ComPAT: A Compiler Principles Course Assistant

Shubin Cai, Honglong Chen, Youyi Huang, Zhong Ming

Abstract: Students in compiler principle course often encounter difficulties in understanding abstract programming concepts and need first-aid assistants in the context of independent learning. An innovative Large Language Model-based assistant ComPAT (Compiler Principles course AssisTant) is proposed in this paper. ComPAT uses vector database for text segmentation and similarity queries, meta prompting for answer refinement and challenging questions for heuristic learning. A compiler principles domain-specific dataset with 1520 questions is constructed. Compared to ChatGPT3.5, ERNIE3.5, and ChatGLM4, ComPAT has shown superior performance in handling fill-in-the-blank and question-and-answer. ComPAT demonstrates an innovative and promising use of LLM-based teaching/learning assistant system in computer education.