SYNC: Efficient Neural Code Search Through Structurally Guided Hard Negative Curricula

Atharva Naik, Soumitra Das, Jyothi Vedurada, Somak Aditya

Keywords: Neural Code Search, Curriculum Learning, Hard Negative Mining, Abstract Syntax Tree, Structure-aware Embeddings

TL;DR: This paper proposes an AST-guided hard negative sampling method for training efficient neural code search models using curriculum learning.

Abstract: Efficient code snippet search using natural language queries can be a great productivity tool for developers (beginners and professionals alike). Recently neural code search has been popular, where a neural method is used to embed both the query (NL) and the code snippet (PL) into a common representation space; which is further used to obtain the most relevant PL satisfying the intent in the query. Transformers-based pre-trained language models (such as CodeBERT, GraphCodeBERT, UniXCoder) have been especially effective to learn such representation. These models often make mistakes such as retrieving snippets with incorrect data types, and incorrect method names or signatures; even when exposed to the underlying structural information of the code (such as Abstract Syntax Tree and other static analysis outputs) during pre-training. The generalization ability beyond the training data is also limited (as the code retrieval datasets vary in the ways NL-PL pairs are collected). In this work, we propose a structure-aware hard negative sampling method and a mastering-rate based curriculum learning technique (SYNC) that enhances the pre-trained representation using both soft (random) and the (synthesized) hard negative samples. Our experiments on three state-of-the-art pre-trained language models for programming languages, over four Python code retrieval datasets, show the efficacy of the approach (under both in-distribution and out-of-distribution settings).

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)