ICSE-2019 A Neural Model for Generating Natural Language Summaries of Program Subroutines
Paper: A Neural Model for Generating Natural Language Summaries of Program Subroutines
ast-attentiongru
1. Introduction
- Code is treated as both a textual token representation and an AST representation; to distinguish these two views, both are fed as input jointly.
- Two experimental settings are designed: (i) the standard setting, code tokens plus AST; (ii) the challenging setting, AST only, without code tokens.
2. Model
The authors propose ast-attendgru, an attention-based encoder–decoder architecture. For code representation, they use the token sequence and the structural AST; each branch is encoded separately. The architecture is shown below:

Dataset: a self-collected Java method corpus—51 million Java methods from 50,000 projects—yielding roughly 2.1M methods after preprocessing.
- baselines:
- attendgru
- SBT (ICPC 18)
- CODE-NN (Iyer 19)
<HR align=left color=#987cb9 SIZE=1>