuqh, xiongyf, lige, zhanglucs}@pku.edu.cn Adept Mind Research, Toronto, ON, Canada doublepower.mou@gmail.com
Pseudocode No The paper describes the model architecture and training process in detail but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/zysszy/Grammar CNN
Open Datasets Yes We conducted experiments on an established benchmark dataset, Hearth Stone, for python code generation (Ling et al. 2016). Our first (and main) experiment is based on an established benchmark dataset, Hearth Stone (Ling et al. 2016, HS).
Dataset Splits Yes We follow the train-dev-test split as in Ling et al. (2016). We used the standard train-dev-test split for the datasets Zettlemoyer and Collins (2005).
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments. It only discusses the neural network architecture and training settings.
Software Dependencies No The paper mentions using the 'Adam optimizer (Kingma and Ba 2015)' but does not specify versions for programming languages, libraries, or other software dependencies.
Experiment Setup Yes For the neural network, we set the number of CNN layers L to 21... The layers of difference CNN modules are set to the same dimension, chosen by validation from {128, 192, 256} for each predictor network. We applied dropout (drop rate= 0.5) and ℓ2 penalty to regularize the fully connected layers. The network is trained by the Adam optimizer (Kingma and Ba 2015) with default hyperparameters. We use beam search to approximate the global inference, and the beam size is 5 in our experiments. The number of layers L is 7 in this experiment.