Structured Understanding of Unstructured Data: A Cross-Domain Review of LLM-Augmented Analysis Systems

This review surveys cross-domain academic research on structured understanding of unstructured data using large language model (LLM)-augmented analysis systems. We categorize existing work into five application domains: knowledge graph construction and reasoning, health-related social media analysis, label representation in classification, retrieval-augmented financial analytics, and LLM-driven human-computer interaction. Empirical findings show that LLMs, when integrated with retrieval modules, multi-modal knowledge graphs, or label embeddings, can perform structured extraction, classification, inference, and summarization with competitive accuracy. Benchmark evaluations across domains report LLMs exceeding human annotation consistency, enabling zero-shot classification in high-dimensional output spaces, and improving financial answer generation accuracy by over 20%. We highlight system-level design patterns, including RAG pipelines, low-coherence projection for label space compression, and chain-of-thought prompting, which support task-specific reliability. The review emphasizes only peer-reviewed and publicly reproducible work without addressing future trends or commercial applications.