Multi-modal Code Summarization Fusing Local API Dependency Graph and AST

Abstract

Automatically obtaining descriptions of the functions of code snippets in natural language, i.e., code summarization, is an important issue in software engineering. In recent years, abstract syntax tree (AST)-based code summarization models for modelling syntactic structures have continued to emerge. In addition to syntactics, the semantics of the code are gradually gaining attention. In this paper, we propose a code summarization approach that incorporates local-ADG and AST, called GTsum. In particular, we introduce a novel local-ADG-based approach that can effectively filter out irrelevant semantics in the modelling of the semantic structure of code. GTsum learns semantics and syntactics in the local-ADG and AST through graph convolutional networks (GCNs) and then fuses them using Transformer. We evaluate our model on two Java language datasets with several metrics. The results demonstrate that our model achieves state-of-the-art performance compared to the existing models.