LSem2Vec: Two-Stage Approach for Source Code Embedding Using LLMs
A novel technique for embedding source code, known as LSem2Vec, is presented in a paper on arXiv (identifier 2409.14644). This approach leverages large language models (LLMs) to derive semantic meaning and employs a sentence embedding model for vector representation. By utilizing a two-stage method, it effectively tackles software engineering issues such as source code clone detection and clustering without the need for expensive supervised training or fine-tuning. LSem2Vec specifically addresses the problem of inaccurate information in embeddings, a frequent challenge. The method, which stands for LLM-extracted code Semantics to Vector embedding, distinguishes between semantic extraction and vectorization, making it suitable for diverse software engineering applications. The paper highlights the significance of source code embeddings in AI-driven software development but lacks benchmark results or experimental specifics.
Key facts
- LSem2Vec is a two-stage approach for source code embedding
- It combines large language models (LLMs) and sentence embedding models
- It eliminates the need for task-specific training or fine-tuning
- It targets tasks such as source code clone detection and clustering
- It addresses erroneous information in source code embeddings
- The paper is available on arXiv with identifier 2409.14644
- The announcement type is 'replace-cross'
- The method is designed as a simpler and effective alternative to existing approaches
Entities
—