2 Hits in 6.5 sec

MMF3: Neural Code Summarization Based on Multi-Modal Fine-Grained Feature Fusion [article]

Zheng Ma, Yuexiu Gao, Lei Lyu, Chen Lyu
2022 arXiv   pre-print
Aims: This study intends to improve the model's prediction performance for high-quality code summarization by accurately aligning and fully fusing semantic and syntactic structure information of source  ...  They generally represent different modalities of a piece of code, such as an Abstract Syntax Tree (AST) and a token sequence, as two embeddings and then fuse the two ones at the AST/code levels.  ...  The reason is that both MMTrans and AST-Trans only explore the structural information of the AST modality of the source code and do not consider the information of the Token modality of the source code  ... 
arXiv:2209.08978v1 fatcat:l4uedr7kwncwvct7st75ypchcy

Deep Learning Meets Software Engineering: A Survey on Pre-Trained Models of Source Code [article]

Changan Niu and Chuanyi Li and Bin Luo and Vincent Ng
2022 arXiv   pre-print
In particular, the development and use of pre-trained models of source code has enabled state-of-the-art results to be achieved on a wide variety of SE tasks.  ...  [2022]) 48.2 (AST-Trans [2022]) 1.6 M1 36.1 (SPT-Code [2022]) 34.7 (AST-Trans [2022]) 2.1 MN A2 60.1 (TreeBERT [2021]) E1 39.0 (TreeBERT [2021]) 57.5 (GNN+GRU [2020]) 34.4 (GNN+GRU [2020]) 6.1 7.0 CG  ...  In addition, code has syntactic structures (as defined in Abstract Syntax Trees (ASTs)) and semantic structures (as defined in Control Flow Graphs (CFGs)).  ... 
arXiv:2205.11739v1 fatcat:htgchuepzzgnnhvsehj5k3oype