On the effectiveness of large language models for query expansion in code search

Language Models (LMs) are deep learning models trained on massive amounts of text data. One of their main advantages is their superior language understanding capabilities. This study explores the application of Large Language Models (LLMs) understanding capabilities in code search query expansion. To this end, we collected a query corpus from multiple data sources and trained multiple LMs (GPT2, BERT) on this query corpus using a self-supervised task. The trained LM models are then used to expand the input query. We evaluate the performance of these LLMs on the CodeSearchNet dataset using two state-of-the-art code search methods (GraphCodeBERT and CoCoSoda) and compare these LLMs with currently popular expansion methods. Experimental results show that LLM-based query expansion methods outperform existing query reformulation methods in most cases.