Automatic Annotation and Analysis of Oral History using LLMs: An Empirical Study of Densho Digital Collection

Abstract

Oral histories serve as vital records of lived experience, particularly within communities affected by systemic injustice and historical erasure. Effective and efficient annotation and analysis of oral history archives can promote access and use of the oral histories. However, large scale analysis of these archives remains limited due to their unstructured format, emotional complexity, and the high cost of manual annotation. This paper presents a scalable LLM-based framework to automate semantic and sentiment annotation for oral history archives, with a focus on Japanese American Incarceration Oral History (JAIOH). Using large language models (LLMs), this study seeks to construct a high-quality dataset, systematically evaluate the performance of multiple LLMs, and investigate effective prompt engineering strategies on annotation in historically sensitive contexts. Our multiphase approach combines expert annotation, prompt design, and LLM evaluation using ChatGPT, Llama, and Qwen. We labeled 558 sentences from 15 narrators for sentiment and semantic classification, then developed prompts and evaluated across zero shot, few shot, and retrieval augmented generation (RAG) strategies based on the labeled data. On this benchmark, ChatGPT achieved the highest macro-F1 score for semantic classification (88.71%), followed by Llama (84.99%) and Qwen (83.72%). For sentiment analysis, Llama performed slightly better (82.87%) than Qwen (82.66%) and ChatGPT (82.29%), with all models showing comparable results. Based on the overall evaluation, we selected the best performing prompt configurations for each task and used them to automatically annotate 92,191 sentences from 1,002 interviews in the JAIOH collection. This study develops an LLM-based framework for automatic annotation and analysis of oral history collections. The evaluation results indicate that the proposed LLM-based annotation framework provides promising performance on semantic and sentiment annotation of the JAIOH collection. It contributes a reusable pipeline and practical guidance for applying LLMs in culturally sensitive archival analysis. All code, annotated data, prompt templates, and experimental details are available at the Repository . 1

The authors' abstract, as published at the source. Journal on Computing and Cultural Heritage, 2026 · DOI ↗