Self-Supervised Learning for Code Retrieval and Summarization through Semantic-Preserving Program Transformations
share
Code retrieval and summarization are useful tasks for developers, but it is also challenging to build indices or summaries of code that capture both syntactic and semantic essential information of the code. To build a decent model on source code, one needs to collect a large amount of data from code hosting platforms, such as Github, Bitbucket, etc., label them and train it from a scratch for each task individually. Such an approach has two limitations: (1) training a new model for every new task is time-consuming; and (2) tremendous human effort is required to label the data for individual downstream tasks. To address these limitations, we are proposing Corder, a self-supervised contrastive learning framework that trains code representation models on unlabeled data. The pre-trained model from Corder can be used in two ways: (1) it can produce vector representation of code and can be applied to code retrieval tasks that does not have labelled data; (2) it can be used in a fine-tuning process for tasks that might still require label data such as code summarization. The key innovation is that we train the source code model by asking it to recognize similar and dissimilar code snippets through a contrastive learning paradigm. We use a set of semantic-preserving transformation operators to generate code snippets that are syntactically diverse but semantically equivalent. The contrastive learning objective, at the same time, maximizes the agreement between different views of the same snippets and minimizes the agreement between transformed views of different snippets. Through extensive experiments, we have shown that our Corder pretext task substantially outperform the other baselines for code-to-code retrieval, text-to-code retrieval and code-to-text summarization tasks.
READ FULL TEXT