comparison of encoding methods in code generation neural networks

comparison of encoding methods in code generation neural networks

program synthesis is the study of generating program codes based on user’s intent. neural program synthesis uses neural networks in order to generate program codes at output. the most widely-used architecture in such problems is encoder-decoder. in this paper, we use a predefined code generation neural network and apply different methods of encoding embedded natural language data to the neural network.

the encoding methods include various types of recurrent neural networks and convolutional neural networks. the architectures in-use consist of series and parallel connection of layers in the encoder. we then train and test the code generation neural network and evaluate the results comparing the methods with each other. the results show that using higher number of layers connected to each other in series in the encoder part of the code generation neural network reduces the accuracy gradually, while using parallel encoding architectures could result in boosting the performance. the results also indicate the advantage of applying recurrent neural network architectures to the encoder in such code generation tasks.