ETD

• Code Summarization
• Code Understanding
• Data Lineage
• Large Language Models

This thesis investigates how Large Language Models (LLMs) can support software understanding by generating natural-language explanations and structured artifacts directly from source code, with a stronger emphasis on data lineage. The work considers two complementary tasks. The first is behavior-focused code summarization, where models must produce short descriptions of what a function does under strict generation constraints; the thesis test an evaluation protocol that combines human assessment and LLM-as-a-judge scoring across multiple quality dimensions. The second task targets column-level backward data lineage for ETL-like scripts, where the goal is to infer and represent dependencies between input and output columns as a structured edge list, enabling traceability of transformations in data pipelines. Across both tasks, the thesis discusses dataset construction, annotation choices, prompt and output constraints, and the methodological challenges of evaluating model outputs when only static code is available (e.g., ambiguity, underspecification, and format compliance). Overall, the thesis provides a systematic framework for studying LLM-based software explanation and offers practical guidance for designing reliable evaluation setups for code summarization and data lineage extraction.