Research

Research Overview

My research focuses on software engineering and program analysis, with the goal of improving the reliability, maintainability, and developer productivity of real-world software systems.

I am particularly interested in combining automated analysis with learning-based techniques, including large language models (LLMs), to support developers across the software lifecycle.

Research Areas

Program Analysis and Software Evolution

Analysis of software changes, refactorings, and recurring modification patterns
Techniques to support code review, regression testing, and program comprehension
Improving the transparency and interpretability of software evolution

Software Correctness

Automated detection and explanation of software defects
Integration of static analysis, statistical methods, and machine learning
Developer-facing tools that provide actionable feedback

AI & LLM-assisted Software Engineering

Combining large language models (LLMs) with traditional program analysis
Applications to code completion, code summarization, and human–AI collaboration
Emphasis on reliability and trustworthiness in developer tools

AI-assisted Programming Education

Intelligent tutoring systems for programming education
Automated feedback to support effective software development practices
Reinforcing fundamental software engineering principles

Selected Recent Publications and Projects

Intelligent Code Completion by a Unified Multi-task Learning with a Large Language Model (SERA 2025)

Shradha Maharjan, Meng Xia, Tae-Hyuk Ahn, Myoungkyu Song

This work introduces CODECOM, a deep learning-based code completion technique that integrates a large language model with program analysis to provide more accurate and context-aware code suggestions. By leveraging source code tokens, abstract syntax trees, and program dependencies, CODECOM significantly outperforms existing approaches and helps developers write correct code more efficiently.

Automated Code Summarization by Training Large Language Models with Crowdsourced Knowledge (SERA 2025)

Meng Xia, Shradha Maharjan, and Myoungkyu Song

This paper presents DEEPKNOWLEDGE, an automated code summarization approach that leverages a large language model trained with crowdsourced knowledge from GitHub and Stack Overflow. By integrating real-world coding examples and developer discussions, the approach generates more accurate and context-aware summaries that explain code behavior, implementation rationale, and usage guidelines. Extensive evaluation on large-scale Java datasets shows that DEEPKNOWLEDGE significantly outperforms state-of-the-art techniques, improving BLEU scores by up to 39%, and demonstrating its effectiveness in supporting program comprehension and maintaining up-to-date documentation in evolving software systems.

SYNC: Synergistic Annotation Collaboration between Humans and LLMs for Enhanced Model Training (SERA 2025)

Tommy Le, Will Taylor, Shradha Maharjan, Meng Xia, and Myoungkyu Song

This paper introduces SYNC, a human-LLM collaborative annotation framework designed to improve the accuracy and reliability of data annotation for Stack Overflow datasets. SYNC combines multiple automated annotation strategies—including TF-IDF-based lexical matching, transformer-based semantic similarity, and code-aware embeddings using UniXcoder—with human verification and refinement. By integrating automated efficiency with human oversight through web, mobile, and desktop interfaces, the approach produces higher-quality annotations that better support downstream tasks such as code retrieval, summarization, and analysis in large language model-based programming assistants.

SYNCode: Synergistic Human-LLM Collaboration for Enhanced Data Annotation (Information 2025, extended version of SERA 2025)

Meng Xia, Shradha Maharjan, Tommy Le, Will Taylor, Myoungkyu Song

This paper presents SYNCode, a human-LLM collaborative framework for high-quality data annotation in code-centric domains such as Stack Overflow. The approach combines lexical and semantic analysis with code-aware representations to generate initial annotations, which are then iteratively validated and refined by human annotators through an interactive system. Experimental results show that SYNCode improves annotation accuracy and scalability while reducing bias, supporting downstream software engineering and language model-based analysis tasks.

A Machine-Learning Approach to API Usage Learning for Detecting API Misuse Errors

PIs: Myoungkyu Song (UNO), Harvey Siy (UNO), Youngjin Kwon (UNMC), Kwangsung Oh (UNO), and Na Zhong (UNO)

Nebraska Research Initiative (NRI), University of Nebraska System. Project Period: 2023-2025.
This Nebraska Research Initiative-funded project investigates learning-based approaches for modeling correct API usage and detecting API misuse errors. By combining program analysis with machine learning and large language models, the project supports research on code completion, code summarization, and developer-facing program understanding tools, contributing to multiple recent peer-reviewed publications.

Collaboration

I actively collaborate with faculty and students across software engineering, AI/ML, programming languages, automated program analysis, and large language models. I am also committed to extending these efforts beyond academia and actively seek partnerships with industry leaders and national laboratories to translate research innovations into practical solutions and validate research outcomes in real-world software ecosystems.