Automated Software Engineering-2021 Automating just-in-time comment updating

· Automated Software Engineering· · code-summarization

Paper: Automating just-in-time comment updating

CUP: A New Tokenizer and Separate Encoders

Abstract

The task of helping developers update comment text when code changes is termed Just-in-Time (JIT) Comment Updating; the authors propose CUP for this setting.

They introduce several tailored enhancements: a specialized tokenizer and a joint attention mechanism.

The method outperforms the previous best baseline by 7×.

Introduction

  • Poorly updated comment text can mislead developers and is often useless.

  • This approach does not regenerate comments from scratch, so two aspects matter: (1) maintain consistency between old and new comments; (2) the old comment is the basis for the update, so an end-to-end model must learn both the code edit and a representation of the prior code.

    • A new tokenizer and copy mechanism
    • Two encoders—one for the code change and one for the old comment
    • A joint code–comment vocabulary with word embeddings from a pretrained fastText model
    • A joint attention mechanism
  • Data: 1,496 Java projects on GitHub, yielding 108K samples pairing code comments with their modifications

Model

avatar

Data Flattening

  • Tokenization

    • Remove spaces, retain punctuation, and split compound identifiers; when two adjacent tokens are not separated by whitespace, insert a <con> marker.
    • Three common tokenization strategies: leave compound words intact; split them directly; or append a special symbol </t> (e.g., "inputBufer" -> "inputBufer</t>"). The first fails to remove OOV words, the second loses formatting, and the third cannot handle cases where sub-tokens of a compound word are predicted as independent tokens.

  • Code-change representation

    • Each edit comprises three segments.
    • avatar

Seq2Seq Model

avatar
  • Each encoder stacks four layers: embedding, contextual embedding, joint attention, and model layers.
    • Embedding layer: shared vocabulary; embeddings initialized from fastText pretraining.

Data

  • 1,496 repositories with at least 500 commits each.
  • Method-level edits: GumTree is used to derive method-level change structure between two revisions.
  • Data-processing details are lengthy; consult the original paper when implementing.

Experiment

avatar
avatar
avatar

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