Advancing query-focused summarization for unstructured and structured sources

The rapid growth of digital information presents major challenges in identifying and synthesizing content relevant to user needs. Automatic summarization provides an effective solution by condensing diverse sources into concise outputs, yet generic approaches often fail when users seek answers to specific questions. Query-Focused Summarization (QFS) addresses this gap by tailoring summaries to user queries, but current methods face persistent limitations in scalability, faithfulness, and the handling of structured data. This thesis advances QFS across both unstructured text and structured tables. For unstructured text, we explore integrating retrieval to eliminate reliance on preselected document sets and evaluate the faithfulness of generated summaries through a comparative meta-evaluation framework. For structured tables, we introduce query-focused multi-table summarization, developing new datasets and approaches that combine reasoning and summarization based on large language models (LLMs). To enhance table reasoning capabilities of LLMs, we propose a structured planning mechanism that replaces ambiguous natural language plans with explicit, executable representations, enabling more robust and scalable multi-step table reasoning. Across both domains, we demonstrate consistent improvements over strong baselines and highlight key insights into retrieval, faithfulness, and table reasoning. Collectively, these contributions establish QFS as a more practical, reliable, and user-centric technology, offering a foundation for future research at the intersection of retrieval, summarization, and structured reasoning.