Abstract:

Historical inquiry has long relied on official archives and elite writings, often marginalizing individual memories. Oral history offers a distinctive window onto overlooked social life and personal recollection; however, its non-standardized, highly context-dependent, and multi-layered textual characteristics pose challenges for structured information extraction and systematic analysis. Using oral history texts concerning ration coupons as a case study, this research explores and validates a human-AI collaboration methodology that “disciplines” large language models(LLMs) into scholarly assistants capable of strict instruction following. We design a progressive four-stage experiment—basic instructions, rule-based instructions, programmatic constraints, and few-shot learning—to iteratively optimize how to leverage LLMs, semantic understanding and instruction-following capabilities for efficient and precise structured information extraction. The findings show that the maturity of prompt engineering substantially affects output quality, and that carefully designed programmatic constraints can markedly improve the accuracy of LLM-based analyses. We further compare LLMs optimized for different tasks within a common technical framework, documenting variation in logical adherence, confirming the value of few-shot learning while identifying its point of diminishing returns, and revealing inherent limitations of LLMs in tasks requiring exact computation. The study distills an LLM “disciplining” framework for oral history text analysis that incorporates core strategies such as rule-based transduction/normalization and prudent task allocation between humans and models. The framework delivers efficient and accurate structured analysis of oral history texts and offers a reproducible, scalable intelligent research paradigm for digital humanities.

Key words: