Code Generation Method based on Structured Tree Input and AST Decoder Attention Augmentation

Automatic code generation based on natural language input is important to research in the field of software engineering. In the past, it was mostly a seq2seq structure and used the RNN model. Input and output are regarded as simple sequences, and syntactic structure information in source information is often ignored. This paper proposes a code generation method Tx(Tree-Tree). It uses structured trees to replace simple word sequences so that the model can better learn the syntactic and semantic information in the source information. Therefore, it can alleviate the long dependency problem caused by too long source information. At the same time, the enhanced attention mechanism is adopted in the decoder to distinguish the influence of different historical actions on the current predicted action. The model is validated on three datasets: DJANGO, CONALA, and ATIS. Compared with some typical models, Tx(Tree-Tree) improves both accuracy and BLEU.