A Code Generation Method for Syntax Tree Decoder

Abstract

The invention discloses a code generation method oriented to a syntax tree decoder, comprising the following steps: parsing data through an encoder to generate an AST, applying a sliding window on the AST structure to extract the AST structural features, and performing deep-level node analysis on the AST. Traverse to obtain the order relationship between child, parent and sibling nodes in the AST; add multi-layer connections to the AST, design a dense network block structure, and use the dense connection method to train a deep convolutional network to form a dense convolutional neural network; design transition Layer structure, add the designed transition layer to the dense convolutional neural network and perform convolution and pooling operations, optimize the dense convolutional neural network, output feature maps, and generate codes. The invention shortens the connection of each network layer in syntax tree decoding, increases the connection of each network layer, alleviates the gradient disappearance problem caused by network deepening, enhances feature propagation, promotes feature sharing, and effectively reduces the number of parameters.

Description

technical field [0001] The invention relates to the technical field of code generation, in particular to a code generation method oriented to a syntax tree decoder. Background technique [0002] With the development of neural network technology, it has become possible to automatically generate the source code of the executable programming language using description specifications written in natural language and other symbols. Among them, the implementation of the syntax tree decoder, the existing methods mainly rely on the recurrent neural network (RNN) encoder, then to the Seq2Seq model, and then to the CNN, the model can not only learn the long and short-term memory of long sequences, but also capture the description specification. more features. With the deepening of description specification semantics and the complexity of logic, which leads to the deepening of CNN, the problem of gradient disappearance and degradation occurs in the grammar-based structure CNN decoder d...

AI Technical Summary

Problems solved by technology

Although the application scheme can achieve a dynamic description of the target code, it cannot reduce the number of parameters and cannot alleviate the problem of gradient disappearance caused by network deepening.

Images