It Will Never Work in Theory: Live!

Our next set of online lightning talks is happening April 25-26, 2023. Check out the speakers and get your ticket now!

Deep Code Search

Reviewed by Hamza Alvi / 2021-12-02
Keywords: Code Search, Deep Learning

Kim is a developer, and while implementing a new feature, they remember that they wrote something similar in another project. They open that project and starts searching for the code; It takes some time, but finding that code makes their current task easier. They wonder if the time spent searching was worth it, because they might have taken the same amount of time to reimplement the feature.

To help developers with tasks like this, Gu2018 proposes a tool called DeepCS that takes natural language queries and searches for relevant code in a large codebase. The CODEnn model that DeepCS uses to find relevant code snippets consists of three modules:

  1. A code embedding network (CoNN) that learns to embed code into vectors.
  2. A description embedding network (DeNN) that learns to embed natural language descriptions into vectors.
  3. A similarity module that measures the similarity between code and description vectors.

CoNN and DeNN use recurrent neural networks to embed inputs into vectors, while the similarity module uses cosine similarity measure to find the closeness between embedded inputs. Gu2018 used more than 18 million methods extracted from Java projects on GitHub to train the CODEnn model with Hinge loss as a loss function. (Hinge loss ensures that the learned vectors for description are close to the vector of corresponding code and far from vectors other code.)

DeepCS works in three steps. In the first, DeepCS takes a codebase as input and computes code vectors of methods using the CoNN module. It then takes a user query and computes the embedding vector using the DeNN module, and finally finds cosine similarity between query and code vectors obtained in previous steps and returns the ten methods with the highest similarity. To evaluate its performance, the authors collected 9950 projects having at least 20 stars from GitHub that contained more than 16 million methods. As a query, the authors used the top 50 voted Java questions from Stack Overflow. DeepCS had relevant code snippets present in the top 10 results for 86% of queries, compared to 66% for the previous state-of-the-art system. DeepCS also had 49% relevant code snippets in the top 10 results compared to only 28% for previous tools.

Gu2018 Xiaodong Gu, Hongyu Zhang, and Sunghun Kim: "Deep Code Search". Proc. International Conference on Software Engineering (ICSE), 2018, 10.1145/3180155.3180167.

To implement a program functionality, developers can reuse previously written code snippets by searching through a large-scale codebase. Over the years, many code search tools have been proposed to help developers. The existing approaches often treat source code as textual documents and utilize information retrieval models to retrieve relevant code snippets that match a given query. These approaches mainly rely on the textual similarity between source code and natural language query. They lack a deep understanding of the semantics of queries and source code. In this paper, we propose a novel deep neural network named CODEnn (Code-Description Embedding Neural Network). Instead of matching text similarity, CODEnn jointly embeds code snippets and natural language descriptions into a high-dimensional vector space, in such a way that code snippet and its corresponding description have similar vectors. Using the unified vector representation, code snippets related to a natural language query can be retrieved according to their vectors. Semantically related words can also be recognized and irrelevant/noisy keywords in queries can be handled. As a proof-of-concept application, we implement a code search tool named DeepCS using the proposed CODEnn model. We empirically evaluate DeepCS on a large scale codebase collected from GitHub. The experimental results show that our approach can effectively retrieve relevant code snippets and outperforms previous techniques.