ICSME-2021 Ensemble Models for Neural Source Code Summarization of Subroutines

· ICSME· · code-summarization

Paper: Ensemble Models for Neural Source Code Summarization of Subroutines

EnsMdl: Better results by ensembling diverse models

Abstract

  1. Models may share similar architectures yet contribute differently to prediction quality.
  2. The authors study the orthogonal properties of various code summarization methods through model ensembling.

Introduction

  1. Different information sources, once modeled with encoders, can provide orthogonal gains for prediction.

Background & Related Work

  1. Code summarization feature extraction

    1.1 Text-based: treat code as a sequence and feed it directly to the model

    1.2 Flat structure: flatten the AST and use it as input

    1.3 Context-based: API usage, project context, and similar signals

    1.4 GNN-based: construct a graph from the AST

  2. Ensemble models

    Ensembling has achieved strong results in machine translation, stock market forecasting, and other domains.

    Prior work shows that simple ensembles of non-AST and AST-based models already perform well, indicating room for more sophisticated ensemble methods.

Simple Ensembles

  1. Ensemble procedure: stacking and bagging

    avatar

  2. Baseline models: Seq2Seq, Transformer, Seq2Seq-AST-Flat, Seq2Seq-AST-GNN, Seq2Seq-AST-FC, Seq2Seq-AST-Flat-FC

Results

avatar

<HR align=left color=#987cb9 SIZE=1>